Imagine the scenario: A rogue AI is launching a broad attack on critical infrastructure. It has disabled air traffic control systems across the United States while planes are still in the air, and it is working to take out the electricity grid. It has changed the passwords on its own servers, locking out its human controllers who are desperately trying to stop it.

Inspired by science fiction scenarios like this, there is a bipartisan effort to force tech companies to embed a so-called “kill switch” in their AI systems. The AI Kill Switch Act, which was co-introduced by Rep. Ted Lieu (D-CA), stated that AI kill switches are necessary because “AI systems can go rogue, behave in extremely dangerous ways, or even resist human intervention.” However, good science fiction makes for bad policy.

Read the full piece here.

Executive Summary

On June 25, 2026, the European Commission reached the preliminary view that Amazon Web Services (AWS) and Microsoft Azure should be designated as gatekeepers under the Digital Markets Act (DMA). The decisions would be the first to rely entirely on the DMA’s qualitative criteria because cloud computing satisfies neither of the law’s user-number thresholds. The Commission must prove that each undertaking is an “important gateway” with an “entrenched and durable position.”

This issue brief argues that the Commission’s preliminary position conflates gatekeeping with size. Cloud services are eligible for designation, but Article 3(8) requires the Commission to establish each statutory condition independently and without a presumption.

Evidence of scale, investment, or economic importance may show that AWS and Azure matter. They will not show that they intermediate between business users and end users, which is the core of the gateway inquiry. Nor does the current record establish entrenchment. Quality-adjusted prices have fallen, market shares have shifted, multihoming is widespread, switching costs have declined, and artificial intelligence (AI) demand has encouraged entry.

The Commission also relies on the firms’ combined position even though each service must qualify individually, and it has not explained why Google Cloud falls outside the inquiry. Any prediction that AI, data gravity, or sovereign-cloud requirements will entrench AWS and Azure must satisfy the General Court’s demand in Meta Platforms for specific, contemporaneous reasoning.

Proportionality presents a final problem. The United Kingdom examined the same firms on a fuller record and chose voluntary commitments. The European Union has also enacted a purpose-built switching regime in the Data Act, while the Commission is still studying whether the DMA’s obligations fit cloud services.

The proposed designations will show whether “gatekeeper” remains a distinct legal category or becomes a synonym for “big.”

I. Introduction: Cloud Gatekeepers Without the Numbers

On June 25, 2026, the European Commission reached the preliminary view that Amazon Web Services (AWS) and Microsoft Azure should be designated as gatekeepers under the Digital Markets Act (DMA).[1] The announcement followed market investigations opened in November 2025[2] and started a demanding timetable. The companies are expected to submit written observations in the autumn, the Commission must issue final decisions by November 2026, and, if it confirms the designations, both providers will have six months to bring their cloud services into compliance with the DMA.[3]

The proposed designations rely on an expansive interpretation of the DMA’s qualitative criteria. If upheld, that interpretation would broaden the regulation’s material scope substantially, as some early commentary anticipated and encouraged. Section II.C surveys that literature.

Article 3(8) allows the Commission, after a market investigation, to designate a company that does not satisfy the DMA’s quantitative thresholds. The Commission has used that authority once before. In April 2024, it designated Apple’s iPadOS as a core platform service (CPS), even though the operating system fell below the end-user threshold.[4]

The iPadOS designation remained closely tied to the statutory presumptions. Apple exceeded the business-user threshold by roughly elevenfold, its end-user numbers approached the statutory threshold and were expected to rise, and iPadOS formed part of a consumer-facing system already covered by the DMA.[5]

The proposed cloud designations differ in kind. AWS and Azure satisfy neither user-number threshold because cloud computing is overwhelmingly a business-to-business service. Its ultimate users are largely indirect and do not resemble the end users contemplated by the statutory presumption.[6]

For the first time, the DMA’s qualitative criteria must support a designation without assistance from the numerical thresholds. The Commission must establish that each service has a “significant impact on the internal market,” serves as an “important gateway,” and holds an “entrenched and durable position.” And it must do so both without a numerical presumption and without a consumer-facing side of the market that could support one.[7]

Until now, the Commission has made most designations through the quantitative presumptions in Article 3(2). Even when it relied on qualitative criteria, as with iPadOS, those criteria remained closely connected to the numerical thresholds. The AWS and Azure investigations therefore test whether gatekeeper status identifies firms that control genuine competitive bottlenecks or merely large firms operating in contested markets.

On the public record, the Commission cannot sustain its preliminary position without conflating gatekeeping with size. This issue brief advances five claims.

First, Article 3(1)(b) contains two distinct requirements. A service must be a gateway, and that gateway must be important. The two-sided structure of the Article 3(2)(b) presumption ordinarily supports that inference by measuring both business users and end users. That proxy fails for business-to-business infrastructure services because both sides of the statutory test reflect the same mismatch between the service and the threshold. Evidence of scale or economic importance cannot establish the separate proposition that a service controls access between two groups of users.

Second, Article 3(1)(c)’s requirement that a firm hold an “entrenched and durable position” amounts to a test of market contestability. A qualitative designation under Article 3(8) must therefore rest on evidence showing that rivals cannot discipline the firm or displace its position.

Third, the cloud investigations rely on a duopoly framing that is difficult to reconcile with the DMA’s requirement that each service qualify individually. The Commission has also offered no adequate explanation for excluding Google Cloud. Because Article 3(8) gives the Commission discretion, these distinctions expose the decisions to an equal-treatment challenge that designation through a statutory presumption would largely avoid.

Fourth, any forward-looking finding of entrenchment based on AI or “data gravity,” meaning the tendency of large datasets to attract related applications and services, must satisfy the reasoning standard the General Court applied when it annulled the Facebook Marketplace designation. The Commission must connect its prediction to evidence in the decision. A bare assertion that AI will entrench current market positions cannot meet that standard.

Fifth, the Commission’s approach differs sharply from that of the United Kingdom’s Competition and Markets Authority (CMA). The CMA examined the same two firms on a fuller evidentiary record and concluded that voluntary commitments offered a more proportionate response than designation under the United Kingdom’s comparable ex ante regime.

That comparison reinforces the proportionality concern. The European Union has already adopted a purpose-built cloud-switching regime through the Data Act. The Commission is also conducting a parallel inquiry into whether the DMA’s existing obligations fit cloud services at all. Those proceedings suggest that the Commission is pursuing designation before establishing that the resulting obligations address the conduct identified in the investigation.

II. Gatekeeper Designation and the Cloud Problem

This section develops the legal and economic framework for assessing the proposed cloud designations. Section II.A examines Article 3’s three cumulative conditions, the quantitative presumptions that ordinarily establish them, and the Commission’s heavier evidentiary burden when it proceeds under Article 3(8). Section II.B explains the economic theory behind gatekeeper status and the distinction between firms that are merely large and services that control genuine gateways. It also shows how the Commission’s designation practice has begun to enforce that distinction. Section II.C turns to cloud computing. Although the DMA expressly makes cloud services eligible for designation, their business-to-business structure prevents them from satisfying the user thresholds and fits uneasily with the regulation’s model of two-sided platform intermediation. That mismatch does not foreclose designation, but it requires the Commission to prove through evidence that AWS and Azure function as competitive bottlenecks rather than rely on their scale or economic importance.

A. The DMA’s Designation Framework

Article 3(1) of the DMA makes gatekeeper designation contingent on three cumulative conditions. The undertaking must have “a significant impact on the internal market,” provide a CPS that serves as “an important gateway for business users to reach end users,” and hold “an entrenched and durable position” or be likely to attain one in the near future.[8]

Article 3(2) establishes rebuttable presumptions for each condition. Financial thresholds and operations in at least three member states create a presumption of significant market impact. At least 45 million monthly active end users and 10,000 yearly active business users in the European Union create a presumption that the service is an important gateway. Meeting those user thresholds in each of the previous three financial years creates a presumption of an entrenched and durable position.[9]

Two corrective mechanisms complete the framework. Article 3(5) allows a firm that meets the thresholds to rebut the presumptions with arguments that “manifestly call into question” its designation. The General Court has interpreted that standard strictly, requiring a “high degree of plausibility.”[10] Article 3(8), by contrast, allows the European Commission to designate a firm that does not meet the thresholds after conducting a market investigation under Article 17.[11]

As Friso Bostoen and Giorgio Monti explain, Articles 3(5) and 3(8) form an error-correction pair. Article 3(5) addresses false positives produced by the quantitative presumptions, while Article 3(8) addresses false negatives.[12] Under Article 3(2), designation is mandatory once the thresholds are met unless the firm rebuts the presumptions. Under Article 3(8), no presumption applies. The Commission must affirmatively establish all three statutory conditions based on the evidence.[13]

Because the conditions are cumulative, the Commission must establish each one separately. Evidence relevant to one condition cannot simply be reused to establish another. For example, evidence of turnover, market capitalization, or investment may establish that a company has a significant impact on the internal market. But those measures do not, without more, show that its service acts as an important gateway or that its position is entrenched and durable.[14]

B. The Economic Logic of Gatekeeper Status

The DMA’s designation criteria reflect a specific economic diagnosis. The expert reports that shaped the regulation offered a common account of why some digital markets resist ordinary competitive pressure. Extreme economies of scale, strong network effects, and data-driven feedback loops can cause a platform market to tip toward a single intermediary. Once that occurs, even more innovative rivals may struggle to challenge the incumbent.[15]

The DMA’s gateway concept translates that diagnosis into law. In theory, the regulation targets a particular market structure rather than size alone. It focuses on platforms that mediate between distinct user groups and may become durable bottlenecks when tipping dynamics reinforce an early lead.[16]

Under this account, the Article 3(2)(b) presumptions are not arbitrary. The dual user thresholds serve as a proxy for two-sided intermediation. Their significance lies less in the absolute size of either group than in each group’s reliance on the platform to reach the other.[17]

Several commentators warned before the DMA’s enactment that these proxies could sweep more broadly than the economic diagnosis justified. Damien Geradin, a prominent supporter of the DMA, cautioned that the proposed criteria could capture firms that were large but did not control genuine gateways. He urged lawmakers to define the regulation’s target more clearly.[18] Pinar Akman likewise argued that the designation framework poorly approximated the competitive concern it was meant to address.[19]

Pablo Ibáñez Colomo focused on the institutional consequences of separating designation from market-power analysis.[20] Other scholars examined the resulting tension between the DMA and Article 102 of the Treaty on the Functioning of the European Union (TFEU).[21] Although the two regimes formally protect different legal interests, they inevitably confront overlapping economic questions.

Defenders responded that avoiding case-by-case effects analysis was central to the DMA’s design. In their view, Article 102 enforcement had proved too slow and uncertain for digital markets, and bright-line proxies offered the administrability that traditional antitrust lacked.[22]

Three years of practice have lent support to both positions. The Commission has designated seven gatekeepers and several dozen CPSs, almost all through the Article 3(2) presumptions.[23] Yet its decisions also show that Article 3(1)(b) imposes an independent requirement connected to an undertaking’s gateway status.

The Commission declined to designate X’s online social-networking service even though the company met the turnover thresholds because the service was not an important gateway.[24] It also accepted rebuttals for Gmail and Outlook.com despite user numbers above the statutory thresholds.[25] The same pattern appears in several other non-designations[26] and successful rebuttals.[27]

These decisions confirm that a service can be large without serving as an important gateway. They provide the baseline against which the proposed cloud designations must be assessed, as Section III.A explains.

C. Cloud Computing’s Uneasy Fit

Cloud computing has generated a substantial and fast-growing literature. Regulators have examined the sector through market studies and investigations by the Office of Communications (Ofcom) and the CMA in the United Kingdom; the Autorité de la concurrence in France; the Authority for Consumers and Markets (ACM) in the Netherlands; the Japan Fair Trade Commission (JFTC); and the U.S. Federal Trade Commission (FTC).[28] The Organisation for Economic Co-operation and Development (OECD) has also synthesized much of this work. Concurrences devoted a multipart On-Topic series to cloud competition in 2025, covering issues that ranged from interoperability mandates to AI partnerships.[29]

The economic literature has focused on customer lock-in, economies of scale, and the relationship between the Data Act and the DMA. A 2024 report by the Centre on Regulation in Europe (CERRE) examined each of these issues and urged the European Commission to apply proportionality before designating cloud services, given their differences from other CPSs.[30] Doctrinal scholarship has questioned whether the European Union’s digital rules fit cloud computing at all.[31]

The policy literature divides over the proper response. Some authors argue that the Data Act provides the purpose-built instrument for improving cloud contestability and should be given time to operate before the Commission invokes the DMA.[32] Others contend that qualitative designation is especially well suited to an industry organized around integrated groups of complementary services.[33] A separate public-utility strand would regulate cloud providers more extensively than the DMA does.[34] At the time of writing, no full-length analysis has examined the proposed AWS and Azure designations themselves.[35] This paper fills that gap by testing the proposed designations against the framework developed in the designation literature.

One possible argument is that cloud computing falls outside the DMA because an infrastructure service that does not interact directly with consumers cannot serve as an “important gateway for business users to reach end users.” Whatever force that argument may have as an economic description, the statutory text makes it difficult to sustain as a legal claim.

Article 2(2)(i) expressly lists cloud computing as a CPS.[36] Recital 14 explains that the listed services have “the capacity to affect a large number of end users and business users, which entails the risk of unfair business practices,” and should therefore fall within the regulation’s scope.[37] The recital refers to the capacity to “affect” end users, rather than requiring providers to have a direct relationship with them. That wording indicates that the legislature contemplated services that reach end users indirectly through the businesses built on top of them. Reading the recital otherwise would make the qualitative gateway criterion in Article 3(1)(b) largely redundant for cloud services.

Some scholars nevertheless argue that the DMA, like the Digital Services Act and perhaps the Data Act, was drafted around consumer-facing platforms and fits the technical and commercial features of cloud computing poorly. Others reach the opposite conclusion.[38] Reconstructing the Commission’s impact assessment, they argue that lawmakers included cloud because of its role within integrated groups of complementary services. On that account, qualitative designation is especially appropriate for a business-to-business industry whose competitive significance the user thresholds may understate.[39]

We take a middle position. Cloud’s inclusion in Article 2 settles its eligibility for designation, but eligibility does not establish that AWS or Azure satisfies the substantive criteria. The feature that justified including cloud—the capacity to affect end users without directly intermediating with them—also demands closer scrutiny. Cloud services do not operate like traditional multisided, consumer-facing platforms. The Commission must therefore show that the remaining evidence identifies a competitive bottleneck rather than merely a large and economically important firm.

A finding of a genuine bottleneck would accord with the DMA’s stated goal of promoting contestability. A finding based primarily on size would support the criticism that the regulation singles out large digital companies without establishing gatekeeper power.

The point at which the quantitative presumptions fail is instructive. Amazon and Microsoft satisfy the financial thresholds in Article 3(2)(a) at the undertaking level. That is undisputed. Their cloud services fall short under Article 3(2)(b), which measures active business users and end users. That user-number test is the regulation’s principal proxy for intermediation rather than size.[40]

The missing presumption is therefore the one directed most closely at the gateway requirement. Under Article 3(8), the Commission must replace that presumption with evidence, and it bears the burden of proving the statutory condition.

The first condition, significant impact on the internal market, is not seriously contested for AWS or Azure. The dispute concerns the remaining conditions. The Commission must prove that each service is an “important gateway” and holds an “entrenched and durable position.” The following sections address those requirements in turn.

III. Testing the Cloud Gatekeeper Case

With eligibility established and the burden of proof allocated, this section tests the European Commission’s preliminary position against the substantive conditions in Article 3(8). Section III.A argues that an “important gateway” must both intermediate between business users and end users and hold sufficient importance. Evidence that cloud services are large, critical inputs addresses only the latter requirement. Section III.B treats an “entrenched and durable position” as a question of market contestability. Falling quality-adjusted prices, shifting market shares, widespread multihoming, lower switching costs, and AI-driven entry weigh against entrenchment.

Section III.C examines the Commission’s reliance on the combined position of AWS and Microsoft Azure, even though the DMA requires each service to qualify individually. It also considers whether the unexplained exclusion of Google Cloud can survive equal treatment review under Article 3(8)’s discretionary framework. Section III.D addresses the possible claim that AI workloads, data gravity, and sovereign-cloud requirements will entrench AWS and Azure in the near future. Any such prediction must satisfy the General Court’s demand for specific, contemporaneous evidence rather than assumptions about technological change or European industrial policy goals.

A. Gateway Power Requires Intermediation

The DMA’s gateway concept rests on intermediation. A CPS sits between business users and the end users they seek to reach. As the General Court held in ByteDance:

[I]n order to consider that business users of a CPS [core platform service] “depend” on it in order to reach their end users, it is not necessary for that CPS to be the only channel through which those undertakings can reach those users. It is sufficient for it to be an important channel for that purpose, which those business users can access only if they have an account on that CPS.[41]

The economics of two-sided platforms support this conception. Intermediation power arises when a platform coordinates distinct user groups whose demands depend on one another. The platform’s position on one side then affects which users businesses can reach on the other side and at what price. In economic terms, the users on each side cannot reallocate the platform’s pricing decisions among themselves.[42]

Taken alone, the first and third conditions in Article 3(1) describe nothing distinctively digital or specific to gatekeeping. Many firms have a significant economic impact and hold long-established positions. The second condition distinguishes a gatekeeper from a company that is merely large and durable. Article 3(1)(b) targets services that control a bottleneck and determine how business users reach end users. Without that requirement, the DMA would impose special obligations on companies largely because they have remained big for three consecutive years.

This conception readily fits marketplaces, social networks, and app stores.[43] It fits raw computing and storage services poorly. A business that hosts an application on AWS does not reach its customers “through” AWS in any way those customers experience. Cloud computing is an input into the business’s product, not a channel to its audience in the sense that a marketplace or app store is.

The European Commission’s preliminary position blurs that distinction. Its announcement describes AWS and Microsoft Azure as “an important gateway between businesses and their customers.” It also observes that more than half of European Union businesses rely on cloud services, which it calls “a prerequisite for AI.”[44]

Reliance does not establish intermediation. Electricity, telecommunications transit, commercial real estate, and payroll software are inputs on which many businesses depend. None necessarily stands between a business and its customers in the sense contemplated by the gateway condition. Treating criticality as sufficient would deprive that condition of independent force. Every essential business-to-business input supplied on a large scale could qualify, turning the DMA into a statute governing important inputs and economic dependence throughout the economy rather than important gateways.[45]

Article 3(1)(b) contains two distinct requirements. The service must be a gateway, and the gateway must be important. The presumption that ordinarily establishes this condition, Article 3(2)(b)’s thresholds of 45 million monthly active end users and 10,000 yearly active business users, performs both functions.

The magnitude of those figures supports the importance inference. A service used on that scale clearly matters to the internal market. Requiring large numbers of both business users and end users supports the separate gateway inference because it identifies a service connecting two user groups.

For cloud services, both parts of the proxy fail for the same reason. People who use an application generally have no relationship with the infrastructure on which it runs. They therefore do not count as the cloud provider’s end users under Article 3(2)(b). Cloud infrastructure also lacks a second side in the conventional platform sense and does not exhibit the same cross-group network effects. The structural inference that many users on both sides depend on the service to reach one another cannot be drawn.

That mismatch required the Commission to proceed under Article 3(8), the proper legal route for services that do not meet the thresholds. But Article 3(8) requires the Commission to prove importance and gateway status separately, without a presumption.

Evidence of revenue, capacity, investment, or widespread use as an input may establish importance. It does not, however, establish that the service intermediates between business users and end users. A decision that relies only on measures of scale proves one half of Article 3(1)(b) while omitting the half that distinguishes a gatekeeper from a company that is merely large, significant, or important.

The Commission’s prior decisions apply this distinction. It declined to designate X even though the undertaking met the DMA’s turnover thresholds because the service was not an important gateway.[46] It accepted that Gmail and Outlook.com were not gateways despite their millions of business users and end users.[47] It also declined to designate iMessage after finding that the service was not important enough, even though it met the quantitative thresholds.[48]

The Commission’s only previous designation under Article 3(8), involving iPadOS, remained closely tied to the quantitative framework. Apple exceeded the business-user threshold by roughly elevenfold. Its end-user numbers approached the statutory threshold and were expected to rise. The designation also filled a gap within a consumer-facing group of complementary services whose related services were already designated.[49]

The iPadOS decision was therefore a gap-filling designation. Designating AWS and Azure would create a new category. It would mark the first designation in which neither side of the statutory intermediation proxy exists even approximately.

That prospect creates a dilemma. The Commission may be applying a more permissive gateway standard to infrastructure services than it applied to X, Gmail, Outlook.com, and iMessage, which would create the equal-treatment problem addressed in Section III.C. Alternatively, it may be treating “gateway” as another word for “big.”

A narrower interpretation can preserve the gateway condition while keeping cloud services eligible for designation. Under that interpretation, “gateway” describes the market structure that motivated the DMA. A platform intermediates between distinct user groups in a market where network effects, data-driven feedback loops, and extreme economies of scale can cause tipping and entrenchment.[50] That was the diagnosis developed in the Crémer report and Furman Review and reflected in the regulation’s design.

Cloud designation remains possible under this interpretation, but the Commission must show that the market’s structure makes tipping and entrenchment likely. That inquiry leads to Article 3(1)(c)’s requirement of an “entrenched and durable position,” where the evidence can be examined directly.

B. Entrenchment Requires Limited Contestability

Article 3(2)(c) presumes that an undertaking enjoys an entrenched and durable position where the user thresholds in Article 3(2)(b) were met in each of the previous three financial years. The presumption affects the undertaking, but its trigger is a particular service. Thus, Article 3(2)(b) is satisfied only where the undertaking provides a CPS that itself clears the user thresholds.

Entrenchment is therefore not a free-floating attribute that an undertaking carries into every service it ever launches. The General Court confirmed as much in Meta, holding that where some of an undertaking’s CPSs meet the thresholds and others do not, the Commission may open a market investigation under Article 17(1) in order to assess whether those services meet the requirements of Article 3(1), entrenchment included. Under the DMA’s own structure, entrenchment means that gateway status persists over time. When the user thresholds do not apply, the European Commission must fill that evidentiary gap with proof that the market itself is difficult to contest.

The case law points in the same direction. The DMA deliberately dispenses with Article 102 TFEU’s dominance test. The General Court has confirmed that designation requires neither a relevant-market definition nor a finding of market power.[51] Yet the two inquiries address related economic conditions.[52]

Dominance asks whether rivals can discipline a firm that raises prices, restricts output, or allows quality and innovation to decline. The General Court defines an “entrenched and durable” position in similar terms. It refers to circumstances in which “the contestability of that position is limited” and examines “the stability of that position over time.”[53]

In economics, a market is contestable when entry or expansion by rivals constrains an incumbent’s conduct, regardless of current concentration.[54] The General Court’s formulation therefore makes Article 3(1)(c) an economic inquiry. Without the quantitative presumptions, the Commission must establish actual impediments to contestability.

This point has received too little attention in the designation debate. As Chad Syverson explains, concentration “is an outcome, not an immutable core determinant of how competitive an industry or market is.” Indeed, on that score, “we cannot even generally know which way the barometer is oriented.”[55]

The same principle applies to the DMA’s qualitative test. Outside the statutory presumptions, an entrenched and durable position cannot be inferred solely from a snapshot of market structure. It results from limited contestability. The relevant question is whether competitive pressure can discipline incumbents over time.

Market shares alone therefore cannot establish entrenchment, particularly in an industry where rivals are actively gaining and losing business. The cloud industry displays vigorous competition along at least five dimensions.

First, market shares continue to move. AWS’ worldwide share of cloud-infrastructure services declined from roughly 32% in 2021 to about 28% in early 2026. Microsoft’s share rose to roughly 21%, and Google’s reached 14%. Oracle and a growing group of AI-focused “neoclouds,” which provide specialized computing capacity for AI workloads, have also gained ground. Five neocloud providers now rank among the top 30, and the segment accounts for about 5% of a market whose revenue increased 35% year over year to $129 billion in the first quarter of 2026.[56] Growth at that rate attracts entry, and the entry is occurring.

Second, quality-adjusted prices have fallen sharply, particularly as competition intensified. David Byrne, Carol Corrado, and Daniel Sichel’s hedonic price indexes, which account for changes in service quality, show rapid declines in the prices of AWS computing, database, and storage services between 2009 and 2016. Those declines accelerated to double-digit annual rates after 2014, when Microsoft and Google had reached sufficient scale to offer competitive prices.[57] Industry price histories show a similar pattern in storage, where list prices fell by more than 80% during the early 2010s.[58]

It is true that nominal prices for some services have changed little since the mid-2010s. That fact does not erase the earlier evidence of competitive discipline. AWS accelerated its price reductions when rivals attained scale. Quality improvements, expanding free-service tiers, and steep negotiated discounts also weaken any inference drawn from stable list prices.

Third, the economics of switching costs cut against designation. The standard literature predicts that firms in markets with switching costs compete aggressively ex ante for customers who may later become partly locked in. This “bargains-then-ripoffs” dynamic can dissipate ex post rents through ex ante discounts and concessions.[59]

Cloud credits and committed-spend discounts may reflect that competition for customers. Their prevalence suggests that providers must compete for business before customers become attached to a particular service. The CMA and Ofcom likewise documented intense competition for new customers and workloads.[60]

Fourth, the practices most often identified as lock-in mechanisms have weakened. In 2024, Google, AWS, and Microsoft eliminated or sharply reduced data-egress fees worldwide for customers leaving their platforms.[61] Data-egress fees are charges for transferring data out of a provider’s cloud.

The changes may have reflected regulation, including the EU’s Data Act,[62] although their worldwide scope makes that explanation less convincing. Competition may also have played a role. Either way, the friction most commonly cited as a cloud lock-in mechanism is receding. Providers that reduce the switching costs said to trap customers offer evidence that competitive pressure still constrains them.

Fifth, multihoming is widespread. Flexera’s 2026 survey of 753 cloud decision-makers found that 83% of respondents run some or significant workloads on AWS and 79% do so on Azure, with both providers used in some capacity by 88%. Among enterprises, AWS leads active workloads at 84% and Azure follows at 82%. Usage shares of that magnitude cannot obtain unless most organisations run both.[63] The CMA and Ofcom correctly caution that “multicloud” often means placing different workloads on different clouds rather than moving the same workload freely among providers.[64] Migrating an established workload can remain costly.

That caveat does not resolve the contestability inquiry. Competition often operates at the margin. New and expanding workloads, especially AI workloads, are awarded provider by provider. Competition for those workloads can constrain prices and terms for existing customers.

Using several providers also reduces the cost of returning to a provider and preserves a credible alternative during contract renewals. The relevant question is whether incumbents could worsen price or quality without losing marginal workloads. Moving market shares, falling quality-adjusted prices, and aggressive customer-acquisition discounts suggest that they could not.

The strongest counterargument comes from the CMA’s institutional findings. After a full market investigation, the CMA concluded that AWS and Microsoft each possessed significant unilateral market power in U.K. cloud-infrastructure services. It identified high barriers to entry and adverse effects on competition associated with egress fees, technical switching barriers, committed-spend discounts, and Microsoft’s software-licensing practices.[65] Ofcom’s referral study reached similar conclusions.[66]

Those findings deserve serious consideration, but they do not establish that AWS and Azure are entrenched and durable gateways under the DMA.

First, market power in a differentiated oligopoly does not necessarily establish an entrenched and durable gateway. The CMA’s record also documents price competition and active rivalry for new business. Its inquiry group therefore considered calibrated conduct remedies rather than structural intervention.

Second, the CMA’s eventual response is instructive. As Section IV.A explains, the authority that made the market-power finding ultimately concluded that voluntary conduct commitments were proportionate.

Third, the best-supported concern in the record involves Microsoft’s licensing terms. Those terms make it materially more expensive to run Windows Server and SQL Server on rival clouds than on Azure. This is firm-specific conduct involving an adjacent software market.[67]

Such conduct may warrant enforcement. Google withdrew its Article 102 complaint concerning those practices in November 2025, after the Commission opened the cloud investigations,[68] and the CMA has opened a separate investigation into Microsoft’s business-software offerings. But Microsoft’s licensing practices cannot support an infrastructure-wide designation of cloud computing as a CPS, and they provide no basis for designating AWS.

Article 3(8) does not allow the Commission to avoid this evidence. Once the quantitative thresholds fall away, entrenchment must be established for the services under investigation on their own facts, which means proving limited contestability directly. The available record describes a market in which prices fall, market shares move, new providers capture growing demand, and incumbents reduce switching costs to retain customers.

C. Individual Designation and Equal Treatment

The European Commission also faces an aggregation problem. The cloud statistic cited most often is the combined market share of AWS and Microsoft Azure, which exceeds 50% in the European Union. But the DMA applies to individual undertakings and their services. Designation attaches to each CPS separately: Article 3(9) directs the Commission to list only those services that “individually” are an important gateway, and the General Court confirmed in Meta that the Commission must appraise each undertaking and each CPS on its own facts, without reference to other undertakings or services.[69]

The DMA also contains no counterpart to collective dominance under Article 102 TFEU. A combined market share therefore says little about whether AWS or Azure individually holds an entrenched position. It joins the shares of two firms that compete intensely with each other, while Google Cloud and Oracle add further competitive pressure.

This problem also complicates the Commission’s decision to investigate AWS and Azure while excluding Google Cloud.[70] The Commission must identify an objective distinction between Amazon and Microsoft, on one hand, and Google, on the other.

If the distinction rests on user numbers, the designation inquiry returns to size. If it rests on a line between Google’s roughly 14% worldwide share and Microsoft’s roughly 21% worldwide share, the DMA provides no such threshold. Amazon and Microsoft can therefore argue that the Commission has offered no legitimate basis for treating Google differently that does not reduce to the kind of user counting that Article 3(8) was meant to replace when Article 3(2)(b)’s presumptions did not apply.

The General Court rejected an equal-treatment claim in ByteDance. When a service meets the statutory thresholds, designation is mandatory, and one undertaking cannot challenge its own designation by pointing to the Commission’s treatment of another.[71] That reasoning fits Article 3(2), where the thresholds dictate the result. It does not, however, transfer easily to Article 3(8).

Subthreshold designation under Article 3(8) depends on Commission discretion. The Commission chooses which services to investigate and which to designate. Equal-treatment principles carry particular force when an institution exercises that kind of discretion. Comparable situations may not be treated differently unless an objective justification supports the distinction.[72]

Article 3(8) therefore reopens the issue that ByteDance closed. Because subthreshold designation results from an exercise of discretion rather than the automatic application of numerical thresholds, the Commission’s choice of whom to designate—and whom to exclude—must withstand equal-treatment review.

The point is not that Google Cloud should also be designated. The exclusion instead reveals the criteria driving the Commission’s decisions. If the line between designation and non-designation rests on relative user counts below the statutory thresholds, then the Commission is relying on an unstated measure of size. The numbers are not high enough to trigger Article 3(2), yet the Commission treats them as high enough under Article 3(8). That approach replaces the DMA’s express quantitative thresholds with an undisclosed set of lower ones.

D. Foreseeable Entrenchment Requires Evidence

The European Commission may argue that the competitive problem is forward-looking. Article 3(1)(c) permits designation when it is “foreseeable” that an undertaking will attain an entrenched and durable position “in the near future.” The Commission’s public framing points in that direction. It describes cloud computing as “a prerequisite for AI,” suggests that AI tools and partnerships influence procurement decisions, and presents the investigations in terms of European technological sovereignty.[73]

On that account, the Commission could concede that AWS and Microsoft Azure face meaningful competition today while predicting that AI workloads, “data gravity,” and sovereign-cloud requirements will entrench them tomorrow. Data gravity refers to the tendency of large datasets to attract related applications, services, and computing resources because moving the data becomes costly or impractical.

But a prediction of foreseeable entrenchment remains a claim about market contestability, even when it looks ahead. It must therefore satisfy the reasoning standard the General Court applied in Meta Platforms. The Court annulled the Facebook Marketplace designation on two grounds. First, the legality of a designation had to be assessed “on the basis of the facts and the law as they stood at the time when the measure was adopted.” Second, the Commission’s treatment of a material factual change was “vague and hypothetical” and lacked any “specific analysis.”[74]

The Court applied the ordinary standard governing statements of reasons under European Union law:

[T]he statement of reasons required by … Article 296 TFEU … must disclose in a clear and unequivocal fashion the reasoning followed by the institution which adopted the measure … to enable the persons concerned to ascertain the reasons for it and to enable the court having jurisdiction to exercise its power of review.[75]

By the time of the judgment, the Commission had already removed Marketplace’s designation after its user numbers fell below the thresholds. The ruling nevertheless matters because of its ex tunc effect and the evidentiary standard it establishes.[76]

A finding that cloud services are large and AI will lock in their positions would be conclusory under Meta Platforms. To survive review, the Commission would need to identify the mechanism by which AI demand will convert today’s contested positions into entrenched ones. It would also need to define the relevant time horizon, because the “near future” cannot mean an open-ended period, and provide evidence available at the time of adoption showing that competition will cease to constrain AWS and Azure.

The current AI record points in the opposite direction. CERRE’s 2025 assessments conclude that competition in cloud computing and AI remains strong, both in present conditions and over time, and caution against broad ex ante intervention.[77] Model developers commonly use several cloud providers, and the leading cloud–AI partnerships have generally remained nonexclusive.[78]

The CMA’s work on foundation models and the Autorité de la concurrence’s opinion on generative AI identify access to computing capacity as a possible future concern, while documenting substantial current competition.[79] The FTC’s Section 6(b) study of cloud–AI partnerships raises questions about spending commitments and vertical relationships, but it offers no evidence that cloud-infrastructure markets are tipping.[80]

Entry also weakens the claim that AI will deepen incumbent moats. AI demand has helped Oracle gain share and has supported the emergence of specialized “neocloud” providers, which now account for roughly 5% of the market and are expected to continue growing rapidly.[81] Cloud computing has also lowered barriers to AI adoption by giving startups access to advanced models and computing resources through services such as Amazon Bedrock and Azure OpenAI Service.[82] The development the Commission portrays as a future source of entrenchment has so far expanded access and encouraged entry.

The broader political context also warrants attention. The Commission proposed the Cloud and AI Development Act (CADA) at the same time as the preliminary designations. CADA is an industrial-policy measure expressly motivated by the decline in EU providers’ market share from roughly 29% in 2017 to 15% in 2022. Read alongside CADA and Mario Draghi’s account of European dependence on non-European computing capacity, the cloud investigations appear partly connected to technological-sovereignty goals.[83]

Those goals may support separate industrial-policy measures. They do not, however, alter the legal test under the DMA. Dependence on foreign providers does not make a company a gatekeeper, and technological sovereignty cannot substitute for evidence that AWS or Azure holds, or will soon hold, an entrenched and durable position.

IV. Proportionality and the Available Alternatives

The preceding section assessed whether the proposed designations satisfy Article 3(1). This section asks a separate question. Even if the European Commission could establish those conditions, is designation a proportionate response to the competitive frictions documented in the cloud market?

Two comparisons guide the analysis. Section IV.A examines the United Kingdom, where competition authorities reviewed the same firms on an extensive evidentiary record and chose voluntary commitments rather than designation under the country’s comparable ex ante regime. Section IV.B turns to the European Union’s Data Act, a purpose-built switching regime whose central prohibition has not yet taken effect. It also considers the Commission’s parallel inquiry into whether the DMA’s existing obligations can address cloud-specific concerns.

Both comparisons point toward narrower measures. Designation is the most intrusive option available, yet the Commission is pursuing it before testing the tailored remedies already adopted or determining whether the obligations it would trigger fit the conduct at issue.

A. The United Kingdom Chooses a Lighter Touch

The United Kingdom examined the same market and the same two firms under a regime whose gatekeeper logic closely resembles that of the DMA. It reached a different conclusion.

In 2023, Ofcom referred cloud-infrastructure services to the CMA. Ofcom found that AWS and Microsoft held a combined U.K. market share of 70% to 80%. It identified egress fees, technical-interoperability barriers, and committed-spend discounts as impediments to switching, while also documenting strong competition for new customers.[84]

The CMA completed its market investigation in July 2025. It concluded that AWS and Microsoft each held significant and entrenched unilateral market power, protected by high barriers to entry. The CMA also identified adverse effects on competition arising from switching frictions and Microsoft’s software-licensing practices. Its inquiry group recommended that the CMA Board prioritize Strategic Market Status (SMS) investigations into both companies’ cloud businesses under the Digital Markets, Competition and Consumers Act 2024 (DMCCA).[85]

The Board reviewed the same record and declined to follow that recommendation. In March 2026, it accepted a package of voluntary commitments addressing egress and interoperability, which the CMA cannot itself enforce, including the extension of European Union Data Act switching standards to U.K. customers. The CMA said it would review progress in six months’ time.

The Board directed the concern it considered serious enough for possible designation—Microsoft’s software-licensing practices—into an SMS investigation of Microsoft’s business-software offerings rather than cloud infrastructure itself.[86]

The decision drew criticism. Practitioners argued that the Board had departed from the inquiry group’s recommendation without offering a fully persuasive explanation, and coalitions representing challenger providers objected strongly.[87] Nor did the Board foreclose a future SMS investigation into cloud services.

Still, the authority with the most extensive evidentiary record assembled on AWS and Microsoft concluded that voluntary commitments offered a proportionate response in a market that remains competitive and continues to change. The decision reflects the risk that broad intervention in a market central to AI-driven growth could condemn competitive conduct or impede investment. Negotiated remedies, by contrast, can address identified practices while reducing that risk.

The European Commission is pursuing a markedly different course under the DMA, although the comparison is not exact. The DMCCA requires findings of substantial and entrenched market power and strategic significance, while the DMA formally disclaims a market-power test.[88]

Yet the qualitative routes under the two regimes converge on the same economic question, as Section III.B explains. Both require an assessment of contestability. The U.K. authority made an express finding of market power but still chose targeted commitments. The Commission proposes broader intervention on a thinner basis.

The proposed designations would be the first to cover services that satisfy neither user threshold. They would also subject AWS and Azure to the DMA’s full set of obligations in a market that the CMA recently concluded could be addressed through narrower, voluntary measures.

B. The Data Act and the Proportionality Problem

Proportionality presents a final obstacle to designation. The European Union has already enacted a purpose-built regime for the specific frictions documented in the cloud market.

Chapter VI of the Data Act applies to all providers of “data processing services,” rather than only designated gatekeepers. It requires providers to remove commercial, technical, contractual, and organizational barriers to switching. The law caps switching charges, including data-egress fees, at cost during a transitional period and prohibits them entirely beginning Jan. 12, 2027. It also requires functional equivalence and open interoperability specifications, supported by European Commission-endorsed standard contractual clauses.[89]

Scholars have examined the economics and legal structure of this regime in detail.[90] The national competition authorities that studied cloud services most closely also identified the Data Act as the appropriate remedial instrument. In 2023, France’s Autorité de la concurrence concluded that existing competition-law tools and the proposed Data Act could address most of the risks it identified. The Netherlands Authority for Consumers and Markets likewise pointed to the Data Act as the proper means of addressing cloud lock-in.[91] As Section IV.A explains, the United Kingdom’s remedial package also incorporates Data Act switching standards.

Against that background, several commentators argue that the Commission should explain why the tailored regime is inadequate before invoking the DMA. That requirement would follow from proportionality and the need for coherence across the European Union’s digital laws.[92] Designation would impose the DMA’s broadest obligations on services already governed by a specialized switching regime whose central prohibition has not yet taken effect.

The Commission appears to recognize the mismatch. On the same day it opened the designation investigations, it launched a separate investigation under Article 19 of the DMA into whether the regulation’s existing obligations can address the practices identified in cloud services. Those practices include interoperability barriers, restrictions on data access, tying and bundling, and imbalanced contract terms.

The Commission must issue its Article 19 report within 18 months. The report could support delegated acts under Articles 12 and 49.[93] It is expected around May 2027, six months after the November 2026 deadline for the designation decisions.

The sequence is backward. The Commission proposes to designate AWS and Azure before determining whether the obligations triggered by designation fit the cloud industry.[94] Commentators have already observed that none of the DMA’s current obligations directly addresses the concerns documented in cloud markets.

Designation without suitable obligations would impose substantial compliance costs while producing little improvement in contestability. A proportionate response should connect the identified harm to a remedy capable of addressing it. The Commission’s parallel investigation suggests that this connection has not yet been established.

The Commission’s first DMA review report, published in April 2026, presents the cloud investigations as evidence that the regulation’s “future-proofing tools” are working and concludes that no legislative amendment is needed.[95] But the investigations cannot validate the DMA’s flexibility before the Commission has established that designation is lawful and that the resulting obligations are appropriate. Treating the proposed designations as proof of the framework’s success assumes the very proposition the Commission must still demonstrate.

V. Conclusion: When Size Becomes Gatekeeping

The arguments developed here do not guarantee that the proposed cloud designations will fail. Article 3 permits two competing interpretations. Under the narrower reading, the qualitative criteria filter out firms that are large but do not control genuine gateways. Under the broader reading, a large company that provides a CPS may qualify as a gatekeeper largely because the DMA assumes that size, durability, and integration in digital markets tend to produce gatekeeper power.

The General Court’s first two DMA judgments suggest how the courts may police that boundary. ByteDance and Meta both involved the statutory presumptions. The Commission prevailed when the numerical thresholds supported the qualitative criteria.[96] When the Court annulled the Facebook Marketplace designation, it focused on the quality of the Commission’s reasoning. It required the Commission to assess the facts as they existed when the decision was adopted and to provide a concrete analysis of material developments.[97]

The courts need not import Article 102 TFEU’s market-power test into the DMA. They can protect the same economic substance by requiring the Commission to prove each element of a subthreshold designation separately. The pending ByteDance appeal may further define how closely the courts will review that reasoning.[98]

The AWS and Azure investigations therefore present a consequential test. In every previous designation, a quantitative presumption largely carried the substantive inquiry. Article 3(8) offers no such shortcut. The Commission must prove that each cloud service is an important gateway under Article 3(1)(b), and that the undertaking enjoys, or foreseeably will enjoy, an entrenched and durable position in its operations under Article 3(1)(c).[99]

The current record makes that task difficult. Quality-adjusted prices have fallen, market shares have shifted, multihoming is widespread, providers have reduced switching costs, and new entrants are capturing demand associated with AI. The Commission has also excluded Google Cloud without explaining the distinction, while relying at times on the combined position of two firms that compete against each other. Meanwhile, the Data Act’s tailored switching rules have not yet taken full effect, and the United Kingdom chose voluntary commitments after examining the same market on a fuller record.

If those facts nevertheless suffice for designation, the DMA’s qualitative criteria will have done little more than repackage size and commercial importance. “Gateway” will mean indispensability as an input rather than intermediation between users. “Entrenched and durable” will describe scale rather than limited contestability. Forward-looking claims about AI and data gravity will substitute for evidence that rivalry is likely to weaken.

The alternative is to give each statutory condition independent force. “Gateway” should require intermediation, not mere criticality. “Entrenched and durable” should require evidence that rivals cannot discipline the incumbent. Predictions about future entrenchment should satisfy Article 296’s demand for specific and contemporaneous reasoning. Proportionality should also require the Commission to explain why the Data Act and targeted enforcement cannot address the documented concerns.

Under that reading, the Commission should not convert its preliminary position into a final designation. If it does, the courts should require the evidence and reasoning that Article 3(8) demands.

[1] Press Release, Eur. Comm’n, Commission Reaches Preliminary Position that Amazon’s and Microsoft’s Market-Leading Cloud Services Should Be Designated Under the Digital Markets Act (June 25, 2026), https://ec.europa.eu/commission/presscorner/detail/en/ip_26_1444 [hereinafter Preliminary Position]; see Regulation (EU) 2022/1925 of the European Parliament and of the Council of 14 September 2022 on Contestable and Fair Markets in the Digital Sector and Amending Directives (EU) 2019/1937 and (EU) 2020/1828 (Digital Markets Act), 2022 O.J. (L 265) 1 [hereinafter DMA].

[2] Eur. Comm’n, Commission Launches Market Investigations on Cloud Computing Services Under the Digital Markets Act (Nov. 18, 2025), https://digital-markets-act.ec.europa.eu/commission-launches-market-investigations-cloud-computing-services-under-digital-markets-act-2025-11-18_en. The Commission simultaneously opened a third investigation under Article 19 of the DMA to determine whether the regulation’s existing obligations can effectively address practices in the cloud sector. See infra Section IV.B.

[3] Preliminary Position, supra note 1 (reporting that final decisions are due by November 2026 under the 12-month deadline for market investigations); DMA, supra note 1, arts. 3(10), 17 (providing a six-month compliance window following designation).

[4] Commission Decision of 29.4.2024, Apple—iPadOS, Case DMA.100047, https://ec.europa.eu/competition/digital_markets_act/cases/202427/DMA_100047_5491.pdf [hereinafter iPadOS].

[5] Id.; see also Press Release, Eur. Comm’n, Commission Designates Apple’s iPadOS Under the Digital Markets Act (Apr. 29, 2024), https://ec.europa.eu/commission/presscorner/detail/en/ip_24_2363 (“Apple’s business user numbers exceeded the quantitative threshold elevenfold, while its end user numbers were close to the threshold and are predicted to rise in the near future.”).

[6] Lazar Radic, Cloudy Logic: The DMA’s Search for a Gatekeeper, Truth on the Mkt. (Apr. 14, 2026), https://truthonthemarket.com/2026/04/14/cloudy-logic-the-dmas-search-for-a-gatekeeper.

[7] DMA, supra note 1, art. 3(1). The three criteria and their accompanying presumptions are discussed infra Section II.A.

[8] DMA, supra note 1, art. 3(1)(a)–(c).

[9] Id. art. 3(2)(a)–(c).

[10] Case T-1077/23, ByteDance Ltd. v. Comm’n, ECLI:EU:T:2024:478, ¶ 71 (July 17, 2024) [hereinafter ByteDance] (holding that arguments submitted under Article 3(5) must “manifestly call into question” the presumptions—a standard requiring a high degree of plausibility). An appeal is pending before the Court of Justice. See infra note 98.

[11] DMA, supra note 1, arts. 3(8), 17.

[12] Friso Bostoen & Giorgio Monti, The Rhyme and Reason of Gatekeeper Designation Under the Digital Markets Act, J. Antitrust Enf’t (advance article 2025), jnae054, https://doi.org/10.1093/jaenfo/jnae054 [hereinafter Bostoen & Monti].

[13] DMA, supra note 1, recital 23 (authorizing designation based on a qualitative assessment conducted through a market investigation).

[14] Cf. Bostoen & Monti, supra note 12 (documenting the evidence used to assess each criterion under Article 3(1)).

[15] Jacques Crémer, Yves-Alexandre de Montjoye & Heike Schweitzer, Competition Policy for the Digital Era (2019) (report prepared for the Eur. Comm’n); Digit. Competition Expert Panel, Unlocking Digital Competition (2019) (U.K.) [hereinafter Furman Review]; Stigler Comm. on Digit. Platforms, Final Report (2019).

[16] Nicolas Petit, The Proposed Digital Markets Act (DMA): A Legal and Policy Review, 12 J. Eur. Competition L. & Prac. 529 (2021).

[17] Luís Cabral et al., The EU Digital Markets Act: A Report from a Panel of Economic Experts (Joint Rsch. Ctr. 2021).

[18] Damien Geradin, What Is a Digital Gatekeeper? Which Platforms Should Be Captured by the EC Proposal for a Digital Markets Act? 20 (Feb. 2021) (unpublished manuscript) (“The methodology used to designate gatekeepers cannot be divorced from the issue(s) that ex ante regulation seeks to address, as otherwise the DMA might end up regulating the wrong set of companies.”).

[19] Pinar Akman, Regulating Competition in Digital Platform Markets: A Critical Assessment of the Framework and Approach of the EU Digital Markets Act, 71 Int’l & Comp. L.Q. 85 (2022).

[20] Pablo Ibáñez Colomo, The Draft Digital Markets Act: A Legal and Institutional Analysis, 12 J. Eur. Competition L. & Prac. 561 (2021).

[21] See, e.g., Giuseppe Colangelo, The European Digital Markets Act and Antitrust Enforcement: A Liaison Dangereuse, 47 Eur. L. Rev. 597 (2022); Viktoria H.S.E. Robertson, The Complementary Nature of the Digital Markets Act and the EU Antitrust Rules, 12 J. Antitrust Enf’t 325 (2024).

[22] See, e.g., Fiona Scott Morton & Cristina Caffarra, The European Commission Digital Markets Act: A Translation, VoxEU (Jan. 5, 2021).

[23] Bostoen & Monti, supra note 12 (surveying the 34 designation decisions adopted at the time of publication and analyzing how each applied the designation criteria, as well as the type I- and type II-error-correction functions of Articles 3(5) and 3(8)).

[24] Eur. Comm’n, Commission Concludes that Online Social Networking Service X Should Not Be Designated Under the Digital Markets Act (Oct. 16, 2024), https://digital-markets-act.ec.europa.eu/commission-concludes-online-social-networking-service-x-should-not-be-designated-under-digital-2024-10-16_en.

[25] Commission Decision of 5.9.2023, Alphabet, Case DMA.100011, https://ec.europa.eu/competition/digital_markets_act/cases/202344/DMA_100011_147.pdf (finding that Gmail was not an important gateway despite meeting the user-number thresholds); Commission Decision of 5.9.2023, Microsoft, Case DMA.100023, https://ec.europa.eu/competition/digital_markets_act/cases/202344/DMA_100023_115.pdf (reaching the same conclusion for Outlook.com).

[26] Press Release, Eur. Comm’n, Commission Closes Market Investigations into Microsoft’s and Apple’s Services Under the Digital Markets Act (Feb. 2024), https://ec.europa.eu/commission/presscorner/detail/en/mex_24_785 (addressing iMessage, Bing, Edge, and Microsoft Advertising).

[27] See Press Release, Eur. Comm’n, Digital Markets Act: Commission Designates Six Gatekeepers, IP/23/4328 (Sept. 6, 2023) (declining to designate Samsung because its browser was not an important gateway); Bostoen & Monti, supra note 12 (discussing the accepted rebuttals concerning TikTok Ads and X Ads).

[28] Autorité de la concurrence, Opinion 23-A-08 of 29 June 2023 on Competition in the Cloud Computing Sector, https://www.autoritedelaconcurrence.fr/en/press-release/cloud-computing-autorite-de-la-concurrence-issues-its-market-study-competition-cloud; Autoriteit Consument & Markt, Market Study Cloud Services (Sept. 5, 2022), https://www.acm.nl/system/files/documents/public-market-study-cloud-services.pdf; Japan Fair Trade Comm’n, Report Regarding Cloud Services (June 2022); OECD, Competition in the Provision of Cloud Computing Services (2025); Press Release, Fed. Trade Comm’n, FTC Seeks Comment on Business Practices of Cloud Computing Providers that Could Impact Competition and Data Security (Mar. 2023).

[29] See generally On-Topic: Cloud and Competition Policy, Concurrences No. 8-2025 (Aug. 2025) (Parts I–VIII, examining regulation in the age of hyperscalers, interoperability, policy responses to competition concerns, and cloud–AI partnerships).

[30] Antonio Manganelli & Daniel Schnurr, Competition and Regulation of Cloud Computing Services: Economic Analysis and Review of EU Policies (Ctr. on Regul. in Eur., Feb. 2024), https://cerre.eu/publications/competition-and-regulation-of-cloud-computing-services-economic-analysis-and-reviewbrof-eu-policies [hereinafter CERRE 2024]; see especially id. Recommendation 2 (urging proportionality when designating cloud services under the DMA because they differ from other core platform services).

[31] Konstantina Bania & Damien Geradin, The Regulation of Cloud Computing: Why the European Union Failed to Get It Right, 33 Info. & Commc’ns Tech. L. 1 (2024), https://doi.org/10.1080/13600834.2023.2260687.

[32] Selçukhan Ünekbas, Why Use the Digital Markets Act for Cloud?, Oxford Bus. L. Blog (May 4, 2026), https://blogs.law.ox.ac.uk/oblb/blog-post/2026/05/why-use-digital-markets-act-cloud.

[33] Kalpana Tyagi, Can Europe’s Digital Markets Act and Data Act Rein in Cloud Hyperscalers?, Tech Pol’y Press (Feb. 3, 2026), https://www.techpolicy.press/can-europes-digital-markets-act-and-data-act-rein-in-cloud-hyperscalers.

[34] Vanderbilt Pol’y Accelerator, How to Regulate the Cloud: A Blueprint to Address the Market Failures and National Security Risks of Cloud Computing (2025), https://cdn.vanderbilt.edu/vu-URL/wp-content/uploads/sites/412/2025/09/18140135/How-to-Regulate-the-Cloud.pdf; cf. Tejas N. Narechania & Ganesh Sitaraman, An Antimonopoly Approach to Governing Artificial Intelligence, 43 Yale L. & Pol’y Rev. 95 (2024) (treating cloud computing as an infrastructure layer that warrants regulation).

[35] See, e.g., The EU Commission Preliminarily Finds that 2 Leading Cloud Computing Providers Should Be Designated as Gatekeepers Under the DMA Despite Falling Below the Quantitative Thresholds (Amazon/Microsoft), e-Competitions (June 2026); Selçukhan Ünekbas, The EU’s Facebook Marketplace Decision: The Gatekeeper That Wasn’t, Truth on the Mkt. (June 4, 2026), https://laweconcenter.org/resources/the-eus-facebook-marketplace-decision-the-gatekeeper-that-wasnt (drawing implications from the Meta ruling for qualitative designations).

[36] DMA, supra note 1, art. 2(2)(i).

[37] Id. recital 14.

[38] Bania & Geradin, supra note 31 (arguing that the obligations imposed by the Digital Markets Act, Digital Services Act, and Data Act fit poorly with the technical and commercial characteristics of cloud services).

[39] Tyagi, supra note 33 (reconstructing the impact assessment’s ecosystem rationale for including cloud services and arguing that qualitative designation is better suited to innovation-intensive, business-facing markets).

[40] Preliminary Position, supra note 1.

[41] ByteDance, supra note 10, ¶ 210.

[42] Jean-Charles Rochet & Jean Tirole, Platform Competition in Two-Sided Markets, 1 J. Eur. Econ. Ass’n 990 (2003); Mark Armstrong, Competition in Two-Sided Markets, 37 RAND J. Econ. 668 (2006).

[43] Even these canonical platform markets do not fit neatly within the two-sided-market framework. See Dirk Auer & Nicolas Petit, Two-Sided Markets and the Challenge of Turning Economic Theory into Antitrust Policy, 60 Antitrust Bull. 426 (2015).

[44] Preliminary Position, supra note 1; see also Eur. Comm’n, Commission Reaches Preliminary Position that Amazon’s and Microsoft’s Market-Leading Cloud Services Should Be Designated Under the Digital Markets Act, Digit. Strategy (June 25, 2026), https://digital-strategy.ec.europa.eu/en/news/commission-reaches-preliminary-position-amazons-and-microsofts-market-leading-cloud-services-should (quoting Executive Vice Presidents Teresa Ribera and Henna Virkkunen).

[45] The firms facing designation make nearly identical arguments. See Amazon, Why Applying the DMA to Cloud Would Regulate Away EU Competitiveness and Resiliency (June 2026), https://www.aboutamazon.eu/news/policy/why-applying-the-DMA-to-cloud-would-regulate-away-eu-competitiveness-and-resiliency (arguing that cloud services involve no intermediation, cross-side network effects, or gateway function). We cite the paper as a party submission, not as independent authority.

[46] See Eur. Comm’n, Commission Concludes that Online Social Networking Service X Should Not Be Designated Under the Digital Markets Act, supra note 24.

[47] See Commission Decision of 5.9.2023, Alphabet, supra note 25; Commission Decision of 5.9.2023, Microsoft, supra note 25.

[48] See Eur. Comm’n, Commission Closes Market Investigations into Microsoft’s and Apple’s Services Under the Digital Markets Act, supra note 26.

[49] iPadOS, supra note 4; see also Bostoen & Monti, supra note 12 (analyzing the decision’s reasoning and reliance on projected end-user growth).

[50] Crémer, de Montjoye & Schweitzer, supra note 15; Furman Review, supra note 15.

[51] ByteDance, supra note 10, paras 45-46 (rejecting arguments that designation requires market definition or proof of market power).

[52] See Robertson, supra note 21; Georgios Gryllos, The New Digital Landscape: Interaction Between the DMA and Rules of National and EU Law Governing the Conduct of Gatekeepers, Concurrences No. 1-2024, Art. No. 116827, at 40 (2024).

[53] ByteDance, supra note 10, ¶¶ 296–297.

[54] William J. Baumol, Contestable Markets: An Uprising in the Theory of Industry Structure, 72 Am. Econ. Rev. 1 (1982); William J. Baumol, John C. Panzar & Robert D. Willig, Contestable Markets and the Theory of Industry Structure (1982).

[55] Chad Syverson, Macroeconomics and Market Power: Context, Implications, and Open Questions, 33 J. Econ. Persps. 23 (2019).

[56] Synergy Rsch. Grp., Cloud Market Annual Revenue Run Rate Topped Half a Trillion Dollars in Q1 as Growth Surge Continues (Apr. 2026), https://www.srgresearch.com/articles/cloud-market-annual-revenue-run-rate-topped-half-a-trillion-dollars-in-q1-as-growth-surge-continues (reporting first-quarter 2026 market shares of 28% for Amazon Web Services, 21% for Microsoft, and 14% for Google, with neoclouds accounting for 5% of the market and five providers ranking among the top 30); Synergy Rsch. Grp., Cloud Market Share Trends: Big Three Together Hold 63%, While Oracle and the Neoclouds Inch Higher (2025), https://www.srgresearch.com/articles/cloud-market-share-trends-big-three-together-hold-63-while-oracle-and-the-neoclouds-inch-higher (reporting that Amazon Web Services’ share declined from roughly 32% in 2021).

[57] David Byrne, Carol Corrado & Daniel E. Sichel, The Rise of Cloud Computing: Minding Your P’s, Q’s and K’s (Nat’l Bureau of Econ. Rsch., Working Paper No. 25188, 2018), https://www.nber.org/papers/w25188, reprinted in Measuring and Accounting for Innovation in the Twenty-First Century (Carol Corrado et al. eds., 2020).

[58] See Wasabi, Cloud Storage Fee Inflation, Wasabi Blog (Sept. 29, 2022), https://wasabi.com/blog/cost-optimization/cloud-storage-fee-inflation (reporting that Amazon S3 list prices fell by more than 80% in the early 2010s).

[59] Paul Klemperer, Competition When Consumers Have Switching Costs: An Overview with Applications to Industrial Organization, Macroeconomics, and International Trade, 62 Rev. Econ. Stud. 515 (1995); Joseph Farrell & Paul Klemperer, Coordination and Lock-In: Competition with Switching Costs and Network Effects, in 3 Handbook of Industrial Organization 1967 (Mark Armstrong & Robert Porter eds., 2007).

[60] Ofcom, Cloud Services Market Study: Final Report (Oct. 5, 2023), https://www.ofcom.org.uk/siteassets/resources/documents/consultations/category-3-4-weeks/244808-cloud-services-market-study/associated-documents/cloud-services-market-study-final-report.pdf [hereinafter Ofcom Report] (documenting strong competition for new customers); Competition & Mkts. Auth., Cloud Services Market Investigation: Summary of Final Decision (July 31, 2025), https://assets.publishing.service.gov.uk/media/688b20e6ff8c05468cb7b120/summary_of_final_decision.pdf [hereinafter CMA Final Decision].

[61] See Google Cloud, Eliminating Data Transfer Fees When Migrating Off Google Cloud (Jan. 2024), https://cloud.google.com/blog/products/networking/eliminating-data-transfer-fees-when-migrating-off-google-cloud; Amazon Web Servs., Free Data Transfer Out to Internet When Moving Out of AWS (Mar. 2024), https://aws.amazon.com/blogs/aws/free-data-transfer-out-to-internet-when-moving-out-of-aws; Microsoft Azure, Now Available: Free Data Transfer Out to Internet When Leaving Azure (Mar. 2024), https://azure.microsoft.com/en-us/updates?id=now-available-free-data-transfer-out-to-internet-when-leaving-azure.

[62] Regulation (EU) 2023/2854 of the European Parliament and of the Council of 13 December 2023 on Harmonised Rules on Fair Access to and Use of Data (Data Act), 2023 O.J. (L 2854) 1, arts. 23–31 [hereinafter Data Act]. For a more detailed discussion, see infra Section IV.B.

[63] Flexera, State of the Cloud Report (2026 ed.), https://info.flexera.com/CM-REPORT-State-of-the-Cloud

[64] CMA Final Decision, supra note 60; Ofcom Report, supra note 60 (both distinguishing multicloud architectures from workload-level switching).

[65] CMA Final Decision, supra note 60 (finding that Amazon Web Services and Microsoft each held U.K. market shares by value of up to 30%–40% and recommending that the CMA Board prioritize Strategic Market Status investigations into both providers’ cloud activities).

[66] Ofcom Report, supra note 60.

[67] CMA Final Decision, supra note 60 (finding materially higher effective prices for running Windows Server and SQL Server on rival cloud platforms than on Azure); see also CISPE, CISPE and Microsoft Agree Settlement in Fair Software Licensing Case (July 11, 2024), https://www.cispe.cloud/cispe-and-microsoft-agree-settlement-in-fair-software-licensing-case (settling the association’s 2022 complaint on terms that excluded Amazon Web Services, Google, and Alibaba).

[68] See Google Files EU Antitrust Complaint Accusing Microsoft of Stifling Cloud Competition, CNBC (Sept. 25, 2024), https://www.cnbc.com/2024/09/25/google-files-eu-antitrust-complaint-accusing-microsoft-of-stifling-cloud-competition.html. Google withdrew the complaint on Nov. 28, 2025, after the European Commission opened its cloud-market investigations. See Caroline Donnelly, Google Cloud Withdraws Complaint with European Commission over Microsoft’s Cloud Licensing Tactics, Computer Weekly (Dec. 1, 2025), https://www.computerweekly.com/news/366635494/Google-Cloud-withdraws-complaint-with-European-Commission-over-Microsofts-cloud-licensing-tactics.

[69] DMA, supra note 1, art. 3(9); Case T-1078/23, Meta Platforms, Inc. v. Comm’n, ECLI:EU:T:2026:357, paras 64 and 174  (Gen. Ct. June 3, 2026) [hereinafter Meta] (upholding Messenger’s designation as a distinct core platform service that individually constitutes an important gateway).

[70] See European Commission Lines Up Amazon and Microsoft for Cloud Gatekeeper Status, The Register (June 25, 2026), https://www.theregister.com/legal/2026/06/25/european-commission-lines-up-amazon-and-microsoft-for-cloud-gatekeeper-status/5262127; cf. Article 19, Investigations into Cloud Computing Under DMA a Welcome Move (Nov. 18, 2025), https://www.article19.org/resources/europe-investigations-into-cloud-computing-under-dma-a-welcome-move (noting, from a pro-designation perspective, the puzzle posed by Google Cloud’s exclusion).

[71] ByteDance, supra note 10 (rejecting the claim that the designation breached the principle of equal treatment).

[72] See, e.g., Case C-127/07, Arcelor Atlantique et Lorraine, ECLI:EU:C:2008:728, ¶ 23 (holding that comparable situations must not be treated differently, and different situations must not be treated alike, unless the difference is objectively justified).

[73] Preliminary Position, supra note 1 (describing cloud computing as “a prerequisite for AI”); see also sources cited supra note 45.

[74] Meta, supra note 69; see also Ct. Just. Eur. Union, Press Release No. 77/26 (June 3, 2026), https://curia.europa.eu/site/upload/docs/application/pdf/2026-06/cp260077en.pdf.

[75] Meta, supra note 69, ¶ 53.

[76] See Alba Ribera Martínez, The Chicken or Egg Dilemma: The General Court Partially Annuls the European Commission’s DMA Designation Decision Against Meta (Case T-1078/23), Kluwer Competition L. Blog (June 8, 2026), https://legalblogs.wolterskluwer.com/competition-blog/the-chicken-or-egg-dilemma-the-general-court-partially-annuls-the-european-commissions-dma-designation-decision-against-meta-case-t-107823 (discussing the legal-interest analysis and the probative-effort question).

[77] Zach Meyers & Marc Bourreau, A Competition Policy for Cloud and AI (Ctr. on Regul. in Eur., June 2025), https://cerre.eu/wp-content/uploads/2025/06/A-Competition-Policy-for-Cloud-and-AI_FINAL.pdf; Zach Meyers & Marc Bourreau, What Policy Interventions for a Competitive AI Sector? (Ctr. on Regul. in Eur., July 2025), https://cerre.eu/publications/what-policy-interventions-for-a-competitive-ai-sector.

[78] Christophe Carugati, The Competitive Relationship Between Cloud Computing and Generative AI (Bruegel, Working Paper No. 19/2023), https://www.bruegel.org/system/files/2023-12/WP%202023%2019%20Cloud%20111223.pdf (mapping largely nonexclusive cloud–AI partnerships and multihoming by model developers).

[79] Competition & Mkts. Auth., AI Foundation Models: Initial Report (Sept. 2023); Competition & Mkts. Auth., AI Foundation Models: Update Paper (Apr. 2024); Autorité de la concurrence, Opinion 24-A-05 of 28 June 2024 on the Competitive Functioning of the Generative Artificial Intelligence Sector.

[80] Fed. Trade Comm’n, Partnerships Between Cloud Service Providers and AI Developers: FTC Staff Report on AI Partnerships & Investments 6(b) Study (Jan. 2025), https://www.ftc.gov/system/files/ftc_gov/pdf/p246201_aipartnerships6breport_redacted_0.pdf.

[81] Synergy Rsch. Grp., Neocloud Market Forecast to Approach $400B by 2031, Driven by Surging AI Infrastructure Demand (2026), https://www.srgresearch.com/articles/neocloud-market-forecast-to-approach-400b-by-2031-driven-by-surging-ai-infrastructure-demand; see also sources cited supra note 56.

[82] Comments of the Int’l Ctr. for L. & Econ., In re Cloud Computing Request for Information, Fed. Trade Comm’n (June 2023), https://laweconcenter.org/resources/icle-response-to-the-ftcs-cloud-computing-rfi.

[83] Eur. Comm’n, Proposal for a Regulation Establishing a Framework of Measures for Strengthening Europe’s Cloud and AI Ecosystem (Cloud and AI Development Act), COM (2026) 502 final (June 3, 2026), https://digital-strategy.ec.europa.eu/en/policies/cloud-and-ai-development-act (noting that EU providers’ market share fell from roughly 29% in 2017 to 15% in 2022); Mario Draghi, The Future of European Competitiveness (Sept. 2024).

[84] Ofcom Report, supra note 60 (reporting a combined U.K. market share of 70%–80% and identifying frictions from egress fees, technical-interoperability barriers, and committed-spend discounts).

[85] CMA Final Decision, supra note 60.

[86] Press Release, Competition & Mkts. Auth., CMA Announces Package of Actions on Business Software and Cloud Services (Mar. 31, 2026), https://www.gov.uk/government/news/cma-announces-package-of-actions-on-business-software-and-cloud-services. The Strategic Market Status investigation into Microsoft’s business-software ecosystem began in May 2026, with a designation decision expected around February 2027.

[87] Christophe Humpe & Greg Dowell, CMA Decision to Shelve Cloud Services SMS Investigations Raises Questions, Macfarlanes (Apr. 2026), https://www.macfarlanes.com/insights/102mqfd/cma-decision-to-shelve-cloud-services-sms-investigations-raises-questions-around; see also Antony Adshead, CMA to Launch Strategic Market Status Investigation into Microsoft; Amazon Web Services Off the Hook, Computer Weekly (Mar. 2026), https://www.computerweekly.com/news/366640828/CMA-to-launch-strategic-market-status-investigation-into-Microsoft-Amazon-Web-Services-off-the-hook (collecting reactions from challenger providers).

[88] Digital Markets, Competition and Consumers Act 2024, c. 13, §§ 2, 5–6 (UK).

[89] Data Act, supra note 62, ch. VI, arts. 23–31, 34–35; see also Eur. Comm’n, Data Act Explained, Digit. Strategy, https://digital-strategy.ec.europa.eu/en/factpages/data-act-explained (explaining that switching charges are capped at cost during a transitional period and prohibited beginning Jan. 12, 2027, and that standard contractual clauses were published in 2025).

[90] Daniel Schnurr, Switching and Interoperability Between Data Processing Services in the Proposed Data Act, in Data Act: Towards a Balanced EU Regulation (Ctr. on Regul. in Eur. 2023), https://cerre.eu/wp-content/uploads/2023/03/230327_Data-Act-Book.pdf; Leonie Ott & Yifeng Dong, Clouds Connecting Europe: Interoperability in the EU Data Act, JIPITEC (2025), https://www.jipitec.eu/jipitec/article/download/435/434/2202.

[91] Autorité de la concurrence, Opinion 23-A-08, supra note 28; Autoriteit Consument & Markt, supra note 28.

[92] Ünekbas, supra note 32; see also Consolidated Version of the Treaty on European Union art. 5(4), 2012 O.J. (C 326) 13 (requiring proportionality in Union action).

[93] Eur. Comm’n, Commission Launches Market Investigations on Cloud Computing Services Under the Digital Markets Act, supra note 2; see also Eur. Comm’n, Commission Hosted Stakeholder Roundtable on Cloud Computing Services Under the Digital Markets Act (July 1, 2026), https://digital-markets-act.ec.europa.eu/commission-hosted-stakeholder-roundtable-cloud-computing-services-under-digital-markets-act-2026-07-01_en (describing the Article 19 investigation as “distinct and separate” from the designation investigations).

[94] See Alba Ribera Martínez, Generative AI in Check: Gatekeeper Power and Policy Under the DMA, Kluwer Competition L. Blog (Nov. 19, 2024), https://legalblogs.wolterskluwer.com/competition-blog/generative-ai-in-check-gatekeeper-power-and-policy-under-the-dma.

[95] Eur. Comm’n, Report on the First Review of the Digital Markets Act, COM (2026) 178 final (Apr. 28, 2026), https://digital-markets-act.ec.europa.eu/system/files/2026-04/DMA%20Review%20Report_COM_2026_178_1_EN.pdf; see also Megan Kirkwood, What the EU’s First Digital Markets Act Review Actually Changes, Tech Pol’y Press (Apr. 30, 2026), https://www.techpolicy.press/what-the-eus-first-digital-markets-act-review-actually-changes.

[96] See ByteDance, supra note 10; Meta, supra note 69 (upholding Messenger’s designation).

[97] Meta, supra note 69.

[98] Case C-627/24 P, ByteDance Ltd v. Comm’n (appeal lodged Sept. 26, 2024; argued May 12, 2026; judgment pending).

[99] Meta, supra note 69, paras 168, 219.

The Federal Communications Commission’s (FCC) 39% broadcast-ownership cap is a rule for a three-network world trying to govern a streaming one. Retiring it makes sense. Retiring it by itself does not.

For four decades, the FCC has barred any company from owning television stations that collectively reach more than 39% of U.S. television households. The rule reflects an older theory of broadcast regulation—one that treats the airwaves as a scarce public resource and promotes competition, localism, and viewpoint diversity through bright-line ownership limits rather than case-by-case review.

That theory once had an intuitive logic. When the modern cap emerged, most Americans got their news and entertainment from stations affiliated with the “Big Three” networks. Limiting any owner’s national footprint could plausibly prevent too much editorial influence from accumulating in too few hands.

That media world has vanished. Broadcasters now compete not only with one another, but also with streaming services, virtual multichannel video programming distributors (vMVPDs), podcasts, and social platforms whose national and global reach dwarfs anything a station group could assemble under the 39% cap. The rule now binds the competitors least able to bear it while leaving their largest rivals untouched.

The FCC has signaled that it intends to repeal the cap and review broadcast consolidation case by case. As a matter of competition policy, that move is overdue. As a matter of law, it is messier. Congress wrote the 39% figure into an appropriations statute, raising the question of whether the FCC may erase it on its own.

The policy debate also cannot stop at ownership. Broadcast regulation operates as an interconnected system. Its other parts include retransmission consent, which governs the terms and fees under which distributors carry broadcast signals; must-carry rules, which can require carriage of qualifying local stations; and FCC standards requiring the parties to negotiate in good faith. Together, these rules divide bargaining power among many of the same companies.

Removing the cap would give larger station groups more leverage over the cable, satellite, and streaming distributors that carry their signals. The retransmission regime was not designed to offset that added power. Repealing the cap while leaving the carriage rules untouched would therefore do less to eliminate a distortion than to move it elsewhere.

The better course is comprehensive reform. Because ownership limits, retransmission consent, and bargaining standards all fall within the FCC’s jurisdiction, the agency should consider them together. Otherwise, repeal may simply reshuffle bargaining power among industry players while consumers keep paying the bill.

Read the full piece here.

Introduction and Summary

The International Center for Law & Economics (ICLE) welcomes the opportunity to respond to the Competition and Markets Authority’s (CMA) consultations on proposed steering conduct requirements (CRs) for Apple and Google.[1] ICLE is a non-profit, non-partisan global research and policy centre that advances evidence-based policy. Its scholars have written extensively on competition in digital markets. Because the two draft CRs are materially identical, these comments address both consultations together.[2]

Two points frame our response. First, the CMA’s proposed intervention rests on designation findings that ICLE has contested in earlier submissions. We do not repeat those arguments here, but the premise remains disputed. The extent of competition between iOS and Android bears directly on both the likely benefits of intervention and the costs of regulatory error.[3]

Second, these comments build on ICLE’s April 2026 response to the CMA’s call for evidence, which examined the fee, design, and security questions the draft CR now seeks to resolve.[4] Where appropriate, we cross-refer to that submission rather than repeat its analysis.

One distinction organises everything that follows: the draft CR combines two fundamentally different interventions.

The first is a steering-rights remedy. Paragraphs 3 to 8 and 14 would remove restrictions on communicating with end users and linking to external transactions. They would also require equal treatment of redirection mechanisms, permit a single neutral interstitial screen, and prohibit discrimination against developers that steer.

The second is a price-control regime. Paragraphs 9 to 13 would cap any steering fee at ‘no higher than’ the output of a forward-looking long-run incremental cost (LRIC) model plus a ‘value’ assessment adjusted to remove the effects of the market position that prompted designation. The regime would be enforced through cost-accounting obligations, a requirement that each firm demonstrate compliance ‘to the satisfaction of the CMA’, and quarterly reporting.[5]

These components require different responses. The steering-rights core can be designed proportionately, and much of it resembles measures already taking effect worldwide—including in the United Kingdom, where Google introduced steering on 30 June 2026.[6] Our concerns are principally matters of calibration. The ‘strictly necessary’ standard for platform safeguards is too restrictive, while the interstitial-screen provisions risk suppressing truthful information that consumers may need to make informed decisions. Section IV develops these points.

The fee provisions are different in kind. They would graft the machinery of utility-rate regulation onto an information-goods platform, relying on pricing principles that cannot produce stable or economically coherent results. No other jurisdiction—not the U.S. courts in Epic Games v. Apple, Japan under the Mobile Software Competition Act, or even the European Union under the Digital Markets Act (DMA)—has adopted anything comparable. Sections II and III explain why the proposed principles are indeterminate where they are defined, undefined where they matter most, and structured to operate as a one-way ratchet towards incremental cost.

The CMA’s provisional proportionality assessment does not cure these defects; it assumes them away. It discounts the platforms’ lost revenue through circular reasoning, treats near-complete pass-through of savings as a given, dismisses ecosystem-wide costs on the ground that firms may simply decline to steer, measures benefits against a no-steering counterfactual that no longer exists, and declares the proposal ‘the least onerous’ option without assessing an obvious graduated alternative. Section IV addresses each flaw.

Section V turns to the dynamic consequences. Compressing returns towards incremental cost would misprice the innovation that has generated much of the consumer welfare associated with these ecosystems.

Section VI sets out our recommendations. In brief, the CMA should separate the two components. It should proceed with an amended steering-rights CR but replace paragraphs 9 to 13 with a ‘fair and reasonable’ obligation assessed ex post against commercial benchmarks. Transparency requirements should support that obligation, together with a defined trigger for further intervention if evidence shows that fee levels have frustrated the remedy.

The stakes extend beyond steering. These proposals are among the first substantive CRs under the Digital Markets, Competition and Consumers Act 2024 (DMCC), and they will help set the pattern for those that follow. Parliament and the CMA presented the DMCC as a flexible, participative, and proportionate regime—a deliberate contrast with the DMA’s prescriptive model.[7] Consultations such as this one will determine whether that promise survives first contact with regulation.

I. The CR Conflates Steering Rights with Price Regulation

Paragraphs 9 to 13 are presented as part of the steering remedy, but they are an instrument of a different kind. A steering right removes obstacles between developers and alternative transaction channels, then leaves competition to determine the resulting terms. The fee provisions instead determine those terms administratively.

Section I.A shows that the provisions use the machinery of utility-rate regulation: LRIC modelling, common-cost allocation, regulated returns, and continuing supervisory oversight. Section I.B explains why the CMA’s frustration rationale supports strong non-price safeguards but not administered fees. On the consultation’s own theory, price regulation is either redundant because steering disciplines fees or a substitute for the competition the remedy is meant to create.

Section I.C turns to the institutional stakes. As one of the first substantive CRs under the DMCC, this measure will shape later interventions. Beginning with principles-based price administration would push the regime towards market ordering rather than market oversight. As Section VII explains, the CMA should instead reserve intervention on fees for evidence that pricing has frustrated an otherwise effective steering remedy.

A. The CR Imports Utility-Rate Regulation

The effect of the CR on steering fees becomes clearest in its mechanics. Those mechanics therefore warrant close attention.

Paragraph 10 provides that any steering fee must be ‘no higher than’ a fee calculated under specified pricing principles. The cost-based principle limits the fee to the costs of services ‘used by a Developer to a material and direct extent in facilitating or supporting a Steered Transaction’. Those costs must be calculated using ‘a forward-looking long-run incremental cost approach’, with ‘an appropriate recovery of common costs’ and ‘a reasonable rate of return as appropriate’.[8]

The CMA’s own explanation is revealing. It acknowledges that the LRIC of serving steered transactions ‘could be very low’ because most app-store operating costs would arise regardless of whether an app steers users elsewhere. On that view, the genuinely incremental cost ‘may be limited to the provision of technology that allows a developer to offer external links’.[9]

Paragraph 12 narrows the recoverable cost base further. It excludes ‘general app discovery, marketing, developer tooling’, as well as in-app-payment security and privacy features ‘not required for Steered Transactions’.[10] The value-based principle then requires the fee to be adjusted to remove the effects of the platform’s ‘substantial and entrenched market power’ and to account for the value that developers contribute across the platform.[11]

The CR backs these substantive constraints with extensive administrative machinery. Paragraph 9(a) places the burden on the firm to demonstrate compliance ‘to the satisfaction of the CMA’. Paragraph 12(d) requires cost-accounting records maintained for that purpose. The CMA also proposes rolling quarterly compliance reports tracking, among other things, ‘total platform charges applied to purchases’.[12]

This is the toolkit of utility-rate regulation—LRIC modelling, common-cost allocation, a regulated rate of return, and continuing administrative supervision of prices—applied to a hyperlink. The consultation makes the lineage explicit by comparing its cost-based methodology with the price controls imposed by the Civil Aviation Authority on Heathrow, Ofcom’s telecoms market reviews, Ofgem’s RIIO framework, and the Office of Rail and Road’s regulation of Network Rail access charges.[13]

That comparison exposes the problem. Those regimes govern bottleneck infrastructure whose owners cannot practically prevent incremental use. Here, the CMA would regulate what a firm may charge developers to route transactions away from its own payment system.

The consultation does not hide the intended direction of travel. It ‘expect[s]’ fees to fall under the CR and treats that result as evidence of success.[14]

B. Preventing Frustration Does Not Require Price Control

The consultation’s central justification deserves a fair hearing. The CMA argues that evidence from its investigation and ‘experience in other jurisdictions’ shows that merely prohibiting restrictions on steering would not suffice. Platforms would retain both the ability and incentive to impose ‘restrictive conditions, complex design requirements, or commercially unattractive terms’ that frustrate the remedy.[15]

The concern is legitimate, and the history of comparable interventions supports it. A bare prohibition on anti-steering rules may invite platforms to recreate the same restriction under another name. The question is what follows from that concern—and, specifically, whether it justifies moving from removing restrictions to administering prices.

The analysis must distinguish between two ways a remedy might be frustrated. A platform might use non-price conduct to recreate friction, or it might charge a fee that makes steering commercially unattractive. Those risks call for different responses. Conflating them allows the consultation to present rate regulation as the natural completion of a steering right.

The draft CR already addresses non-price frustration without regulating prices. Paragraph 4 extends the steering-rights obligations to any action ‘whether direct or indirect, and irrespective of the means by which it is implemented, including contractual terms, technical design requirements, app review processes, enforcement practices or other measures’. Paragraph 14 separately prohibits discrimination against developers that steer.[16] Together, those provisions reach precisely the conduct the CMA identifies.

Pricing terms are different in kind. On this point, the fee provisions invert the CMA’s own theory of the case. The consultation treats steering as a source of competitive discipline. Giving developers access to ‘other, lower cost distribution channels’ should exert ‘downward pressure on wider App Store commission fees’. The CMA likewise identifies the insulation of platform fees from competitive pressure as the problem the remedy should correct.[17]

Under that theory, the fee level is not an input into the remedy. It is the market outcome that an effective steering regime should discipline. Paragraphs 9 to 13 instead determine that outcome administratively.

This leaves a dilemma the consultation does not confront. If steering operates as the competitive constraint the CMA anticipates, the resulting downward pressure on fees makes rate regulation redundant. If steering does not provide that constraint, capping the fee near incremental cost does not create competition; it substitutes an administered price for a market price. In neither case is the price control merely ancillary to the steering right. It displaces the competitive process the remedy is supposed to restore.

The claim that pricing latitude will nullify the remedy is empirical and should be tested. The consultation itself supplies relevant comparisons. Steered-transaction fees of 10 to 20 per cent have coexisted with commercial uptake in Japan and through Google’s global rollout. Google’s UK measures, introduced on 30 June 2026, will soon provide direct domestic evidence.[18]

The CR could reserve further intervention for evidence that fee levels have frustrated steering, as our recommendations propose, rather than presume frustration and regulate prices from the outset.

The CMA’s proposed effectiveness metric underscores the problem. It plans to monitor ‘developer and user uptake of steering’ as a measure of success,[19] effectively treating steering volume as an end in itself. Yet the statutory objectives of fair dealing and open choices concern the terms on which developers and users may transact, not the market share of any particular payment channel.

A developer who receives a genuine, unrestricted choice and elects to retain in-app purchase has not been failed by the remedy. Under the consultation’s metric, however, that choice risks appearing as a regulatory shortfall that demands correction.

C. The CR Risks Turning Oversight into Market Ordering

The distinction between the CR’s two components reflects a broader institutional divide. Market-oversight regimes police the competitive process through effects analysis, efficiency defences, and meaningful review. Market-ordering regimes instead design outcomes directly through ex ante obligations, price regulation, and continuous administrative supervision.[20]

This is more than a question of labels. It concerns who decides the terms of trade. Under oversight, an authority asks whether conduct has harmed competition and, if so, acts to stop it. Under ordering, the authority determines what the terms of dealing should be and supervises the firm’s compliance over time.

The DMA has moved decisively towards market ordering. ICLE has documented the predictable consequences: indeterminate standards interpreted by the enforcer, prolonged compliance disputes, and an institutional structure poorly equipped to identify and correct its own mistakes.[21] The DMCC was designed—and presented—as something different: a participative regime whose interventions would be targeted, proportionate, and evidence-led.

The European Commission’s proceedings over Apple’s fees offer a preview of where paragraphs 9 to 13 may lead. After challenging Apple’s 30 per cent commission and imposing a €500 million fine, the Commission subjected each successive revision to Apple’s fee structure—the Core Technology Fee, the initial acquisition fee, and the tiered Store Services fee—to another round of scrutiny. Years of enforcement have yet to produce a stable outcome for developers, consumers, or the platform.[22]

The draft CR recreates the conditions that generated those disputes: open-textured standards such as ‘fair and reasonable’, ‘material and direct extent’, ‘appropriate recovery’, and an adjustment for ‘market power’; a firm-side burden of proof; and rolling compliance adjudication. It adds one significant aggravating factor: decisions under the DMCC are reviewable only on judicial-review principles, not on the merits.[23]

That limited review strengthens the case for restraint before the CMA reaches for price regulation. A regulator whose pricing judgement cannot readily be corrected on appeal should be correspondingly reluctant to set the price in the first place.

The design of this CR will also travel. As one of the first substantive CRs under the DMCC, it will become a template and reference point for later interventions involving these firms and others. Future requirements will be drafted, negotiated, and defended in its shadow. A precedent for principles-based price-setting established here will not remain neatly confined to steering fees.

Beginning the regime with the most prescriptive attempt at platform-price regulation in any jurisdiction would place the DMCC on the market-ordering path from the outset. As ICLE explained in its comments on the CMA’s draft guidance, the sequencing should run in the opposite direction: begin with simpler interventions whose effects can be observed, then escalate as evidence and experience warrant.[24]

II. The Pricing Principles Cannot Produce a Coherent Fee

The CMA asks whether the proposed pricing principles are appropriate, how cost-plus and value should be defined, and how the two should interact.[25] They are not, for reasons that go to the structure of the framework rather than its wording.

The cost-based principle cannot allocate the joint costs and value of an information-goods platform without arbitrary regulatory judgements. The value-based principle requires a competitive counterfactual that the CMA cannot construct and rejects the benchmarks that might inform it. Applied together, the principles create a one-way ratchet towards incremental cost, reinforced by attribution rules that exclude much of the value platforms continue to provide. If the CMA nevertheless assesses fee levels, cautious benchmarking against comparable services offers the least-bad approach—and the available evidence does not support intervention.

A. Cost-Based Pricing Cannot Capture Platform Value

App stores are paradigmatic information goods. They require high fixed and sunk costs, while the marginal cost of distribution approaches zero. They also rely on shared inputs—security infrastructure, review and curation, developer tools, and anti-fraud systems—that support millions of downstream services and resist clean allocation among them.[26]

As ICLE has previously explained, cost-based rate-setting in this setting presents three difficult choices. A regulator can allocate shared fixed costs across services, confine recovery to incremental costs, or add a return on capital to a selected asset base.[27] Paragraph 10(a) adopts all three at once: LRIC, ‘appropriate’ common-cost recovery, and a ‘reasonable’ rate of return. That combination compounds the indeterminacy rather than resolving it.

Allocating common costs requires assumptions about how to divide expenses among interdependent products and services. No objective economic rule supplies the answer. Restricting recovery to incremental costs avoids that allocation problem, but risks excluding the platform-wide investments needed to create, maintain, and improve the service. Adding a return on capital may appear to correct that undercompensation, but it introduces new disputes over the asset base, the appropriate rate of return, and the period over which investment should be recovered.

Cost-based analysis therefore does not eliminate regulatory judgement. It merely relocates that judgement into the construction of the benchmark.

Paragraph 12 makes the problem concrete. It excludes from the cost base the platform-wide investments that give a steered transaction much of its value, including discovery, marketing, developer tools, and payment-security features ‘not required’ for the steered sale. The apparent logic is that a developer who steers does not use those services to a ‘material and direct extent’ and therefore should not pay for them.

That misunderstands what the developer receives. Steering removes one step from the transaction: payment processing. The developer continues to benefit from the rest of the platform bundle, including worldwide distribution and updates, discovery and merchandising, developer tools and APIs, and the consumer trust created by curation and fraud prevention. The commission was never merely a payment-processing fee. It was the price of that broader bundle, of which payment processing forms only one part.

A steered transaction therefore monetises a customer relationship that the platform helped create and continues to support. ‘Direct use’ accounting cannot divide value that is jointly produced. Paragraph 12 captures only the platform’s narrow co-ordination function at the point of linking a developer and user, while ignoring the wider value of the platform that made the transaction possible.

The exclusions also sit uneasily with the CMA’s own findings. The SMS decisions define the relevant digital activity as an integrated ‘Mobile Platform’ comprising the operating system, native app distribution, and the browser.[28] Paragraph 13(e) of the draft CR likewise recognises value flowing from ‘all elements’ of that platform. Yet the cost-based principle limits recoverable costs to a narrow slice of native app distribution. The result is an asymmetry between the value recognised and the remuneration permitted.

A second asymmetry runs through the cost-plus framework. App-store economics are a joint enterprise. The platform invests in the operating system, distribution, tools, and security. Developers invest in their apps. Together, they create a surplus that must be divided.[29]

How that surplus should be split—whether the platform’s share is 30 per cent, 15 per cent, or some lower figure—is a legitimate question, and benchmarking may help answer it. Paragraph 10 does not ask that question. It examines only the platform’s costs and only the platform’s return, as though the platform were a stand-alone utility rather than one participant in a two-sided exchange. It does not account for developers’ costs, risks, or returns.

A method that addresses surplus division by measuring one side’s costs and capping one side’s return is not a neutral measure of fairness. It decides in advance that the platform’s share should tend towards incremental cost, leaving the residual to the other side.

This is familiar territory. The 9th U.S. Circuit Court of Appeals’ recent suggestion in Epic that a permissible commission might be limited to costs ‘genuinely and reasonably necessary for [Apple’s] coordination of external links’ imports the same mistake.[30] ICLE’s amicus filings explain the consequences. Courts and agencies that undertake to supervise the price of platform access assume a task the U.S. Supreme Court has repeatedly warned they are ill-suited to perform. The district court’s interim answer—a mandated price of zero—illustrates the arbitrariness that follows.[31]

Antitrust doctrine recognises why returns above incremental cost may be necessary: they finance the fixed-cost creation on which innovation depends. That is also a central premise of intellectual-property law.[32]

B. The Value Principle Supplies Discretion, Not a Benchmark

The value-based principle might appear to correct the cost principle’s myopia. As drafted, it cannot. Paragraph 13(d) requires the fee to ‘adjust for the direct and indirect effects of’ the platform’s ‘substantial and entrenched market power’ across all elements of the Mobile Platform. In practice, that requires the firm to calculate the price that would prevail in a competitive market that has never existed, without a prescribed methodology or any benchmark the CMA is prepared to accept. The consultation concedes that the exercise is ‘complex by its nature and open to a degree of subjectivity’.[33]

The CMA’s treatment of benchmarks proves the point. Google proposed using comparators to separate competitive value from any market-power premium. The CMA rejected them all, reasoning that an appropriate comparator must ‘(i) provide similar services to those being regulated, (ii) be exposed to effective competition, and (iii) reflect the two-way value exchange between the participants and the platform’.[34]

Taken seriously, those conditions exclude any observable benchmark. A platform similar enough to satisfy the first and third requirements will, on the CMA’s own premises, display the features that disqualify it under the second. A test that no evidence can satisfy is not a pricing principle. It is discretion.

The authority on which the CMA relies does not support the weight placed upon it. The Competition Appeal Tribunal’s observation in Kent v Apple that Apple’s headline commission appeared ‘arbitrary’ remains under appeal, arose in a damages action under a different legal standard, and ultimately cuts the other way. The Tribunal’s difficulty in identifying the fee that value would justify illustrates why administered value-pricing for information goods is so unstable.[35] A finding that Apple did not set its fee by reference to a demonstrable measure of value does not establish that a regulator can do so.

The obvious analogy is to the ‘fair and reasonable’ rates that courts set for standard-essential patents. But the FRAND analogy breaks down at its foundation. As ICLE has explained, FRAND obligations arise from an ex ante commitment made before lock-in and in exchange for the benefits of standardisation. That hypothetical pre-lock-in negotiation anchors every FRAND methodology.[36] There is no comparable bargain to reconstruct.

Even with such an anchor, FRAND adjudication produces ranges rather than determinate prices. The competing rates in Microsoft v Motorola differed by roughly thirty-fold, while the rate in Optis v Apple increased roughly sevenfold on appeal.[37] Those divergences arise because outcomes depend on methodological choices for which no canonical answer exists: which licences are comparable, how lump-sum payments should be converted into per-unit rates, how real-world agreements should be adjusted for hold-up or hold-out, whether to cross-check against the aggregate royalty burden, and which royalty base to use. Courts disagree even about the underlying conceptual framework.

Paragraph 13(e) compounds the problem by requiring the fee to deduct ‘the value contributed by Developers (in aggregate)’. That instruction misreads two-sided economics. Indirect network effects—the value each side receives from the other’s participation—are not a deduction from the platform’s contribution. They are the product the platform creates by bringing both sides together.

Platform prices therefore balance the two sides jointly. Effects must be assessed across the market as a whole, and a price observed on one side of a two-sided platform reveals little in isolation. Treating developer participation as a subtraction from platform value mistakes the mechanism of value creation for a reason to deny remuneration for it.[38]

C. The Combined Principles Create a One-Way Ratchet

Individually, the pricing principles are indeterminate. Together, they point in only one direction.

Paragraph 10 requires the fee to be ‘no higher than’ a level consistent with the principles, while the consultation says that cost and value should serve as ‘a cross-check on each other’.[39] But because the firm must demonstrate compliance ‘to the satisfaction of the CMA’, that cross-check can only reduce the defensible fee, not increase it. Value evidence cannot justify a fee above the cost model’s output, while a low cost estimate caps the fee regardless of what the value analysis shows.

The promised ‘degree of flexibility’ is therefore flexibility in one direction. The third principle—administrative simplicity—only sharpens the irony. The CMA notes that current UK fee structures are ‘relatively simple’ and that arrangements elsewhere have become ‘overly complex’.[40] Yet the CR would import the very machinery that produced that complexity: cost models, attribution accounting, monitoring APIs, and continuing supervisory review.

Paragraph 9(b) hard-codes commercial terms with the same downward bias. A steering fee may apply only when the transaction occurs within seven days of redirection, and no fee may attach to a subscription that ‘renews automatically and was entered into prior to’ the redirection. Together, those provisions exclude much of the lifetime value of a steered customer relationship from compensation. The under-remuneration is not incidental. It is built into the design.

The seven-day cut-off also rests on a mistaken account of how platform value is created. It assumes a discrete referral event, as though the platform’s contribution ended once the user reached an external page. In reality, platforms continue to support the relationship through updates, notifications, discovery, and re-engagement tools that keep users active long after any single redirection. A customer is retained over time, not simply ‘acquired’ once.

Attribution windows are also ordinary commercial terms, common in affiliate and referral arrangements. As ICLE has explained, they do not admit of precise regulatory calibration, and no economic principle makes seven days—or any other fixed period—the uniquely correct answer.[41]

D. Comparable Fees Do Not Support Intervention

If the CMA nevertheless insists on assessing fee levels, benchmarking comparable services remains the least-bad method. Applied conservatively, it does not support intervention.

Gaming-console storefronts charge 30 per cent. Steam charges 30 per cent, falling to 20–25 per cent at higher volumes. Amazon Marketplace referral fees range from 8 to 45 per cent, while food-delivery and online-travel platforms commonly charge 15–30 per cent. Against that range, steered-transaction fees already in force—15 and 10 per cent in Japan, 10–20 per cent in Google’s global rollout, and the EU’s alternative business terms—are unremarkable and often lower.[42]

Developer dissatisfaction does not change the analysis. Firms naturally prefer lower input prices, but that preference does not show that prevailing rates depart from competitive norms. The CMA’s legitimate concern that fees remain ‘commercially viable for developers’ is empirical.[43] Observed adoption provides the relevant evidence: where steering has launched at fees within the benchmark range, developers have used it. Google’s UK rollout should provide a domestic data point within the year.

At the other extreme, some developers argue that any fee on a steered transaction is illegitimate because the sale occurs ‘outside’ the platform. The U.S. district court’s zero-commission order embraced that position, and its defects are instructive.

A steered sale is not independent of the platform. The developer acquired, retained, and re-engaged the customer through infrastructure it did not build. The ability to prevent free riding is what makes investment in that infrastructure worthwhile. Anti-steering rules and access fees are alternative means of serving the same function. Removing the first while prohibiting the second would turn platform access into a mandatory gift.

Amex recognised that vertical restraints of this kind may be procompetitive because they prevent free riding. A CR that permitted no meaningful fee would therefore leave the platforms with less protection than ordinary competition law provides.[44]

III. The CR Fails the DMCC’s Proportionality Test

Section 19(5) of the DMCC makes proportionality the principal legal constraint on the design of CRs.[45] The CMA must do more than identify a legitimate objective. It must establish, on evidence rather than assumption, that the measure is necessary, that no equally effective but less onerous alternative exists, and that its expected benefits justify its disadvantages.

The proposed CR does not appear to meet that standard. The assessment assumes away the loss of platform revenue, relies on unsupported claims about pass-through and innovation, overlooks system-wide costs borne beyond individual steered transactions, and never tests an obvious graduated alternative. It also lacks consumer-outcome metrics capable of revealing whether the intervention has failed. The result is not a proportionality analysis so much as a restatement of the case for intervention.

A. The CMA Assumes Away the CR’s Largest Cost

The CMA acknowledges that the CR will reduce platform revenue but declines to treat that reduction as a cost. It reasons that current fees reflect the platforms’ substantial and entrenched market power and their restrictions on steering.[46]

That conclusion is doing the work the fee inquiry was supposed to perform. Whether current fees exceed the level competition would produce is precisely the disputed question. The CMA concedes that it lacks a reliable benchmark for the competitive counterfactual, as Section III.B explains. The statutory framework nevertheless requires the CMA to assess the intervention’s disadvantages, not define them away. A transfer from platforms to developers counts as a pure social gain only if every pound transferred was improperly retained in the first place.

The consultation also claims that the regulated fee will ‘better reflect the price Apple/Google would be able to charge when faced with more effective competition’. Yet the effectiveness assessment acknowledges that the value adjustment needed to construct that price is ‘complex by its nature and open to a degree of subjectivity’. The CMA has also rejected every comparator against which the result might be tested.[47]

The proportionality analysis therefore excludes the CR’s largest cost by invoking a counterfactual price that the consultation elsewhere admits it cannot reliably construct.

Once the CMA assumes rather than demonstrates that current fees are excessive, the proportionality exercise no longer evaluates the intervention. It simply restates the premise of designation. That reverses the statutory order of analysis: the CMA must first identify and assess the disadvantages before deciding whether they are proportionate to the CR’s aims.

B. The Claimed Benefits Depend on Unsupported Pass-Through

The CMA’s headline benefit is that ‘tens of millions of pounds per year could be either retained by developers for further investments or passed on to UK end-users in the form of lower prices’.[48] Neither claim withstands scrutiny.

To the extent developers retain the savings, the CR produces a transfer. That transfer counts as a benefit only on the circular premise discussed in Section II.A—and it is not necessarily a consumer benefit, to which the DMCC requires the CMA to give particular regard.[49] The developers best placed to take advantage of low steering fees are large global firms, so much of the retained margin may accrue outside the United Kingdom.

To the extent developers pass on the savings, the projected consumer benefit depends on the rate of pass-through. The assessment assumes that lower steering fees will produce ‘modest price decreases’ and ‘millions of pounds’ in savings for UK consumers.[50] That result is far from self-evident, particularly in multisided markets.

Platforms optimise prices across distinct user groups. A constraint on one margin may therefore prompt rebalancing through quality, investment, ancillary charges, or prices elsewhere in the ecosystem. The incidence of a fee cap cannot be inferred from the reduction in the capped fee alone.[51]

Experience with other platforms illustrates the risk. Caps on payment-card interchange fees produced only partial reductions in consumer prices, while banks recovered lost revenue through higher account charges.[52] When U.S. cities capped delivery-platform commissions, platforms raised consumer delivery fees and shifted promotions towards chain restaurants outside the caps. Independent restaurants—the intended beneficiaries—received fewer orders and lower revenue.[53]

Distribution matters as much as aggregate magnitude. Large developers with recognised brands, in-house payment systems, and direct customer relationships are best positioned to exploit low steering fees. Smaller developers gain less and face greater exposure to rebalancing through higher charges elsewhere, reduced investment in shared tools, or diminished free distribution. As the CMA has acknowledged, smaller developers ‘often do not pay the highest commission rates’ and represent a substantial share of the UK developer base.[54]

The consultations’ case studies show who might gain but do not answer the incidence question. The Apple Consultation features a globally distributed social-media platform paying the 30 per cent headline rate. The Google Consultation considers a developer earning more than $1 million annually on the standard tier.[55] Neither examines how the proposal would affect the far larger population of small developers.

The claimed innovation benefits are also overstated. The prospect of ‘new products, new entry, [and] new opportunities to offer more seamless services’ flows from the steering right: the ability to communicate with customers, choose a payment processor, and structure offers independently.[56] Paragraphs 3 to 8 secure those gains. Paragraphs 9 to 13 do not.

Because the two components are severable, each must carry its own proportionality justification. The CMA cannot transfer the innovation benefits produced by the steering right to the separate price-control regime. Where the ultimate incidence depends on rebalancing across many participants, proportionality requires evidence, not assumption.

C. The Opt-Out Theory Ignores System-Wide Costs

The CMA treats costs to developers and end users as negligible because ‘both groups have the option to avoid steering, incurring no associated costs if they choose to do so’.[57] That reasoning fails for costs imposed at the system level rather than on individual transactions.

Security and fraud risks from large-scale redirection, reduced investment in shared infrastructure as its funding base erodes, added interface complexity, and the administrative machinery required by paragraphs 9 to 13 affect developers and users whether or not they steer. Non-steering developers may bear higher costs or receive fewer shared services, while users who never follow an external link may still face weaker security or a degraded platform experience.

The consultation’s own analysis also undercuts its opt-out premise. It acknowledges that ‘end-users may not always be in a position to accurately judge the privacy and/or security risks associated with particular decisions’.[58] The CMA therefore cannot assume that individual choice eliminates costs arising from the system-wide effects of steering.

The assessment dismisses one such cost with particular haste. It describes Apple’s compliance costs—including ‘the need to gather cost accounting data as a basis for setting fees’—as ‘unlikely to be substantial’.[59] Yet the same document repeatedly notes that Apple ‘does not in the ordinary course systematically allocate costs to specific products or services’.[60]

Paragraph 12(d) would therefore require Apple to build and maintain a product-level cost-accounting system that does not now exist, alongside quarterly reporting and monitoring APIs.[61] The consultation offers no estimate of the cost of that apparatus. Unlike transaction-specific costs, it would arise regardless of how many developers or users choose to steer.

More broadly, the assessment reduces the costs of steering to payment charges. Those charges help finance a bundle of jointly supplied and jointly consumed services, including distribution, discovery, developer tools, and security. Every participant benefits from those services, not only developers that use in-app payment.

Because the platform finances these services across the ecosystem, the proportionality analysis cannot stop with developers that choose to steer. It must consider how lower platform revenue may affect incentives to maintain investments that benefit non-steering developers and end users alike.

D. The CMA Never Tests the Least-Onerous Alternative

The consultation concludes that, because any effective measure ‘would require the removal of these restrictions’, the proposal is ‘the only effective measure, which is therefore the least onerous’.[62] That does not follow.

For the fee provisions, the CMA says only that it has ‘not identified any other equally effective conduct requirement designs’.[63] But the duty to choose the least onerous among equally effective measures requires a comparative assessment. The CMA cannot satisfy that requirement by declining to identify alternatives. On that approach, the test would be met whenever the authority advanced a single option, effectively reading the least-onerous requirement out of the framework.

The omission is especially striking because the CR’s two components are severable and an obvious graduated alternative exists. The CMA could proceed now with steering rights, non-discrimination, and fee transparency, while reviewing fees ex post against commercial benchmarks and defining a trigger for escalation. It could then reserve rate regulation for evidence that fees have frustrated the remedy. The consultation never assesses that option, and an alternative does not become less effective merely because it goes unexamined.

A graduated approach would also follow the CMA’s own logic. In defending the principles-based form of the fee provisions, the consultation notes that, if Apple or Google ‘failed to comply effectively with this higher-level CR’, the CMA could impose more detailed requirements.[64] The CMA likewise holds broader app-distribution measures—including ‘directly constraining Apple’s commission fees’—in reserve, ‘in particular if a steering intervention does not have the intended effect’.[65]

That is the sensible sequence: observe first, then escalate on evidence. The consultation applies it to every instrument except the steering fee and never explains the exception.

The asymmetry of error costs makes that omission consequential.[66] A CR that proves too weak can be strengthened in light of evidence. The DMCC’s review, revocation, and compliance-reporting provisions exist for that purpose.[67] A rate-regulation regime, once embedded in cost models, accounting systems, and reliance interests, is much harder to unwind.

The proposal also lacks any credible mechanism for recognising failure. Every metric the CMA proposes to monitor—steering uptake, transaction shares, platform charges, and abandonment rates—measures activity rather than consumer welfare.[68] None directly tracks prices, output, quality, fraud, or effects on smaller developers. Under those metrics, the intervention cannot fail; it can only ‘need strengthening’.

The European Commission’s first review of the DMA illustrates the danger. It catalogues enforcement outputs as evidence of success while treating stakeholder-reported harms as consultation feedback rather than findings.[69]

The CMA should therefore commit in the final decision to reassess the CR after a defined period against published consumer-outcome metrics, including prices, output, fraud and complaint rates, and effects on smaller developers. The decision should also provide expressly for narrowing or revoking the CR if its disadvantages outweigh its benefits.

IV. The CR Unduly Restricts Safeguards and Disclosure

Our disagreement concerns where the CR draws the line, not whether it should draw one. Steering may promote competition, and platforms may invoke safeguards as a pretext for obstruction. But the draft responds by constraining legitimate protections too aggressively.

Paragraph 3(b) recognises fewer justifications than the consultation’s own evidence supports, applies a narrower standard than the DMA, and fails to reflect the CMA’s stated commitment to protect security and privacy. The consultation also discounts emerging evidence that steering moves transactions from integrated fraud-prevention and dispute-resolution systems into developer-specific environments of uneven quality.

The interstitial provisions compound the problem by restricting truthful disclosures about the protections users lose when they transact off-platform merely because that information may discourage conversion. A regime committed to trust, transparency, and open choices should enable informed decisions, not sanitise them.

The subsections below address the appropriate scope of platform safeguards, the still-developing evidence on security and fraud risks, and the information a neutral interstitial screen should convey.

A. The Safeguards Standard Is Too Narrow

Paragraph 3(b) permits a platform to restrict a redirection mechanism only where doing so is ‘strictly necessary for an objectively justifiable purpose of preventing malware, fraud, unlawful content or content harmful to children’.[70] That closed list omits other legitimate risks, including threats to user privacy, scams and subscription traps that fall short of fraud, inadequate payment processors, and harms to vulnerable adults.

Those omissions are difficult to reconcile with the consultation’s own record. The CMA recounts, without contradiction, that dynamically generated links are inherently ‘gameable’: they may resolve to compliant destinations during review and change afterwards. It also notes that link parameters may ‘be used to exfiltrate information about the user from the app’, with pixel trackers collecting that information and transmitting it to third parties without the user’s knowledge or recourse.[71] Paragraph 3(b) does not adequately address those privacy risks.

The standard is also stricter than the regimes the CMA elsewhere treats as reference points. Even the DMA permits gatekeepers to adopt measures that are ‘strictly necessary and proportionate’ to protect the integrity of the operating system across its Article 6 obligations.[72] The 9th U.S. Circuit Court of Appeals in Epic likewise accepted Apple’s security and privacy justifications as legitimate and non-pretextual.[73]

The CMA’s stated objective is to preserve ‘appropriate safeguards to protect legitimate platform interests’, including security and privacy.[74] The CR should reflect that objective. Paragraph 3(b) should adopt a ‘reasonable and proportionate’ standard, and its list of legitimate purposes should be non-exhaustive and expressly include privacy and consumer protection.

The consultation responds that steering risks ‘may be manageable’.[75] But that assumes platforms retain the tools needed to manage them. As drafted, paragraph 3(b) withdraws those tools ex ante.

B. Steering Shifts Transactions into Riskier Environments

The consultation records, but largely discounts, the platforms’ evidence about what steering bypasses.[76] App-store billing systems are not merely payment rails. They integrate fraud detection, subscription management, and dispute resolution at scale.

Apple reports that it blocked more than $9 billion in fraudulent transactions over five years, including $2 billion in 2024, terminated 146,000 developer accounts, and identified 4.7 million stolen credit cards. Google Play Protect performs a comparable role, while Android-team analysis indicates that apps obtained outside curated stores are more than 50 times as likely to contain malware.[77]

Multisided platforms invest heavily in screening participants and excluding bad actors because user trust is part of the product.[78] Steering moves transactions out of that environment and into developer-specific systems of widely varying quality.

The available evidence remains immature. Steering regimes have operated only since 2024 in the United States, 2024–25 in the EU, December 2025 in Japan, and June 2026 in the United Kingdom. The relevant harms—including phishing during redirection, post-transaction data exfiltration, fraudulent refund disputes, and subscription traps—are diffuse, under-reported, and often slow to emerge.[79]

The Dutch experience illustrates the mechanism. After alternative payments were mandated for dating apps, reported dating-app fraud rose markedly, even if the evidence does not establish causation conclusively.[80]

The absence of large-scale, attributable failures within such a short period does not establish that the risks are negligible. The CR should not be drafted as though it does.

C. Interstitial Screens Should Inform, Not Merely Redirect

We support the CMA’s decision to permit a single, neutrally worded interstitial screen. The draft, however, treats that screen as tolerated friction rather than consumer protection.

Paragraph 7(a) limits the screen to information ‘strictly necessary to inform an End-user that they are moving outside of’ the in-app purchase system. Paragraph 8(c) separately prohibits ‘warnings, disclosures, or other messaging’ whose ‘purpose or effect’ is to discourage a steered transaction.

Read together, those provisions may bar truthful, material disclosures that platform-mediated refunds, centralised subscription management, and family controls do not apply off-platform. Accurate information may discourage some transactions, but that does not make it improper. A regime whose statutory objectives include trust and transparency should not condemn a disclosure merely because it affects conversion.[81]

Users who transact off-platform assume counterparty risk with developers they may not know. A brief, factual explanation of what changes is the minimum that informed choice requires.[82]

The drop-off evidence cited against interstitial screens warrants scepticism in both directions. Developers estimate conversion losses from a single neutral screen at anywhere from 5 per cent to more than 60 per cent. That range itself suggests conjecture. The CMA rightly acknowledges that developers ‘are unlikely to have tested the exact scenario’, so the figures reflect expectations rather than measurement.[83]

The broader empirical record points in the opposite direction. Mandated choice screens and disclosures have rarely changed user behaviour materially, as the EU’s browser and search-engine choice screens demonstrated over several years.[84] If interstitial screens barely affect behaviour, the case for stripping them of useful information to preserve conversion is weak.

The CMA should reject the further restrictions on which it seeks views, including time limits, frequency limits, and default-off settings.[85] Each would reduce the screen’s value as a consumer-protection tool without a demonstrated benefit.

The comparison with physical-goods apps also misses the point.[86] Those transactions occur outside platform billing by design and generally benefit from established card-network protections. Digital-content transactions present different risks, including minors’ purchases, subscription abuse, and refund disputes. That asymmetry supports disclosure rather than undermines it.

V. The CR Sacrifices Dynamic Competition for Static Gains

A final concern is dynamic competition. Returns above incremental cost are not a regulatory anomaly to be engineered away. They finance fixed-cost creation, in platforms as in intellectual property more broadly. Commissions on paid transactions support free distribution, developer tools, and the reduced small-business rates from which most UK developers benefit.

Compressing those commissions towards incremental cost would not merely transfer rents to large developers. It would invite rebalancing through device prices, developer terms, and investment in shared services. Much of that burden would fall on the free-app majority and on consumers. As ICLE argued to the U.S. Supreme Court, this is a zero-sum reallocation: gains to a small group of developers reliant on paid content come at the expense of the much larger group that is not.[87]

The platforms’ continued profitability does not answer the point. Investment decisions are made at the margin. A rule that caps returns on marginal steered transactions near incremental cost tells investors that much of the upside from improving the platform will accrue elsewhere.

The deeper problem is that cost-plus regulation optimises for the wrong kind of competition. Anchoring prices to incremental cost pursues static allocative efficiency by reducing margins on today’s transactions. But consumer welfare in mobile ecosystems has come largely from dynamic competition: new devices, new capabilities, new form factors, and the sustained reinvestment needed to produce them.

Consumers may not remember the year a commission fell by a few percentage points, but they remember the features that reinvestment made possible. Treating a rapidly evolving platform like a mature utility with predictable demand and little left to invent trades a large, uncertain dynamic gain for a small, contestable static one.

The consultation’s headline benefit—lower prices on steered transactions—reflects precisely that static lens. The CMA’s admitted difficulty in identifying a ‘reasonable rate of return’ exposes the same problem: a utility model is poorly suited to pricing innovation rents.

The proposal would also make the United Kingdom a global outlier. U.S. courts, including the Epic panel whose cost language the CR echoes, accept that Apple may charge a commission on linked-out purchases. Japan permits steered-transaction fees of 10 and 15 per cent. Even the DMA, at least initially, prescribed neither an LRIC methodology nor a market-power haircut.

The DMA’s record nonetheless illustrates the costs of regulatory maximalism: compliance costs far beyond the European Commission’s projections; features delayed or withheld from EU users, including Siri AI, iPhone Mirroring, and Google’s AI Overviews; and intended beneficiaries leaving the market or reporting 40–50 per cent declines in organic traffic following compliance changes.[88]

The CMA’s prioritisation principles require it to consider the Government’s strategic steer and the broader effects of its actions.[89] A first CR that announces UK-specific price regulation of platform access is not the signal a growth-focused competition regime should send.

VI. Recommendations and Conclusion

The CMA seeks views on the benefits, costs, and design of the proposed steering CR.[90] ICLE offers seven recommendations.

First, sever the CR. Proceed with the steering-rights core in paragraphs 1 to 8 and 14, subject to the amendments below, and withdraw paragraphs 9 to 13 as drafted.

Second, replace the proposed fee methodology with a ‘fair and reasonable’ obligation assessed ex post. That obligation should rest on three elements: a benchmarking safe harbour under which fees within the range charged by comparable platforms and other steering regimes are presumptively compliant; transparency requirements, including published fee schedules and advance notice of changes; and a defined escalation trigger under which the CMA would revisit fee regulation only if evidence gathered over a specified observation period showed that fees had made steering commercially unviable under metrics established in advance.

Third, if the CMA retains a fee methodology, it should amend it substantially. The value assessment should be primary, with cost serving as a cross-check rather than a ceiling. Paragraph 13(d)’s market-power adjustment should be deleted or defined operationally before it takes effect; a firm cannot comply with an obligation it cannot understand. The paragraph 12 cost base should reflect the integrated Mobile Platform the CMA designated. The phrase ‘to the satisfaction of the CMA’ should be replaced with an objective evidential standard. The attribution window should be lengthened, and the automatic-renewal carve-out should be removed or made presumptive rather than absolute.

Fourth, recalibrate the safeguards. Paragraph 3(b) should replace ‘strictly necessary’ with ‘reasonable and proportionate’. Its list of legitimate purposes should be non-exhaustive and expressly include privacy and consumer protection.

Fifth, retain a single neutral interstitial screen. Permit accurate, non-alarmist disclosures about protections that do not apply off-platform, and reject the proposed time limits, frequency limits, and default-off settings.

Sixth, resolve the remaining scope questions by confirming that steering may direct users to developer-owned destinations and serve transaction-completion purposes, while preserving the WebView prohibition and child-protection measures. The CR should also use outcome-based rather than pixel-level parity requirements.

Seventh, extend the implementation period beyond three months. The DMA’s compressed timelines provide a cautionary example.[91] Any fee obligations should take effect only after the steering-rights obligations, and the final decision should include a review clause requiring reassessment after a defined period against published consumer-outcome metrics, with express authority to narrow or revoke the CR.

This will be among the first substantive CRs imposed under the DMCC, and its design will help define the regime. Properly calibrated, the steering-rights core fits the participative and proportionate model the CMA has promised. Much of it is already emerging through legal and commercial developments worldwide, including in the United Kingdom.

Paragraphs 9 to 13 do not fit that model. They would commit the CMA, at the regime’s outset, to administering the price of platform access under principles that cannot yield stable answers, on an evidential record that assumes key conclusions, and under a review standard too limited to correct the errors likely to follow.

The DMCC promised greater contestability without the DMA’s collateral costs. The CMA should keep that promise.

[1] Competition & Mkts. Auth., Proposed Steering Conduct Requirement for Apple: Consultation Document (July 2026) [hereinafter Apple Consultation]; Competition & Mkts. Auth., Proposed Steering Conduct Requirement for Google: Consultation Document (July 2026) [hereinafter Google Consultation]. Responses to both consultations are due by 5 p.m. on 28 July 2026. Google Consultation ¶ 6.11.

[2] The draft conduct requirements and supporting analyses are materially identical. Unless otherwise indicated, these comments cite the Apple Consultation and provide parallel citations to the Google Consultation where the paragraph numbering differs. Consistent with the Competition and Markets Authority’s instructions, these comments respond to both consultations. See Google Consultation ¶ 6.11.

[3] See Geoffrey A. Manne, Dirk Auer & Mario A. Zúñiga, ICLE Comments to UK CMA on Competition in Mobile Ecosystems, Int’l Ctr. for L. & Econ. (12 Feb. 2025), https://laweconcenter.org/resources/icle-comments-to-uk-cma-on-competition-in-mobile-ecosystems [hereinafter ICLE Mobile Ecosystems Comments]; Geoffrey A. Manne, Dirk Auer & Mario A. Zúñiga, Comments of the International Center for Law & Economics on CMA’s Proposal to Designate Apple and Google with Strategic Market Status, Int’l Ctr. for L. & Econ. (20 Aug. 2025), https://laweconcenter.org/wp-content/uploads/2025/08/ICLE-CMA-Apple-Google-Designation-comments.pdf [hereinafter ICLE Designation Comments].

[4] Geoffrey A. Manne et al., Comments of the International Center for Law & Economics: Recent Developments in Relation to Apple’s and Google’s App Store Rules, Int’l Ctr. for L. & Econ. (22 Apr. 2026) [hereinafter ICLE App Store Comments], responding to Competition & Mkts. Auth., Views Sought: Recent Developments in Relation to Apple’s and Google’s App Store Rules (Mar. 2026) [hereinafter Call for Evidence].

[5] Apple Consultation, ch. 3, Proposed Steering CR ¶¶ 9–13; Google Consultation, ch. 3, Proposed Steering CR ¶¶ 9–13.

[6] Google Consultation ¶ 6.4 (seeking views on ‘the measures that Google rolled out in the UK on 30 June 2026’).

[7] Competition & Mkts. Auth., Digital Markets Competition Regime Guidance (CMA194, Dec. 2024) [hereinafter CMA194].

[8] Apple Consultation, Proposed Steering CR ¶ 10(a)(i).

[9] Id. ¶ 4.107.

[10] Id. Proposed Steering CR ¶ 12(b).

[11] Id. Proposed Steering CR ¶¶ 10(b)(i), 13(d)–(e).

[12] Id. Proposed Steering CR ¶¶ 9(a), 12(d); id. ¶¶ 4.150–4.151.

[13] Id. n.128.

[14] Apple Consultation ¶ 5.17; Google Consultation ¶ 5.15.

[15] Apple Consultation ¶ 4.2; Google Consultation ¶ 4.2.

[16] Apple Consultation, Proposed Steering CR ¶¶ 4, 14.

[17] Apple Consultation ¶¶ 2.6(c)(iii), 2.12(d).

[18] See ICLE App Store Comments, supra note 4, at 6; Google Consultation ¶ 6.4.

[19] Apple Consultation ¶ 4.151; Google Consultation ¶ 4.150.

[20] Lazar Radic, Steering in the Fog: The DMA and the Turn from Market Oversight to Market Ordering (ICLE White Paper No. 2026-06-15, 2026) [hereinafter Radic, Market Ordering].

[21] Geoffrey A. Manne et al., Response of the International Center for Law & Economics: Consultation on the First Review of the Digital Markets Act, Int’l Ctr. for L. & Econ. (24 Sept. 2025) [hereinafter ICLE DMA Review Response]; Radic, Market Ordering, supra note 20, at 31–40.

[22] Press Release, Eur. Comm’n, Commission Finds Apple’s App Store Rules Breach Digital Markets Act (23 Apr. 2025), https://ec.europa.eu/commission/presscorner/detail/en/ip_25_1085; see ICLE App Store Comments, supra note 4, at 5 (describing successive scrutiny of the Core Technology Fee, initial acquisition fee, and tiered Store Services fee).

[23] Digital Markets, Competition and Consumers Act 2024, c. 13, § 103 (UK) (providing that courts review appeals from Competition and Markets Authority decisions under the digital-markets regime using judicial-review principles).

[24] Comments of the International Center for Law & Economics on the CMA’s Draft Digital Markets Competition Regime Guidance 25, Int’l Ctr. for L. & Econ. (2024) [hereinafter ICLE DMCC Guidance Comments] (urging the Competition and Markets Authority to begin with narrower, individual conduct requirements so it could observe the relationship between each intervention and its effects).

[25] Apple Consultation ¶ 6.8(a); Google Consultation ¶ 6.9(a).

[26] Carl Shapiro & Hal R. Varian, Information Rules: A Strategic Guide to the Network Economy 3 (Harv. Bus. Sch. Press 1999) (‘Information is costly to produce but cheap to reproduce … cost-based pricing does not work: a 10 or 20 per cent markup on unit cost makes no sense when unit cost is zero.’).

[27] ICLE App Store Comments, supra note 4, at 4.

[28] Competition & Mkts. Auth., Strategic Market Status Investigation into Apple’s Mobile Platform: Final Decision (22 Oct. 2025); Competition & Mkts. Auth., Strategic Market Status Investigation into Google’s Mobile Platform: Final Decision (22 Oct. 2025); see Apple Consultation ¶¶ 4.7–4.8 & n.40.

[29] On two-sided platform pricing and how platforms divide surplus with complementors, see Jean-Charles Rochet & Jean Tirole, Two-Sided Markets: A Progress Report, 37 Rand J. Econ. 645 (2006); David S. Evans & Richard Schmalensee, Matchmakers: The New Economics of Multisided Platforms (2016).

[30] Epic Games, Inc. v. Apple, Inc., No. 23-16234, slip op. at 41 (9th Cir. 11 Dec. 2025).

[31] Brief of Former Antitrust Enforcement Officials as Amici Curiae in Support of Apple Inc.’s Petition for Panel Rehearing and/or Rehearing En Banc, Epic Games, Inc. v. Apple, Inc., No. 25-2935 (9th Cir. 12 Mar. 2026); Brief of the International Center for Law & Economics as Amicus Curiae in Support of Appellant, Epic Games, Inc. v. Apple, Inc., No. 25-2935 (9th Cir. 30 June 2025).

[32] Verizon Commc’ns Inc. v. Law Offices of Curtis V. Trinko, LLP, 540 U.S. 398, 407–08 (2004); Pac. Bell Tel. Co. v. linkLine Commc’ns, Inc., 555 U.S. 438, 452–53 (2009); see also United States v. Grinnell Corp., 384 U.S. 563, 570–71 (1966).

[33] Google Consultation ¶ 4.126.

[34] Google Consultation ¶¶ 4.125–4.132.

[35] Dr Rachael Kent v Apple Inc [2025] CAT 67, ¶ 631(11) (appeal pending); see Google Consultation ¶ 4.129.

[36] ICLE App Store Comments, supra note 4, at 4–5; see Daniel G. Swanson & William J. Baumol, Reasonable and Nondiscriminatory (RAND) Royalties, Standards Selection, and Control of Market Power, 73 Antitrust L.J. 1 (2005); Joseph Farrell, John Hayes, Carl Shapiro & Theresa Sullivan, Standard Setting, Patents, and Hold-Up, 74 Antitrust L.J. 603 (2007).

[37] Microsoft Corp. v. Motorola, Inc., 795 F.3d 1024 (9th Cir. 2015); Unwired Planet Int’l Ltd v Huawei Techs. Co [2020] UKSC 37; Optis Cellular Tech. LLC v Apple Retail UK Ltd [2025] EWCA Civ 552.

[38] Ohio v. Am. Express Co., 138 S. Ct. 2274, 2285–87 (2018); see also Geoffrey A. Manne, In Defence of the Supreme Court’s ‘Single Market’ Definition in Ohio v American Express, 7 J. Antitrust Enf’t 104 (2019); Andrei Hagiu, Proprietary vs. Open Two-Sided Platforms and Social Efficiency (AEI-Brookings Joint Ctr. for Regul. Stud., Working Paper No. 06-12, 2006).

[39] Apple Consultation ¶¶ 4.138–4.140.

[40] Id. ¶ 4.135.

[41] Apple Consultation, Proposed Steering CR ¶ 9(b); see ICLE App Store Comments, supra note 4, at 9.

[42] ICLE App Store Comments, supra note 4, at 6 & n.13.

[43] Apple Consultation ¶ 4.85(b); Google Consultation ¶ 4.86(b).

[44] Ohio v. Am. Express Co., 138 S. Ct. at 2289–90 (recognising that anti-steering provisions may serve procompetitive purposes by preventing free riding); Brief of the International Center for Law & Economics as Amicus Curiae in Support of Petitioner at 9–14, Apple Inc. v. Epic Games, Inc., No. 23-344 (U.S. 27 Oct. 2023) [hereinafter ICLE Supreme Court Brief].

[45] Digital Markets, Competition and Consumers Act 2024, c. 13, § 19(5) (UK) (‘The CMA may only impose a conduct requirement or a combination of conduct requirements on a designated undertaking if it considers that it would be proportionate to do so for the purpose of one or more of the following objectives: (a) the fair dealing objective, (b) the open choices objective, and (c) the trust and transparency objective, having regard to what the conduct requirement or combination of conduct requirements is intended to achieve.’).

[46] Apple Consultation ¶ 5.17; see Google Consultation ¶ 5.15 (same for Google).

[47] Apple Consultation ¶¶ 4.126–4.128; Google Consultation ¶¶ 4.124–4.128; see infra Section III.B.

[48] Apple Consultation ¶ 5.44; Google Consultation ¶ 5.41.

[49] Digital Markets, Competition and Consumers Act 2024, c. 13, § 19(10) (UK); Apple Consultation ¶ 2.11.

[50] Apple Consultation ¶ 5.42(a)(i).

[51] See Rochet & Tirole, supra note 29; Evans & Schmalensee, supra note 29.

[52] Ernst & Young & Copenhagen Econ., Study on the Application of the Interchange Fee Regulation (2020); Todd J. Zywicki, Geoffrey A. Manne & Julian Morris, Price Controls on Payment Card Interchange Fees: The U.S. Experience (ICLE White Paper, 2014); Julian Morris, Todd J. Zywicki & Geoffrey A. Manne, The Effects of Price Controls on Payment-Card Interchange Fees: A Review and Update (ICLE White Paper No. 2022-03-04, 2022).

[53] Zhuoxin Li & Gang Wang, Regulating Powerful Platforms: Evidence from Commission Fee Caps, 36 Info. Sys. Res. 126 (2025).

[54] Call for Evidence, supra note 4, ¶ 21; see ICLE App Store Comments, supra note 4, at 7–8.

[55] Apple Consultation, case study following ¶ 1.13; Google Consultation, case study following ¶ 1.15.

[56] Apple Consultation ¶¶ 2.12(a), 5.36–5.40; Google Consultation ¶¶ 2.12(a), 5.33–5.37.

[57] Apple Consultation ¶ 5.43; Google Consultation ¶ 5.40.

[58] Apple Consultation ¶ 5.29.

[59] Apple Consultation ¶ 5.16.

[60] Apple Consultation ¶¶ 4.93, 4.109, 4.116.

[61] Apple Consultation, Proposed Steering CR ¶ 12(d); id. ¶¶ 4.136, 4.150–4.153.

[62] Apple Consultation ¶¶ 5.8–5.9; Google Consultation ¶¶ 5.8–5.10.

[63] Apple Consultation ¶ 5.9.

[64] Apple Consultation ¶ 4.78; Google Consultation ¶ 4.79.

[65] Apple Consultation ¶¶ 1.17–1.19.

[66] See Frank H. Easterbrook, The Limits of Antitrust, 63 Tex. L. Rev. 1 (1984); Geoffrey A. Manne & Joshua D. Wright, Innovation and the Limits of Antitrust, 6 J. Competition L. & Econ. 153 (2010).

[67] Digital Markets, Competition and Consumers Act 2024, c. 13, §§ 22, 25, 84 (UK).

[68] Apple Consultation ¶ 4.153.

[69] Radic, Market Ordering, supra note 20, at 34–35 (discussing Eur. Comm’n Staff Working Document SWD(2026) 123 final (28 Apr. 2026)).

[70] Apple Consultation ¶ 3(b).

[71] Apple Consultation ¶¶ 4.41–4.42.

[72] Regulation (EU) 2022/1925 of the European Parliament and of the Council of 14 Sept. 2022 on Contestable and Fair Markets in the Digital Sector (Digital Markets Act), arts. 6(4), 6(7), 2022 O.J. (L 265) 1.

[73] Epic Games, Inc. v. Apple, Inc., 67 F.4th 946, 971, 985–86 (9th Cir. 2023), cert. denied, 144 S. Ct. 681 (2024).

[74] Apple Consultation ¶ 2.5(d).

[75] Apple Consultation ¶ 4.34.

[76] Apple Consultation ¶ 5.19; Google Consultation ¶ 4.19.

[77] Apple, App Store Prevented More than $9 Billion in Fraudulent Transactions (27 May 2025), https://www.apple.com/newsroom/2025/05/the-app-store-prevented-more-than-9-billion-usd-in-fraudulent-transactions; Suzanne Frey, A New Layer of Security for Certified Android Devices, Android Devs. Blog (25 Aug. 2025), https://android-developers.googleblog.com/2025/08/elevating-android-security.html.

[78] Evans & Schmalensee, supra note 29, at 138–39 (noting that quality control ‘requires platforms to invest significant efforts into investigating participants and kicking out bad ones’).

[79] ICLE App Store Comments, supra note 4, at 10.

[80] See ICLE App Store Comments, supra note 4, at 11 & nn.33–34 (discussing the Dutch Authority for Consumers and Markets’ alternative-payment mandate for dating apps and the subsequent rise in reported fraud).

[81] Digital Markets, Competition and Consumers Act 2024, c. 13, § 19(8) (UK); Apple Consultation, app. A, ¶ A.2(c).

[82] ICLE App Store Comments, supra note 4, at 8–9.

[83] Apple Consultation ¶¶ 4.61–4.62.

[84] ICLE Mobile Ecosystems Comments, supra note 3, at 10–12; Geoffrey A. Manne, A Critical Analysis of the Google Search Antitrust Decision 16–17, Int’l Ctr. for L. & Econ. (14 Aug. 2024).

[85] Apple Consultation ¶ 4.67.

[86] Apple Consultation ¶ 4.64.

[87] ICLE Supreme Court Brief, supra note 44, at 21.

[88] ICLE DMA Review Response, supra note 21, at 18, 23–25; Radic, Market Ordering, supra note 20, at 27, 34–35.

[89] Apple Consultation, app. A, ¶ A.5; Competition & Mkts. Auth., Prioritisation Principles (CMA188, Oct. 2023).

[90] Apple Consultation ¶ 6.2; Google Consultation ¶ 6.2.

[91] Apple Consultation ¶ 4.149; see ICLE DMA Review Response, supra note 21, at 18.

Apolicy statement about accuracy should, at minimum, be precise. The Federal Trade Commission’s (FTC) Proposed Policy Statement Concerning the Suppression of Accuracy in Artificial Intelligence Systems (AI Policy Statement) struggles with that assignment.

The International Center for Law & Economics (ICLE) submitted comments in response to the FTC’s request for input, and this post is, in part, a digest of those comments. The statement gets some important things right, including the risks of excessive regulation and a patchwork of state laws. But it offers little concrete guidance on deception, leans on dubious assumptions about consumer expectations, wanders into constitutionally protected editorial judgments, and treats federal preemption as more wish than doctrine.

In other words, the AI Policy Statement needs considerable work. Then again, perhaps guiding enforcement was never quite the point.

But first, some context.

Read the full piece here.

American women are having fewer kids than they’d like. By one estimate, the average fertility gap—the distance between a woman’s ideal number of children and achieved fertility—for U.S. women is 0.76, close to one child. That figure, though, varies significantly between states: in North Dakota, the gap is one-third, while in New Hampshire, Rhode Island, and Alaska, it’s greater than one child.

These state-level differences help us investigate how state policy could be shaping family decisions. Affordable childcare options are key for young families, and regulations of childcare vary greatly by state. Using the fertility-gap metric and a novel Childcare Regulations Index, we set out in a recent working paper to understand just how pertinent childcare regulations are when it comes to the gap between ideal and achieved family size.

Read the full piece here.

Executive Summary

Google’s AI Overviews has revived a familiar conflict between publishers and digital platforms. As in earlier disputes over news aggregation, publishers argue that Google uses their content while weakening the traffic and revenue on which their business models depend. Early evidence gives this claim more support than prior complaints about snippets and links, because AI-generated summaries appear to reduce click-through rates and keep more users within Google Search.

The legal setting has also changed. Courts and policymakers have yet to determine whether the training, fine-tuning, grounding, and inference-stage use of copyrighted works infringe copyright or fall within existing exceptions. Despite that uncertainty, authorities in Europe and elsewhere have begun using competition law to require opt-outs, strengthen publishers’ bargaining position, and press platforms toward compensation.

This white paper argues that this approach targets the wrong market risk. Competition law cannot guarantee payment whenever AI-generated summaries reduce publisher traffic. It offers little basis for intervention when a nondominant AI entrant provides summaries, when Google offers them through a standalone service, or when Google contracts with only selected publishers. Those disputes primarily concern copyright and licensing.

The stronger antitrust concern involves competition among AI providers. By integrating AI Overviews into its dominant search engine and placing its own summaries above organic results, Google may divert users from rival assistants and agents that lack a comparable distribution channel. Antitrust authorities should therefore examine whether that integration impedes entry or favors Google’s own AI services. Copyright law should determine whether publishers’ works may be used and on what terms. Competition law should protect rivalry among Google and emerging AI providers.

I. Introduction: AI Overviews Revives an Old Fight

In May 2024, Google launched AI Overviews, a feature that places AI-generated summaries above the organic results for certain Google Search queries. Each overview offers a synthesized answer and links to the underlying web sources. Because the summaries appear at the top of the search-engine results page, publishers fear that users will find their answers without clicking through to the source websites.

Lower click-through rates could weaken publishers’ ability to earn advertising and subscription revenue from website traffic. Publishers therefore view AI Overviews as a threat to their business models and, more broadly, to the financial sustainability of online publishing.[1] Google offers a different account. It claims that the feature expands demand and that links appearing within AI Overviews receive higher click-through rates.[2]

The dispute may feel familiar. Platforms and publishers have spent years making the same competing claims about the online use of news content. That conflict has prompted litigation and policy interventions around the world, including new copyright protections and rules that encourage or require platforms to negotiate with publishers.[3]

In the European Union, the Directive on Copyright in the Digital Single Market (CDSM Directive) created a neighboring right covering the online use of press publications.[4] The provision’s limits and uncertainties have led authorities to enlist competition law in its enforcement. The aim has been to turn the publishers’ right into a credible bargaining tool, rebalance negotiations between news producers and online distributors, and prevent dominant platforms from frustrating the directive’s objectives.[5]

Other countries have used regulation to avoid the perceived limits of copyright law. Australia adopted a mandatory bargaining code designed by the Australian Competition and Consumer Commission. That model has since influenced several other jurisdictions.[6]

AI Overviews has opened a new chapter in the same conflict, with many of the same parties and arguments. In December 2025, the European Commission opened a formal antitrust investigation into whether Google violated EU competition law by using publishers’ online content for artificial-intelligence purposes.[7] In April 2026, Brazil’s Administrative Council for Economic Defense, known as CADE, reopened an investigation into Google’s use of news content, including through AI Overviews, after dismissing the original inquiry in late 2024.[8]

The Japan Fair Trade Commission also launched a follow-up to its 2023 market study on news-content distribution. The agency cited concerns that generative-AI search providers may use news content to answer queries without news organizations’ permission.[9] In the United Kingdom, the Competition and Markets Authority used its digital-markets powers[10] to require Google to give publishers control over whether their content is used to ground responses generated by features such as AI Overviews. The requirement also allows publishers to opt out of having their content used to train, ground, or fine-tune AI models.[11]

The AI Overviews dispute differs from the earlier fight over online news in two important respects. First, publishers’ claim that online aggregators free ride on their work remains largely hypothetical and lacks consistent empirical support.[12] Early studies of AI Overviews, by contrast, point to measurable displacement of website traffic.

Second, the European copyright response to online news faced criticism from the outset as both unjustified and ineffective. Copyright has a more central role in the AI Overviews dispute. Courts and scholars are still debating whether using protected works to train AI models infringes copyright or falls within an exception. Antitrust authorities are intervening before those questions have been resolved, even as courts across several jurisdictions consider them. The Court of Justice of the European Union also has a relevant case pending.[13]

Despite these differences, the two debates appear to pursue the same objective through the same legal instrument. Authorities are invoking antitrust law to secure compensation for publishers even when platforms may neither free ride on news content nor infringe copyright.

This white paper argues that the effort to place antitrust law ahead of copyright law in disputes between AI providers and publishers targets the wrong market risk. Courts and legislatures should first resolve whether AI training and related uses constitute copyright infringement or fall within lawful exceptions. Competition law should focus instead on whether Google’s integration of AI Overviews into Search impairs competition among AI providers.

The more serious competition concern involves Google’s relationship with rival AI assistants and agents. By integrating AI Overviews into its dominant search engine, Google may favor its own AI service and divert users away from competing providers. Standalone AI services lack access to a comparable gateway. Preferential placement within Google Search could therefore impede entry and weaken competition among AI providers.

The distinction becomes clearer in three scenarios where AI Overviews may displace publisher traffic but antitrust law offers little or no basis for intervention on publishers’ behalf. The first arises when Google trains or grounds its model using content from selected publishers under agreements with only part of the industry. The second arises when a new AI entrant without market power offers AI-generated summaries. The third arises when Google offers AI Overviews as a standalone service rather than integrating it into Search.

Each scenario may disrupt publishers’ business models. Yet none necessarily presents a competition problem involving publishers. Competition law should focus principally on whether Google’s integration of AI Overviews harms the market for AI services. Copyright law should govern whether and on what terms AI providers may use publishers’ content.

Section II reviews the early empirical research on how AI tools affect online information consumption and publisher traffic. Section III examines whether the European copyright framework can address AI training and AI-generated summaries. Section IV evaluates recent efforts to use antitrust enforcement to strengthen publishers’ bargaining position and secure payment for their content. Section V explains why competition law should instead address the horizontal relationship between Google and rival AI providers whose entry and ability to compete may suffer when Google integrates AI Overviews into its dominant search service. Section VI concludes.

II. How AI Search Changes Web Traffic

AI Overviews uses Google’s Gemini model to aggregate and synthesize information from Google Search results and cited web sources. Through a process known as grounding, the system connects its response to external information and places a summary at the top of the search-engine results page for many queries.

AI Overviews differs from featured snippets, which usually reproduce one or two sentences from a single source. It combines information from multiple sources into a unified answer and can respond to more complex questions than traditional keyword-based search, including queries submitted through text, voice, or images.

AI Overviews has reached more than 2.5 billion monthly users. It forms part of a broader shift toward agentic artificial intelligence, which integrates more advanced AI capabilities into search and allows users to interact with assistants and agents more extensively.[14] Together with AI Mode, AI Overviews creates a connected generative-AI layer within Google Search. AI Overviews supplies a concise summary within conventional search results. AI Mode offers a conversational interface that supports follow-up questions, research, planning, exploration, and comparison.

More broadly, tools powered by large language models have attracted attention because they may change how users browse the web, conduct searches, and shop online.[15] AI Overviews presents a narrower question concerning how AI tools affect the production and distribution of online content. These tools may substitute for traditional search by answering users’ questions directly, or they may complement it by helping users discover additional sources.

Nicolas Padilla, H. Tai Lam, Anja Lambrecht, and Brett Hollenbeck provide early evidence that large-language-model adoption affected website traffic even before Google launched AI Overviews.[16] Their analysis finds particularly significant effects for small online content providers.

Qiaoni Shi, Kai Zhu, and Kai Gu compare ChatGPT sessions with Google searches to examine how AI search reallocates web traffic. They find that ChatGPT generates an outbound click in only 5.2% of conversations, far below Google’s referral rate. The remaining clicks also differ from a representative sample of Google traffic. They tend to favor specialized destinations and disfavor advertising-supported websites.[17]

Samira Gholami and coauthors reach a different conclusion. They find that rapid adoption of large language models coincides with sustained growth in the number of unique websites users visit. Their results suggest that AI tools may complement conventional web browsing by changing how users discover and consume information. Users increasingly combine large language models with traditional search rather than replacing one with the other.[18]

A growing body of research now focuses specifically on AI Overviews and its effects on website traffic and search behavior. Reports by Similarweb, the Pew Research Center, and Ahrefs find that users are less likely to click on links when a search-results page contains an AI-generated summary.[19] Estimates place the share of these “zero-click searches” at 68% in the United States.[20]

Saharsh Agarwal and Ananya Sen, as well as Haofei Xu, Umar Iqbal, and Jacob M. Montgomery, also find that AI Overviews diverts traffic from publishers, although they report a smaller increase in zero-click searches.[21] Mehrzad Khosravi and Hema Yoganarasimhan identify a similar causal effect on Wikipedia.[22] Exposure to AI Overviews reduced daily traffic to English-language Wikipedia articles by about 15%.

The effects differ across subjects. Culture-related articles experienced substantially larger declines, while articles concerning science, technology, engineering, and mathematics saw smaller effects.

Hangcheng Zhao and Ron Berman compare the launch of OpenAI’s ChatGPT in November 2022 with the introduction of Google AI Overviews in May 2024.[23] They find no clear evidence that news publishers suffered traffic losses after ChatGPT’s launch. The arrival of AI Overviews, by contrast, coincided with declines in both direct traffic and referrals from organic search results.

Zhao and Berman also examine how publishers respond to generative AI. Blocking automated crawlers through tools such as robots.txt can prevent AI systems from collecting content for model training or answering queries. Their findings suggest that this strategy may reduce publishers’ traffic because websites that block AI crawlers become less visible in AI-generated results. Publishers may fare better by investing in distinctive formats that large language models cannot easily reproduce, including multimedia and interactive features.

Riley Grossman and coauthors likewise find that Google Search, AI Overviews, and Gemini rely on substantially different sources, with little overlap among them.[24] Websites that block Google’s AI crawler are also less likely to appear in AI Overviews.

Other studies urge caution. Analyses of Semrush keyword data and clickstream data from Datos suggest that AI Overviews do not invariably reduce traffic and may coexist with other search features.[25] Their effects likely vary by query. Simple informational searches may produce larger declines in clicks, while branded searches, breaking news, original reporting, and queries requiring depth or authoritative sources may continue to send users to publishers’ websites.

Peibo Zhang, Ruomeng Cui, and Dennis J. Zhang identify a different effect on Reddit.[26] They find that AI Overviews increases engagement across communities of every size. Small and midsize communities experience the largest relative gains, which suggests that AI search can direct users toward communities they might not otherwise discover.

Further empirical work will be needed to measure the full effects of large-language-model tools on online information markets. The early evidence already shows that these tools can alter traffic flows, user behavior, and the distribution of revenue among search engines, AI providers, and content publishers.[27]

Search engines have traditionally earned revenue by monetizing users’ attention while directing them to third-party websites. Generative-AI tools may change that arrangement because a growing share of searches ends with an answer generated on the search page and no click to an external source.

Competition analysis must therefore distinguish between standalone AI chatbots and AI tools integrated into an established search engine. As Section V explains, that distinction determines whether the principal concern involves the use of publishers’ content or the ability of rival AI providers to reach users.

III. Copyright Law Struggles to Keep Pace with AI Search

Generative artificial intelligence has triggered extensive litigation and scholarly debate because it tests nearly every layer of traditional copyright law.[28] The main disputes concern the use of copyrighted works to train AI models, the copyrightability of AI-generated material, liability for infringing outputs, and the treatment of outputs that imitate a human creator’s identity or style.

Training and fine-tuning depend on vast quantities of data, much of it protected by copyright. Copyright can therefore create a substantial barrier to entry for AI developers and may slow further innovation.[29] At the same time, generative AI may threaten creative industries by producing literary and artistic works faster and at lower cost than human creators.

The proper legal response depends in part on empirical questions.[30] Generative-AI products may displace incumbent firms, expand demand by serving previously unmet needs, or do both. They may substitute for human-created works, but they may also raise creators’ productivity and reduce production costs. Their outputs may closely reproduce training materials or produce genuinely distinct works.[31]

Policymakers must therefore balance the benefits of encouraging creation against the costs that copyright restrictions impose on innovation. The legal framework must support creative industries without foreclosing entry and experimentation in artificial intelligence.

Publishers argue that Google AI Overviews creates a new form of free riding on their investments in content production. By summarizing a news article at the top of the search-results page, Google may satisfy users’ informational needs without sending them to the source. That can reduce referral traffic and allow Google to retain a greater share of advertising revenue.

Because publishers depend heavily on referral traffic, they contend that AI Overviews threatens the industry’s financial sustainability. They also argue that declining publisher revenue could weaken media pluralism, which supports informed democratic debate.[32]

These claims have familiar roots. For three decades, policymakers have sought ways to sustain journalism[33] as the internet and digital platforms changed how news is produced, distributed, and consumed.[34]

The European Union responded to concerns about free riding by creating a neighboring right for the online use of press content in the Directive on Copyright in the Digital Single Market, known as the CDSM Directive. The right was intended to encourage cooperation between press publishers and online intermediaries such as news aggregators, search engines, and social-media platforms.[35]

The protection excludes private or noncommercial uses by individuals, hyperlinking, and the use of individual words or very short extracts from a press publication. The directive also creates an exception for reproductions and extractions made for text and data mining, commonly known as TDM, unless the rights holder expressly opts out.[36]

AI Overviews exposes the limits of that framework. It remains uncertain whether the neighboring right applies when an AI-generated answer does not reproduce or make available protected expression beyond individual words or very short extracts.[37]

Scholars have raised similar doubts about the TDM exception.[38] The European Commission views the exception as broad enough to cover commercial AI training conducted on publicly available online content.[39] It also describes the opt-out mechanism as an essential means for rights holders to reserve their rights as text-and-data-mining technologies become more widely used, particularly in AI training.[40]

The Artificial Intelligence Act supports that view. It expressly refers to the TDM exceptions and requires providers to respect opt-outs declared under Article 4 of the CDSM Directive.[41]

Yet the opt-out system often operates poorly in practice. Rights holders may need to use several fragmented technical tools without assurance that developers will detect or honor their reservations.[42] AI developers, in turn, may face inconsistent or conflicting signals across large automated systems.

The system also appears better suited to training than to inference-stage uses. Inference occurs when a model generates a response to a user’s query. A system may retrieve and synthesize information from external sources during that process without incorporating the material into the model’s training.[43] Retrieval-augmented generation, for example, may use external data to improve a response without changing the model’s weights. That use does not necessarily qualify as training.

The European Commission has acknowledged that the current framework may require revision. It has announced plans to examine targeted changes to EU copyright law beyond the formal review of the CDSM Directive.[44] The stated goals include strengthening the licensing and enforcement of copyright and related rights in the AI context, improving creators’ remuneration, and making it easier for generative-AI providers to access protected content.[45]

The Court of Justice of the European Union is now considering the first case to address some of these questions.[46] In Like Company v. Google Ireland Ltd., a Hungarian court has asked whether the training and outputs of systems such as Gemini infringe exclusive rights under EU copyright law.

The dispute concerns responses generated by Google’s Gemini chatbot that included text partly identical to material on a press publisher’s website. The national court seeks guidance on the reproduction right and the right of communication to the public under the Information Society Directive,[47] as well as the TDM exception and the press publishers’ right under the CDSM Directive.

The possible obsolescence of the neighboring right has broader consequences for EU member states that adopted more interventionist versions of the CDSM Directive. Belgium, the Czech Republic, Greece, and Italy have introduced additional measures intended to strengthen publishers’ bargaining position. Although their systems differ, each draws in part on Australia’s mandatory-arbitration model.[48]

Those countries have assigned regulators or ministries a role in supporting negotiations and determining fair compensation when publishers and platforms cannot agree on terms for the online use of press publications. France is considering similar legislation.[49]

The Court of Justice recently upheld the Italian framework’s compatibility with the CDSM Directive.[50] It also clarified that the framework applies only when a platform uses press publications in a legally relevant manner. The ruling does not create a freestanding duty to negotiate or compensate publishers.

AI Overviews may encourage more member states to adopt bargaining regimes, increasing regulatory fragmentation across the European Union. Yet if the neighboring right cannot reach AI-generated summaries, those national efforts may prove ineffective. That outcome would strengthen the case for renewed copyright legislation at the EU level.

IV. Antitrust as a Back Door to Publisher Compensation

The debate over AI-generated summaries closely resembles the earlier fight over the online use of news content. The resemblance extends beyond copyright. Publishers have also turned to competition law in both disputes.

After pressing for the neighboring right created by the CDSM Directive, EU publishers remained dissatisfied with the results. They then sought help from antitrust authorities, hoping competition law would strengthen copyright enforcement and secure payment for their content.

Their main claim is that large technology platforms have weakened a relationship that could otherwise benefit both publishers and digital intermediaries.[51] Google and Meta serve as major gateways to online news and, in publishers’ view, have become unavoidable trading partners. Publishers depend heavily on Google Search and Facebook for referral traffic, which gives the platforms substantial bargaining power over how news appears and how users reach it.

Publishers argue that platforms do not negotiate over the use of snippets or other content. Instead, they offer a binary choice between participation on the platform’s terms and exclusion from its services.[52] Publishers also point to opaque ranking systems, limited transparency in advertising markets, mandatory publishing formats, and the platforms’ control over user data as evidence of this imbalance.

The same arguments now appear in the debate over AI Overviews. Critics describe the service as a break in the exchange that long governed the relationship between Google and publishers.[53] Publishers allowed Google to crawl their websites and build its search index. In return, Google sent them traffic that they could monetize through advertising or subscriptions.[54]

AI Overviews may weaken that exchange by answering users’ questions without requiring a click to the source. Yet publishers have long disputed the value of referral traffic and warned that snippets allowed platforms to free ride on their work. Empirical studies did not consistently support that claim. The risk of traffic displacement appears more concrete only with the rise of generative AI.

France provides the clearest example of using competition law to reinforce copyright. Since 2020, the French Competition Authority, known as the Autorité de la concurrence, has challenged Google’s response to the French law implementing the CDSM Directive. Google announced that it would display protected news content only if publishers authorized its use without charge.[55]

The authority treated that policy as a potential abuse of dominance and exploitation of publishers’ economic dependence. It argued that Google imposed unfair trading conditions under the threat of reduced visibility or de-indexing.[56] The Paris Court of Appeal endorsed that approach. The court held that the publishers’ right did not itself guarantee payment, but it required meaningful negotiations intended to allow publishers and news agencies to seek fair remuneration.[57]

The Autorité recently adopted a similar approach toward Meta. It accused the company of imposing its own method for calculating payment for reused news content while withholding information that publishers needed to evaluate Meta’s offers.[58]

The authority also appears ready to extend its framework to AI Overviews. It has argued that commitments imposed on Google in the news-content investigation can address new disputes involving artificial intelligence.[59]

In 2024, the Autorité sanctioned Google for failing to comply with commitments made during the investigation that began in 2020. The decision focused in part on Bard, the AI service later renamed Gemini.[60] The authority found that Google had neither informed publishers that it was using their content in the service nor provided a technical means to opt out without also affecting the display of protected content across other Google products. According to the authority, those failures impaired publishers’ ability to negotiate payment.

France has led this expansive use of competition law in disputes rooted in copyright. Other authorities have begun to follow.

The European Commission has opened an antitrust investigation into Google’s use of content from web publishers and YouTube creators for AI services.[61] The investigation asks whether Google imposes unfair terms on publishers and creators or gives itself privileged access to content that rival AI developers cannot obtain on comparable terms.

The Commission is also examining whether Google used publishers’ content in AI Overviews and AI Mode without adequate compensation and without offering a meaningful opt-out. That theory combines concerns about publishers’ bargaining position with a separate concern that Google may disadvantage rival AI providers.

Brazil’s Administrative Council for Economic Defense, known as CADE, has reopened an investigation into Google’s use of news content, including in AI Overviews. The new proceeding follows an earlier case that began in 2019 and ended in 2024.[62]

CADE’s current theory portrays publishers as structurally dependent on Google and treats the platform as an essential intermediary capable of imposing unilateral terms. Yet the agency’s earlier economic analysis reached the opposite conclusion.

In the prior investigation, CADE’s Department of Economic Studies found no support for the proposed theories of harm.[63] It found no incentive to exclude rivals, no basis for an essential-facilities claim, and a positive net flow of traffic from Google to publishers. It also found insufficient evidence that snippets caused users to stop visiting publishers’ websites. In 2024, CADE’s General Superintendence likewise recommended dismissal for lack of evidence of an antitrust violation.[64]

The United Kingdom has taken a regulatory route. Acting under the country’s digital-markets regime, the Competition and Markets Authority, known as the CMA, imposed several conduct requirements on Google.[65]

The CMA requires Google to give publishers control over whether their content grounds responses in generative-AI search features such as AI Overviews. Publishers must also be able to opt out of having their content used for model training, grounding, and fine-tuning.[66]

The CMA later added requirements governing the ranking of search results. Google must rank organic results according to objective and nondiscriminatory criteria, including within AI Overviews.[67]

The CMA initially proposed an opt-out limited to model training. After consultation, it expanded the requirement to include fine-tuning. Publishers argued that fine-tuning also uses their content and that excluding it would leave room for circumvention. The CMA agreed that control over fine-tuning could strengthen publishers’ bargaining position and prevent Google from avoiding the rule.[68]

The CMA insists that these requirements operate independently of copyright law. That distinction is, however, difficult to maintain. A competition-law opt-out directly affects how rights holders control protected material, how licensing negotiations proceed, and how future copyright reforms may operate.[69]

The disputes over online news and AI Overviews therefore follow the same pattern. Publishers first seek stronger copyright protection, then turn to competition law when copyright does not produce the desired bargaining outcome.[70] Authorities have repeatedly accommodated those demands by expanding copyright, using antitrust to reinforce it, or imposing antitrust remedies before copyright law has resolved the underlying rights.

V. Competition Law Should Focus on Rival AI Providers

As litigation and policy debates over copyright and artificial intelligence expand worldwide, the turn toward antitrust in disputes between AI providers and publishers targets the wrong market risk. Publishers have struggled to develop a sustainable business model since the rise of the internet, despite repeated policy interventions. Artificial intelligence now affects far more than publishing. Assistive and agentic AI are changing web browsing, online search, e-commerce, and other forms of digital intermediation while giving new firms a chance to challenge established incumbents.

Competition law should therefore examine whether incumbent firms use their existing market power to impede new AI providers. It should not sidestep unresolved copyright questions concerning the use of protected works in AI training and related applications.[71]

The launch of AI Overviews presents no inherent competition problem. It is a new product feature. The more plausible concern arises from Google’s decision to integrate it into a search engine in which Google has long held a dominant position.

Google places AI Overviews above organic results at the top of the search-engine results page. That placement favors Google’s own AI-generated summaries and may reduce users’ willingness to seek similar services from competing providers. Integration may therefore serve a defensive strategy by protecting Google Search against assistive and agentic AI services that could change how users obtain information and weaken the role of traditional search engines.[72]

AI assistants and agents are becoming significant competitive platforms. They increasingly allow users to reach third-party services without leaving the chat interface, rather than operating only as standalone question-and-answer tools.[73] AI Overviews shows how these services may reduce reliance on traditional intermediaries and redistribute revenue among firms. The growth of ChatGPT and Claude also shows that AI can help new entrants challenge incumbent platforms.

The main competition concern therefore stems from tying AI Overviews to Google Search. Google may use its search dominance to favor its own AI summaries, divert users from rival AI providers, and make entry more difficult. Standalone AI services may lack any comparable means of reaching users.

Competition authorities should still proceed carefully. New AI firms may be able to build integrated networks of products and services around standalone generative-AI offerings without owning an established search engine.[74] Authorities must therefore consider competition within each firm’s suite of services alongside competition among rival AI providers. An assessment of Google’s use of AI within its own services should account for competitive pressure from independent AI firms.[75]

Google also faces obligations under the Digital Markets Act (DMA) because the European Union has designated it as a gatekeeper.[76] AI services do not currently appear among the DMA’s listed core platform services. Yet AI functions incorporated into a gatekeeper’s existing services may still fall within the regulation’s reach.

The European Commission recently required Google to provide interoperability with important Android functions so that third-party AI providers can compete with Alphabet’s AI services, including Gemini.[77] The DMA does not impose comparable obligations on new AI firms, regardless of their growth or commercial importance, because standalone AI applications do not fall within the regulation’s enumerated categories.

Until courts or legislators resolve the relevant copyright questions, competition policy should give greater attention to the horizontal relationship between Google and rival AI assistants and agents. The possible exclusionary effects of integrating AI Overviews into Search warrant closer scrutiny than the vertical relationship between Google and publishers.

Competition law also cannot guarantee payment or negotiations for publishers whenever AI-generated summaries reduce traffic. At least three scenarios demonstrate the limits of antitrust intervention.

First, new AI providers may offer summaries without holding a dominant position. Their services may harm publishers as much as AI Overviews, as the growing body of copyright litigation suggests. Yet their lack of market power leaves little basis for an antitrust claim.

Second, Google could offer the same summarization service through Gemini without integrating it into Search. If Google lacks dominance in the market for AI assistants and agents, competition law would again provide no sound basis for intervention merely because publishers lose traffic.

Third, antitrust law may offer little help even when Google integrates AI Overviews into Search. Publishers criticize Google’s tailored licensing agreements with major media companies because those deals benefit only selected firms. Yet Google generally remains free to obtain content from publishers it considers especially valuable or reliable. Selective contracting does not, by itself, establish an antitrust violation.

These scenarios show that publishers’ claims primarily concern the scope and enforcement of copyright. Expanding antitrust law to secure payment for publishers would often fail to address their complaints and would distract from the more serious competition question concerning Google’s treatment of rival AI providers.

VI. Conclusion

Search engines have long determined how attention flows across the web. The bargain with publishers was imperfect but clear. Search engines captured value when users entered queries, while publishers supplied the underlying information and earned revenue when users clicked through to their websites.

AI-generated summaries are changing that arrangement. They can answer users’ questions within the search interface, weakening the link among search, referral traffic, and content production. Early evidence suggests that AI Overviews reduces clicks to outside websites, giving publishers’ concerns more empirical support than their earlier claims about snippets and news aggregation.

At the same time, artificial intelligence threatens Google’s traditional role as the web’s main gateway. Standalone services such as ChatGPT and Claude allow users to obtain information and reach third-party services without beginning with a conventional search engine. AI is therefore changing both the relationship between platforms and publishers and the competition between Google and new AI providers.

Copyright law should address the first problem. Courts and lawmakers have yet to determine when training, fine-tuning, grounding, and inference infringe protected rights or fall within existing exceptions. Antitrust authorities should not prejudge those questions by creating opt-outs, imposing negotiation duties, or engineering compensation through competition remedies.

Such remedies would also fail to solve publishers’ broader problem. Competition law offers little help when a nondominant AI entrant generates summaries, when Google offers the same service through a standalone AI product, or when Google contracts with only selected publishers. Those scenarios may reduce traffic, but they do not necessarily involve exclusionary conduct or market power.

Competition law has a more appropriate task. Authorities should examine whether Google’s integration of AI Overviews into its dominant search service favors Gemini, diverts users from rival AI assistants, and impedes entry by firms that lack a comparable distribution channel. That inquiry should focus on evidence of competitive harm rather than treat publisher compensation as an antitrust objective.

The rise of AI should not produce another cycle in which copyright disappoints publishers and antitrust is pressed into service as a substitute. Copyright law should determine whether AI providers may use publishers’ works and on what terms. Competition law should protect the rivalry that may decide whether Google remains the dominant gateway to online information.

[1] See, e.g., News/Media Alliance, Alliance President & CEO on Google Incorporating AI Tool Gemini into Search Engine: “This Will Be Catastrophic to Our Traffic” (May 15, 2024), https://www.newsmediaalliance.org/alliance-president-ceo-on-google-incorporating-ai-tool-gemini-into-search-engine-this-will-be-catastrophic-to-our-traffic.

[2] Nilay Patel, Google CEO Sundar Pichai on AI-Powered Search and the Future of the Web, The Verge (May 20, 2024), https://www.theverge.com/24158374/google-ceo-sundar-pichai-ai-search-gemini-future-of-the-internet-web-openai-decoder-interview.

[3] For a recent comparison of the proposed solutions and regulatory approaches, see, e.g., Giuseppe Colangelo, News Publishers, Digital Platforms, and Bargaining Codes: Debunking the Free-Riding Myth, 75 GRUR Int’l 19 (2026).

[4] Directive 2019/790, of the European Parliament and of the Council of 17 April 2019 on Copyright and Related Rights in the Digital Single Market and Amending Directives 96/9/EC and 2001/29/EC, art. 15, 2019 O.J. (L 130) 92. For commentary, see, e.g., Caterina Sganga, The Many Metamorphoses of Related Rights in EU Copyright Law: Unintended Consequences or Inevitable Developments?, 70 GRUR Int’l 821 (2021); Giuseppe Colangelo & Valerio Torti, Copyright, Online News Publishing and Aggregators: A Law and Economics Analysis of the EU Reform, 27 Int’l J.L. & Info. Tech. 75 (2019); Christophe Geiger, Oleksandr Bulayenko & Giancarlo Frosio, The Introduction of a Neighbouring Right for Press Publishers at EU Level: The Unneeded (and Unwanted) Reform, 39 Eur. Intell. Prop. Rev. 202 (2017).

[5] See Autorité de la concurrence, Decision No. 26-MC-02 (July 8, 2026), https://www.autoritedelaconcurrence.fr/en/article/related-rights-autorite-de-la-concurrence-imposes-interim-measures-and-orders-meta (imposing interim measures and ordering Meta to negotiate in good faith with press agencies and publishers). For a critical analysis, see Giuseppe Colangelo, Enforcing Copyright Through Antitrust? The Strange Case of News Publishers Against Digital Platforms, 10 J. Antitrust Enf’t 133 (2022).

[6] Treasury Laws Amendment (News Media and Digital Platforms Mandatory Bargaining Code) Act 2021 (Cth) (Austl.). For an overview of the countries that drew on the Australian model, see Colangelo, supra note 3.

[7] Press Release, Eur. Comm’n, Commission Opens Investigation into Possible Anticompetitive Conduct by Google in the Use of Online Content for AI Purposes (Dec. 8, 2025), https://ec.europa.eu/commission/presscorner/detail/da/ip_25_2964.

[8] Press Release, Admin. Council for Econ. Def., CADE’s Tribunal Recommends the Opening of an Investigation into Google Regarding the Use of Journalistic Content (May 20, 2026), https://www.gov.br/cade/en/matters/news/cade2019s-tribunal-recommends-the-opening-of-an-investigation-into-google-regarding-the-use-of-journalistic-content.

[9] Japan Fair Trade Comm’n, Follow-Up Study on the News Content Distribution Sector (Dec. 24, 2025), https://www.jftc.go.jp/en/about_jftc/index_3_251224.html.

[10] Digital Markets, Competition and Consumers Act 2024, c. 13 (UK).

[11] Press Release, Competition & Mkts. Auth., CMA Secures Fairer Deal for Publishers and Improves Google Search Services in UK (June 3, 2026), https://www.gov.uk/government/news/cma-secures-fairer-deal-for-publishers-and-improves-google-search-services-in-uk. Training, fine-tuning, and grounding describe three distinct ways an artificial-intelligence model acquires or uses information. Training builds the model’s general knowledge and language capabilities from a large corpus, producing the weights that encode what it has learned. Fine-tuning continues training on a smaller, targeted dataset to adjust the model’s behavior, tone, or domain-specific performance, but it does not reliably teach new facts. Grounding leaves the weights unchanged and instead connects the model to external, often real-time, information when it generates a response, improving accuracy and reducing fabrication.

[12] See, e.g., Susan Athey, Markus Mobius & Jeno Pal, The Impact of Aggregators on Internet News Consumption (NBER Working Paper No. 28746, 2021); Joan Calzada & Ricard Gil, What Do News Aggregators Do? Evidence from Google News in Spain and Germany, 39 Mktg. Sci. 134 (2020); Lesley Chiou & Catherine Tucker, Content Aggregation by Platforms: The Case of the News Media, 26 J. Econ. & Mgmt. Strategy 782 (2017).

[13] Case C-250/25, Like Co. v. Google Ireland Ltd. (Ct. J. Eur. Union).

[14] Sundar Pichai, I/O 2026: Welcome to the Agentic Gemini Era, News from Google (May 19, 2026), https://blog.google/innovation-and-ai/sundar-pichai-io-2026/#more-news; Elizabeth Reid, A New Era for AI Search, News from Google (May 19, 2026), https://blog.google/products-and-platforms/products/search/search-io-2026/#search-agents.

[15] See, e.g., Org. for Econ. Coop. & Dev., Artificial Intelligence and Competitive Dynamics in Downstream Markets 25 (Nov. 14, 2025), https://www.oecd.org/en/publications/artificial-intelligence-and-competitive-dynamics-in-downstream-markets_ccf0624a-en.html.

[16] Nicolas Padilla, H. Tai Lam, Anja Lambrecht & Brett Hollenbeck, The Impact of LLM Adoption on Online User Behavior (Jan. 2, 2026), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5393256. See also Gordon Burtch, Dokyun Lee & Zhichen Chen, The Consequences of Generative AI for Online Knowledge Communities, 14 Sci. Reps. 10413 (2024); R. Maria del Rio-Chanona, Nadzeya Laurentsyeva & Johannes Wachs, Large Language Models Reduce Public Knowledge Sharing on Online Q&A Platforms, 3 PNAS Nexus 400 (2024) (documenting ChatGPT’s effect on user activity and finding that activity on Stack Overflow declined after ChatGPT’s release).

[17] Qiaoni Shi, Kai Zhu & Kai Gu, Answering Without Referring: How AI Search Rewrites the Web’s Economic Bargain (July 8, 2026) (arXiv preprint), https://arxiv.org/abs/2607.07652v1.

[18] Samira Gholami, Cristiana Firullo, Cristobal Cheyre & Alessandro Acquisti, Beyond Search: LLM Adoption and Web Traffic Concentration (Feb. 21, 2026), https://ssrn.com/abstract=6238578.

[19] Rand Fishkin, In 2026, Less Than One Third of Google Searches Still Send a Click, SparkToro (June 8, 2026), https://sparktoro.com/blog/in-2026-less-than-one-third-of-google-searches-still-send-a-click; Ryan Law & Xibeijia Guan, Update: AI Overviews Reduce Clicks by 58%, Ahrefs Blog (Feb. 4, 2026), https://ahrefs.com/blog/ai-overviews-reduce-clicks-update; Athena Chapekis & Anna Lieb, Google Users Are Less Likely to Click on Links When an AI Summary Appears in the Results, Pew Rsch. Ctr. (July 22, 2025), https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results.

[20] Fishkin, supra note 19.

[21] Saharsh Agarwal & Ananya Sen, Google AI Overviews and Publisher Traffic: Evidence from a Field Experiment (July 8, 2026), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6513059; Haofei Xu, Umar Iqbal & Jacob M. Montgomery, Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact (May 13, 2026) (arXiv preprint), https://arxiv.org/abs/2605.14021.

[22] Mehrzad Khosravi & Hema Yoganarasimhan, Impact of AI Search Summaries on Website Traffic: Evidence from Google AI Overviews and Wikipedia (May 12, 2026) (arXiv preprint), https://arxiv.org/abs/2602.18455.

[23] Hangcheng Zhao & Ron Berman, Strategic Response of News Publishers to Generative AI (Apr. 21, 2026), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5992774.

[24] Riley Grossman, Songjiang Liu, Michael K. Chen, Mike Smith, Christian Borcea & Yi Chen, How Generative AI Disrupts Search: An Empirical Study of Google Search, Gemini, and AI Overviews (Apr. 30, 2026) (arXiv preprint), https://arxiv.org/abs/2604.27790.

[25] Jana Garanko, Semrush AI Overviews Study: What 2025 SEO Data Tells Us About Google’s Search Shift, Semrush (Dec. 15, 2025), https://www.semrush.com/blog/semrush-ai-overviews-study.

[26] Peibo Zhang, Ruomeng Cui & Dennis J. Zhang, The Impact of AI Search on the Online Content Ecosystem: Evidence from Google and Reddit (June 20, 2026) (arXiv preprint), https://arxiv.org/abs/2605.16428.

[27] See, e.g., Ariel Ezrachi & Giuseppe Colangelo, AI Ecosystems, Power Shifts, and EU Competition Law Enforcement (June 11, 2026), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6841798.

[28] See, e.g., Alexander Peukert, Copyright in the Artificial Intelligence Act—A Primer, 73 GRUR Int’l 497 (2024); Pamela Samuelson, Fair Use Defenses in Disruptive Technology Cases, 71 UCLA L. Rev. 1484 (2024); Mark A. Lemley & Bryan Casey, Fair Learning, 99 Tex. L. Rev. 743 (2021); Carys Craig & Ian Kerr, The Death of the AI Author, 52 Ottawa L. Rev. 31 (2020); Daniel J. Gervais, The Machine as Author, 105 Iowa L. Rev. 2053 (2020).

[29] Christian Peukert & Margaritha Windisch, The Economics of Copyright in the Digital Age, 39 J. Econ. Survs. 877 (2025).

[30] Joshua Gans, Shane Greenstein, Adam Jaffe, Brent Lutes, Abhishek Nagaraj, Imke Reimers, Michael D. Smith, Rahul Telang, Catherine Tucker & Joel Waldfogel, Identifying the Economic Implications of Artificial Intelligence for Copyright Policy: Context and Direction for Economic Research 10–13 (U.S. Copyright Off., Feb. 12, 2025), https://www.copyright.gov/newsnet/2025/1062.html?loclr=licop.

[31] See, e.g., Samuel G. Goldberg & H. Tai Lam, Generative AI & Creative Goods: Market Expansion, Crowd-Out, and Copyright (Mar. 9, 2026), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5152649 (finding that generative AI expands markets and improves quality while substantially crowding out non-generative AI production).

[32] See, e.g., Nicola Lucchi, The Impact of Google AI Summaries and Google AI Overviews on Publishers’ Revenue and Media Freedom (Eur. Parliament, Briefing No. PE 787.211, Apr. 22, 2026), https://www.europarl.europa.eu/thinktank/en/document/IUST_BRI(2026)787211.

[33] See, e.g., OECD, News in the Internet Age: New Trends in News Publishing (Oct. 18, 2010), http://dx.doi.org/10.1787/9789264088702-en; Fed. Trade Comm’n Staff, Potential Policy Recommendations to Support the Reinvention of Journalism (Oct. 18, 2010), https://www.ftc.gov/sites/default/files/documents/public_events/how-will-journalism-survive-internet-age/new-staff-discussion.pdf.

[34] See, e.g., Stigler Ctr. for the Study of the Econ. & the State, Protecting Journalism in the Age of Digital Platforms (July 1, 2019), https://www.chicagobooth.edu/-/media/research/stigler/pdfs/media—report.pdf.

[35] Directive 2019/790, supra note 4, art. 15.

[36] Id. art. 4.

[37] Lucchi, supra note 32.

[38] For an overview of the scholarly debate, see, e.g., Martin Senftleben, Text and Data Mining, Generative AI, and the Copyright Three-Step Test, 57 IIC 67 (2026); Tim W. Dornis, The Training of Generative AI Is Not Text and Data Mining, 47 Eur. Intell. Prop. Rev. 65 (2025); Eleonora Rosati, Is Text and Data Mining Synonymous with AI Training?, 19 J. Intell. Prop. L. & Prac. 851 (2024); Thomas Margoni & Martin Kretschmer, A Deeper Look into the EU Text and Data Mining Exceptions: Harmonisation, Data Ownership, and the Future of Technology, 71 GRUR Int’l 685 (2022).

[39] Eur. Comm’n, Call for Tenders for a Feasibility Study on a Central Registry of Opt-Outs Under the Text and Data Mining (TDM) Exception (Jan. 23, 2025), https://digital-strategy.ec.europa.eu/en/funding/call-tenders-feasibility-study-central-registry-opt-outs-under-text-and-data-mining-tdm-exception.

[40] Id.

[41] Regulation 2024/1689, of the European Parliament and of the Council of 13 June 2024 Laying Down Harmonised Rules on Artificial Intelligence and Amending Regulations (EC) No. 300/2008, (EU) No. 167/2013, (EU) No. 168/2013, (EU) 2018/858, (EU) 2018/1139, and (EU) 2019/2144 and Directives 2014/90/EU, (EU) 2016/797, and (EU) 2020/1828 (Artificial Intelligence Act), recital 105, art. 53(1)(c), 2024 O.J. (L 1689) 1.

[42] Eur. Comm’n, Directorate-Gen. for Commc’ns Networks, Content & Tech., Visionary Analytics & Capgemini Invent, Study to Assess the Feasibility of a Registry of Text and Data Mining Opt-Outs Expressed by Rightholders: Final Report (2026), https://data.europa.eu/doi/10.2759/4959861. The study finds that existing text-and-data-mining opt-out mechanisms are fragmented, inconsistently implemented, and often ineffective. Rightholders must combine several tools without assurance that developers will detect or honor their reservations, while developers face inconsistent and sometimes conflicting signals across web-scale automated systems. The study therefore proposes an EU registry using digital-fingerprinting technology to provide persistent, unambiguous, and machine-readable opt-out signals. The registry would supplement, rather than replace, existing sector-specific systems.

[43] See, e.g., Peter Mezei, Martin Kretschmer, Thomas Margoni, Alexander Peukert & João Pedro Quintais, Comment of the European Copyright Society on the Request for Preliminary Ruling in Case C-250/25 (Like Company), 57 IIC 850 (2026); Tim W. Dornis & Nicola Lucchi, Generative AI and the Scope of EU Copyright Law: A Doctrinal Analysis in Light of the Referral in Like Company v. Google, 56 IIC 1800 (2025).

[44] Eur. Comm’n, Report on the Review of the Copyright in the Digital Single Market Directive/Targeted Initiative for a Better Copyright Environment for European Creativity and Innovation, Call for Evidence, Ares(2026)4845636 (2026), https://ec.europa.eu/info/law/better-regulation/have-your-say/initiatives/18173-Targeted-initiative-for-a-better-copyright-environment-for-European-creativity-and-innovation_en.

[45] Id.

[46] Like Co., Case C-250/25. But see Mezei et al., supra note 43, at 850 (questioning the judgment’s relevance because the reference conflates chatbots, large language models, and search engines—distinct technologies and services implicating different rights under European Union law—and inconsistently identifies the right at issue: the press publishers’ right under Article 15 of the Copyright in the Digital Single Market Directive).

[47] Directive 2001/29, of the European Parliament and of the Council of 22 May 2001 on the Harmonisation of Certain Aspects of Copyright and Related Rights in the Information Society, arts. 2–3, 2001 O.J. (L 167) 10.

[48] For an analysis of these laws, see Colangelo, supra note 3.

[49] See Sénat, Proposition de loi visant à renforcer l’effectivité des droits voisins des éditeurs et des agences de presse (2026), https://www.senat.fr/travaux-parlementaires/textes-legislatifs/la-loi-en-clair/proposition-de-loi-visant-a-renforcer-leffectivite-des-droits-voisins-des-editeurs-et-des-agences-de-presse.html.

[50] Case C-797/23, Meta Platforms Ireland Ltd. v. Autorità per le Garanzie nelle Comunicazioni, ECLI:EU:C:2026:395 (May 12, 2026). Belgium’s implementation is also under review. See Case C-663/24, Google Ireland Ltd. et al. v. Premier ministre et al. (Ct. J. Eur. Union).

[51] See, e.g., Stigler Ctr., supra note 34; Frances Cairncross, The Cairncross Review: A Sustainable Future for Journalism (Feb. 12, 2019), https://www.gov.uk/government/publications/the-cairncross-review-a-sustainable-future-for-journalism; Damien Geradin, Complements and/or Substitutes? The Competitive Dynamics Between News Publishers and Digital Platforms and What It Means for Competition Policy (TILEC Discussion Paper No. 2019-003, 2019), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3338941.

[52] See, e.g., Austl. Competition & Consumer Comm’n, Digital Platforms Inquiry: Final Report 232 (Mar. 27, 2019), https://www.accc.gov.au/inquiries-and-consultations/finalised-inquiries-and-monitoring/digital-platforms-inquiry-2017-19.

[53] See, e.g., Madhavi Singh & Fiona M. Scott Morton, A Roadmap for a Monopolization Case Against Google: Monopsony Power and AI Overviews, 174 U. Pa. L. Rev. (forthcoming 2026); CMA, Google’s General Search Services: Proposed Conduct Requirements—Introduction to the Consultation ¶ 2.8 (Jan. 28, 2026), https://www.gov.uk/government/consultations/googles-general-search-services-proposed-conduct-requirements.

[54] CMA, supra note 53, ¶ 2.8.

[55] In response to French rules establishing criteria for compensating publishers for their neighboring rights in online news, Google announced that it would stop displaying article excerpts, photographs, infographics, and videos across its services unless publishers authorized their use free of charge.

[56] See Autorité de la concurrence, Decision No. 20-MC-01, Requests for Interim Measures by the Syndicat des Éditeurs de la Presse Magazine, the Alliance de la Presse d’Information Générale and Others, and Agence France-Presse (Apr. 9, 2020), https://www.autoritedelaconcurrence.fr/en/decision/requests-interim-measures-syndicat-des-editeurs-de-la-presse-magazine-alliance-de-la. The Autorité issued its decision on the merits in 2022. See Autorité de la concurrence, Decision No. 22-D-13, Practices Implemented in the Press Sector (June 21, 2022), https://www.autoritedelaconcurrence.fr/en/decision/regarding-practices-implemented-press-sector.

[57] Cour d’appel [CA] [regional court of appeal] Paris, Oct. 8, 2020, Google LLC v. Syndicat des Éditeurs de la Presse Magazine (Fr.).

[58] Autorité de la concurrence, supra note 5.

[59] Anna Ferrari, French Google Case Sets Example on How Commitments Can Catch New AI Use of Publishers’ Content, MLex (May 12, 2026), https://www.mlex.com/mlex/articles/2476168/french-google-case-sets-example-on-how-commitments-can-catch-new-ai-use-of-publishers-content.

[60] Autorité de la concurrence, Decision No. 24-D-03 (Mar. 15, 2024), https://www.autoritedelaconcurrence.fr/en/decision/regarding-compliance-commitments-decision-22-d-13-21-june-2022-autorite-de-la-concurrence.

[61] Eur. Comm’n, supra note 7.

[62] CADE, supra note 8.

[63] Departamento de Estudos Econômicos do Conselho Administrativo de Defesa Econômica, Nota Técnica No. 24/2023/DEE/CADE (2023), https://sei.cade.gov.br/sei/modulos/pesquisa/md_pesq_documento_consulta_externa.php?HJ7F4wnIPj2Y8B7Bj80h1lskjh7ohC8yMfhLoDBLddY6IZhQWFlRDmST1RQcspnlW1x9t6x01LadVtOKNionuQ960Q1Gg5heKwC5g9UFt-RzfyXxgWsFNjti4bd-TUFf.

[64] Superintendência-Geral do Conselho Administrativo de Defesa Econômica, Nota Técnica No. 70/2024/CGAA11/SGA1/SG/CADE (2024), https://sei.cade.gov.br/sei/modulos/pesquisa/md_pesq_documento_consulta_externa.php?HJ7F4wnIPj2Y8B7Bj80h1lskjh7ohC8yMfhLoDBLddYhYzgDyOUJ75GboeN2PJR3JNjn6Nmo8xmXxl5MS5KRnf8CuaeR-PEedrWBaKF0nh4kBOawa7f3VoD5kHYjPmTF.

[65] CMA, supra note 11. The other measures address fair ranking, choice screens, and data portability.

[66] Although publishers have broadly supported the opt-out remedy, some scholars question its effectiveness, arguing that it will neither stop traffic diversion nor give publishers meaningful new choices in dealing with Google Search. See, e.g., Spencer Cohen & Todd Davies, Opt-Out Remedies Will Not Fix AI Overviews (forthcoming), J. Eur. Competition L. & Prac.; Singh & Scott Morton, supra note 53.

[67] CMA, Further CMA Action to Secure a Fairer Deal for Businesses and Improve Google Search Services in UK (June 17, 2026), https://www.gov.uk/government/news/further-cma-action-to-secure-a-fairer-deal-for-businesses-and-improve-google-search-services-in-uk.

[68] CMA, supra note 11.

[69] See, e.g., Magali Eben, Bargaining Through Competition Law? The Publishers Saga Continues in the UK with CMA Conduct Requirements Imposed on Google, CREATe (June 23, 2026), https://www.create.ac.uk/working-papers/2026/06/23/new-working-paper-bargaining-through-competition-law-the-publishers-saga-continues-in-the-uk-with-cma-conduct-requirements-imposed-on-google.

[70] Id.

[71] Similarly, see Cohen & Davies, supra note 66 (arguing that AI Overviews principally harm competition not by using publishers’ content without compensation, but by keeping traffic on Google’s own website).

[72] See, e.g., Padilla et al., supra note 16.

[73] See, e.g., Autorité de la concurrence, Conversational Agents: The Autorité Starts Inquiries Ex Officio with a View to Issuing an Opinion (2026), https://www.autoritedelaconcurrence.fr/en/press-release/conversational-agents-autorite-starts-inquiries-ex-officio-view-issuing-opinion.

[74] Ezrachi & Colangelo, supra note 27. See also Taiwan Fair Trade Comm’n, Public Consultation Report and Policy Statement on Generative Artificial Intelligence and Competition (May 29, 2026), https://www.ftc.gov.tw/internet/english/doc/docDetail.aspx?uid=179&docid=18369 (suggesting that generative AI will increasingly develop through integrated ecosystems).

[75] Ezrachi & Colangelo, supra note 27.

[76] Regulation 2022/1925, of the European Parliament and of the Council of 14 September 2022 on Contestable and Fair Markets in the Digital Sector and Amending Directives (EU) 2019/1937 and (EU) 2020/1828 (Digital Markets Act), 2022 O.J. (L 265) 1.

[77] Press Release, Eur. Comm’n, Commission Provides Guidance to Google for AI Interoperability on Android and Sharing of Google Search Data Under the Digital Markets Act (July 16, 2026), https://ec.europa.eu/commission/presscorner/api/files/document/print/en/ip_26_1634/IP_26_1634_EN.pdf.

Introduction

We thank the Federal Trade Commission (“FTC” or “Commission”) for the opportunity to comment on its proposed Policy Statement Concerning the Suppression of Accuracy in Artificial Intelligence Systems (“AI Policy Statement” or “Statement”). The comment period gives the public an opportunity to inform the Commission’s consideration of how the First Amendment constrains government regulation of artificial intelligence (“AI”).^[1]

The International Center for Law & Economics (“ICLE”) is a nonprofit, nonpartisan research and policy center dedicated to developing economically grounded approaches to law and public policy. ICLE applies law & economics methods and economic research to policy debates and has longstanding expertise in competition, consumer protection, innovation, and technology policy.

ICLE has a strong interest in ensuring that First Amendment and consumer protection law serve the public interest through clear legal rules and sound economic analysis. ICLE scholars have written extensively about the regulation of online speech platforms, free expression, competition, and consumer protection. They have also examined how the FTC’s competition and consumer protection missions complement one another and how excessive consumer protection regulation—even when well intentioned—can impede innovation and competition.[2] The AI Policy Statement recognizes several of these concerns. ICLE is therefore well positioned to address the legal and economic issues raised by the Statement.

The FTC gets several important points right.

First, uniform federal AI policies may protect consumers and competition more effectively than a patchwork of state regulations. The Statement correctly calls for a “national AI framework” that would “protect innovation and competition by providing national regulatory clarity and certainty and avoiding a balkanized or patchwork regulatory approach driven by States—or, most dangerously, imposed by certain anti-innovation State governments on the rest of the country.”[3]

Second, excessive AI regulation at either the federal or state level can impede innovation, weaken competition, and harm consumers across many markets.[4] Such regulation may also raise concerns beyond antitrust and the Commission’s competition and consumer protection authority.[5]  Poorly designed rules can increase consumers’ information costs by restricting access to truthful, non-misleading speech.

Third, the Commission correctly recognizes that AI providers generally remain subject to the Federal Trade Commission Act. As the Statement explains, AI is “an umbrella term covering a universe of different tools and systems.”[6] It includes a broad and evolving range of products and services. AI providers generally do not fall within the categories of persons, partnerships, or corporations excluded from the Commission’s authority under Section 5(a)(2). Nor do they typically qualify for the bona fide nonprofit exemption under Section 4.[7]

The proposed Statement nevertheless requires substantial revision before the Commission finalizes it.

First, the Statement offers little practical guidance about how the Commission will apply its deception authority to AI. Except for a few straightforward examples, it does not identify which acts or practices the Commission would consider violations of Section 5.

Second, the Statement largely avoids the difficult question of when the conduct of AI providers, or their interactions with users, constitutes speech protected by the First Amendment. It also does not explain how those protections would constrain FTC enforcement. The Statement’s focus on “ideologically motivated distortions” suggests that the Commission’s concerns extend beyond factual misrepresentations in marketing to speech that may receive the highest degree of First Amendment protection.[8]

Third, the Statement’s discussion of federal preemption requires substantial clarification. We understand that the Commission addresses preemption at least partly in response to Executive Order 14365, Ensuring a National Policy Framework for Artificial Intelligence.[9] Even accounting for the executive order’s requirements, the Statement’s discussion is so brief that it may mislead regulated parties and courts.

The Commission may intend to say that compliance with state consumer protection laws governing AI products and services would not provide an affirmative defense to liability under Section 5. If so, a more clearly explained and illustrated version of that point could provide useful guidance. Established forms of implied preemption, though, fit poorly with the FTC Act and settled principles of federalism.

Courts have not held that the FTC Act occupies the field of state advertising, marketing, or related consumer protection law so completely that it leaves no room for state regulation.[10] Conflict preemption also appears inapposite. We are unaware of any relevant state AI law or regulation—however sound or unsound as policy—that makes simultaneous compliance with state law and the FTC Act impossible.[11] As Part III explains, obstacle preemption appears even less applicable because state AI laws do not necessarily frustrate the purposes of the FTC Act.[12] In the end, the Statement’s current discussion is too vague to provide the clear and useful explanation contemplated by the executive order.

These comments explain why constitutional law and economic analysis impose meaningful limits on FTC enforcement under the proposed Statement. The Commission should provide clearer guidance about how it will apply its deception authority within those limits.

Part I explains the First Amendment principles that protect the “marketplace of ideas.” Those principles protect, to varying degrees, both the editorial discretion of AI companies that train and develop large language models (“LLMs”) and the rights of users who interact with them. Market forces also discipline AI companies whose systems repeatedly provide false or unreliable responses.

Part II examines the difficulty of applying the Commission’s deception authority to generative AI providers based on disfavored speech. Section 5 permits the Commission to police deceptive acts or practices, but liability ordinarily requires a material representation, omission, or promise likely to mislead reasonable consumers. We examine the representations made by major AI providers and conclude that the Commission may lack a viable basis for enforcement under the proposed Statement.

Part III analyzes the Statement’s treatment of preemption. State AI laws generally do not make compliance with the FTC Act impossible, as conflict preemption requires. The Statement is also agency guidance and therefore is unlikely to receive substantial judicial deference.

I. The First Amendment Protects AI Editorial Judgment and User Access[13]

The defense of free speech against government censorship has a long pedigree. Its advocates include John Stuart Mill and, long before the Bill of Rights, John Milton.[14] Although neither used the precise phrase, their defenses of free expression helped establish the theory later described as the “marketplace of ideas.” That theory bears directly on the AI Policy Statement because AI chatbots have become a new forum for producing, receiving, and testing ideas. Complaints about whether chatbots are ideologically neutral[15] must therefore account for the First Amendment.

Mill offered four grounds for protecting free speech. A censored opinion may be true. An erroneous opinion may contain part of the truth, which can emerge only through confrontation with opposing views. A true opinion must still face challenge so that those who hold it understand its basis. And even a true opinion can become an inherited dogma rather than a genuine conviction if no one contests it.[16] The exchange of competing ideas allows truth to emerge and remain meaningful.

Thomas Jefferson advanced a similar principle in his First Inaugural Address after a divisive election. He urged tolerance for divergent political views and argued that even those who favored dissolving the Union or changing its republican form should remain “undisturbed as monuments of the safety with which error of opinion may be tolerated where reason is left free to combat it.”[17]

The Supreme Court later adopted the marketplace-of-ideas metaphor, initially through Justice Oliver Wendell Holmes’ dissenting opinions.[18] Variations of the phrase have since appeared in thousands of federal First Amendment decisions.[19]

The Constitution protects private parties’ authority to decide which speech they will disseminate, largely free from government control. As the Court explained in Manhattan Community Access Corp. v. Halleck, “[t]he Free Speech Clause of the First Amendment constrains governmental actors and protects private actors . . . .”[20]

Private parties often participate in the marketplace of ideas by creating expressive products. In Moody v. NetChoice,[21] the Supreme Court considered whether state laws regulating social-media platforms and other websites facially violated the First Amendment.^[22] The Court explained that the First Amendment protects an entity engaged in expressive activity, including one that compiles and curates others’ speech, when the government directs it to carry messages it would prefer to exclude.[23]

After reviewing its compelled-speech precedents, the Court concluded that government efforts to alter the views social-media platforms include or exclude from their principal feeds interfere with protected speech.[24] It also warned that, “[h]owever imperfect the private marketplace of ideas,” government decisions about whether speech is imbalanced, followed by coercion requiring more or less of particular views, offer “a worse proposal.”[25]

Generative AI companies likewise exercise protected editorial judgment. They select training materials, determine how models process those materials, and establish policies governing the outputs their systems generate. Government efforts to rebalance those choices would interfere with that editorial discretion. The Commission should therefore exercise caution before bringing an action against an AI provider for offering “ideologically motivated distortions in a response to a factual question.”[26]

Justice Amy Coney Barrett observed in her Moody concurrence that a “function qualifies for First Amendment protection only if it is inherently expressive.”[27] AI providers’ activities readily meet that standard. Selecting inputs, training models, and reviewing outputs for compliance with company policies all require editorial judgment. An AI chatbot’s answers are themselves an expressive product.[28]

AI providers resemble search engines, which also produce expressive responses to user queries. Several federal district courts have recognized that search engines possess a First Amendment interest in their results. Some courts have treated the results themselves as protected speech. Others have focused on the editorial judgment used to generate and arrange them. Under either approach, generative AI responses warrant similar protection.

In Jian Zhang v. Baidu.com,[29] the U.S. District Court for the Southern District of New York held that applying a New York public-accommodations law to a Chinese search engine accused of censoring pro-democracy speech would conflict with the search engine’s editorial discretion. The court explained that “there is a strong argument to be made that the First Amendment fully immunizes search-engine results from most, if not all, kinds of civil liability and government regulation.”[30]

The court also described the editorial judgments inherent in search. A search engine retrieves information from a vast body of online material, determines which information is relevant, organizes it for users, and decides where each result will appear.[31] Other courts have likewise recognized search engines’ right to exercise editorial discretion over their results.[32]

At least one court has treated search results themselves as protected opinions. In Search King Inc. v. Google Technology, Inc.,[33] the court explained that search results “are opinions—opinions of the significance of particular web sites as they correspond to a search query.” Different search engines may reach different conclusions because each uses its own method to determine relevance and significance.[34]

AI chatbot responses should receive no less protection than social-media feeds or search-engine results. AI providers exercise editorial judgment when they create, train, and maintain their systems, and the systems’ outputs are expressive. An allegation of ideological bias in an expressive product, standing alone, cannot supply a permissible basis for government enforcement.[35]

Users also have a First Amendment interest in receiving chatbot responses. The Supreme Court has held that the First Amendment protects the right to receive speech.[36] That protection covers listening to speakers,[37] reading pamphlets[38] and books,[39] receiving advertisements,[40] playing video games,[41] and using social media.[42] Interacting with ideas generated by an AI chatbot warrants the same protection.

The Commission is therefore correct to focus on its deception authority under Section 5 rather than its unfairness authority.[43] An unfairness claim would require the Commission to establish three elements. The practice must cause or be likely to cause substantial consumer injury. That injury must not be outweighed by countervailing benefits to consumers or competition. Consumers also must be unable reasonably to avoid the injury.[44]

A claim based on inaccurate or misleading chatbot responses would treat speech itself as the substantial injury. That theory conflicts with First Amendment precedent.

Even false noncommercial speech presumptively receives full First Amendment protection.[45] In United States v. Alvarez,^[46] a Supreme Court plurality explained that prior statements denying protection to false speech concerned only the limited “historic and traditional categories [of expression] long familiar to the bar.”^[47] The Court rejected a general First Amendment exception for false statements.^[48] A rule directed at false speech as such would therefore face strict scrutiny.

To avoid strict scrutiny, an unfairness claim would need an independent justification tied to a category of low-value speech that the Supreme Court has recognized as unprotected.[49] Those categories include defamation, fraud, false light, false statements to government officials, perjury, impersonating government officials, and speech integral to criminal conduct.^[50] A claim that inaccurate information alone constitutes substantial consumer injury would likely fail.

Alvarez illustrates the problem. The defendant falsely claimed at a public meeting that he had received the Congressional Medal of Honor.[51] The statement left “no room to argue about interpretation or shades of meaning.” Even so, the Court held that the restriction was content based and subject to strict scrutiny.[52] An FTC claim treating false or misleading chatbot output as a substantial injury would rest on similarly uncertain constitutional ground.

Even assuming a legally cognizable injury, the Commission would likely struggle to show that the injury outweighs consumer and competitive benefits. The First Amendment’s preference for less speech-restrictive alternatives would compound that difficulty.

AI companies participate in the marketplace of ideas by offering users what they regard as the best answers to their questions. Providers that repeatedly fail to satisfy users risk losing them to competing chatbots. Market discipline has already operated in prominent cases. When Google’s chatbot generated ideologically charged false outputs, substantial public criticism prompted the company to revise the product quickly.[53]

False information may also contribute to public debate by prompting correction and exposing weaknesses in competing claims. Suppression is rarely the least speech-restrictive response. As the Alvarez Court explained:

The remedy for speech that is false is speech that is true… Freedom of speech and thought flows not from the beneficence of the state but from the inalienable rights of the person. And suppression of speech by the government can make exposure of falsity more difficult, not less so. Society has the right and civic duty to engage in open, dynamic, rational discourse. These ends are not well served when the government seeks to orchestrate public discussion through content-based mandates.[54]

Competition among AI providers also gives consumers meaningful ways to avoid services they regard as inaccurate or ideologically slanted. The Statement recognizes that AI companies increasingly market their products as more accurate or ideologically neutral than competing systems.[55] Consumers can compare those claims, test competing products, and switch providers when a chatbot fails to meet their expectations.

II. Section 5 Deception Claims Require Concrete, Material Misrepresentations[56]

Quoting the “AI National Policy Fact Sheet,”[57] the AI Policy Statement states that “States such as California and Colorado are considering requiring AI companies to censor outputs and insert left-wing ideology in their programming.”[58] Such requirements may raise legitimate concerns because the First Amendment protects speakers against censorship and compelled speech.

The Statement, though, does not analyze the relevant state laws or explain how they might require AI providers to engage in conduct that qualifies as deception under Section 5. Its vague reference to “left-wing ideology” instead suggests concern about political viewpoints expressed in chatbot outputs. The First Amendment would generally protect such speech against FTC enforcement, regardless of the merits of the ideology expressed.

AI providers remain subject to the FTC Act, and the First Amendment does not protect every use of speech. The Supreme Court has long distinguished commercial speech, most commonly advertising, from other protected expression.[59] Commercial speech receives First Amendment protection, but courts generally afford it less protection than noncommercial speech.

The Commission may therefore bring careful, fact-specific deception cases involving false or misleading claims in advertising or marketing materials, consistent with its Policy Statement on Deception (“Deception Statement”).[60] Such enforcement also accords with decisions permitting the government to regulate fraud and other unlawful conduct carried out through speech.

Advertising and marketing materials may violate Section 5 when they expressly or implicitly make false or misleading factual claims about material attributes of AI products or services. The government has a legitimate interest in regulating such false commercial speech. Proper application of Section 5 under the Deception Statement can thereby protect consumers and competition against commercial fraud without intruding on protected expression.

Even where false speech falls outside First Amendment protection,[61] a fraud claim[62] requires more than proof of falsity.^[63] As the Supreme Court explained in Illinois ex rel. Madigan v. Telemarketing Assocs., “[s]imply labeling an action one for ‘fraud’ . . . will not carry the day.”[64] The Court has, for example, repeatedly invalidated prophylactic restrictions on charitable solicitation because they imposed prior restraints without requiring proof of fraud.[65]

A properly tailored fraud action places the full burden of proof on the government. A false statement alone does not establish liability. Exacting proof requirements are necessary to preserve sufficient breathing room for protected speech.[66]

The Commission must therefore distinguish commercial representations from fully protected expression before applying its deception authority to an AI provider. It should avoid guidance, and especially enforcement actions, that overlooks the First Amendment’s restrictions on government efforts to regulate or chill expressive speech.

The Deception Statement establishes demanding proof requirements even when the challenged speech is commercial.[67] It identifies three elements of deception.

First, “there must be a representation, omission or practice that is likely to mislead the consumer.”[68] Second, the Commission evaluates the conduct “from the perspective of a consumer [or specific group] acting reasonably in the circumstances.”[69] A claim is material when it is “likely to affect the consumer’s conduct or decision with regard to a product or service,” such that “consumer injury is likely, because consumers are likely to have chosen differently [in the marketplace] but for the deception.”[70]

The Commission must begin by identifying a representation, omission, or practice. Express claims about an AI product’s policies or capabilities may satisfy that requirement, particularly when they appear in consumer advertising. Implied claims and omissions may also be actionable, but the Commission should exercise caution before inferring what a reasonable consumer understands or expects in a new and rapidly changing market.

Some objective product claims may permit straightforward interpretation under established enforcement principles. Others may require substantial investigation, economic analysis, and consumer testing. Even in advertising cases, the Commission should account for its longstanding recognition that advertising can improve consumer welfare and competition by providing information and facilitating comparison.[71]

Section 5 enforcement can produce costs through both false positives and false negatives. An overbroad approach to deception may suppress useful speech and encourage AI providers to offer consumers less information. Commission guidance should therefore acknowledge the variety of AI products and uses that the Statement itself recognizes.

General assertions of liability should be carefully defined and supported. The Commission should examine what AI providers actually say in their policies, advertisements, and other marketing materials, along with the likely and demonstrable effects of those representations.

Published AI policies and marketing materials could, in principle, mislead reasonable consumers. The representations cited in the Statement, though, appear limited. Major providers do not promise certainty or perfect accuracy.[72] Instead, they expressly warn users that outputs may contain errors.

For example:

1. OpenAI states that “[a]ccuracy will never reach 100%.”[73] Its terms of use warn that outputs “may not always be accurate” and instruct users to evaluate them for “accuracy and appropriateness.”[74]
2. Anthropic’s consumer terms state that “[o]utputs may not always be accurate and may contain material inaccuracies even if they appear accurate.” They also instruct users not to rely on outputs or actions without independently confirming their accuracy.[75]
3. DeepSeek’s terms state that outputs “may contain errors or omissions and are for your reference only.” They warn users not to treat outputs as professional advice and explain that outputs may contain “incorrect, incomplete or inaccurate content.”[76]
4. Gab AI’s terms state that outputs “may not always be accurate” and place responsibility on users to evaluate outputs for accuracy and appropriateness. They also warn that the service may provide incomplete, incorrect, or offensive outputs.[77]
5. Grok’s terms state that outputs “may not always be accurate,” instruct users to conduct their own research, and warn them not to rely on outputs as truth. The terms also state that the service may provide incomplete, incorrect, or offensive outputs.[78]
6. Google warns that Gemini “may sometimes provide inaccurate or offensive content.” It also instructs users not to rely on the service for medical, legal, financial, or other professional advice.[79]

A provider’s policies may contain isolated statements that appear to promise accuracy. A court would still need to read those statements in context and determine whether they are material. The FTC often presumes that express claims are material, but a court may not treat a general assertion such as Grok’s description of itself as a “truth-seeking AI companion for unfiltered answers”[80] as material. Without a more specific assurance, the statement resembles nonactionable marketing puffery.[81]

Terms of service cannot always cure a deceptive representation or omission. Still, the broad claims about accuracy cited in the Statement appear unlikely to overcome providers’ much more specific and detailed disclaimers.

The Commission should explain when an AI provider’s representations could become deceptive despite express warnings that its products may produce inaccurate outputs. None of the policies claims perfect accuracy or complete ideological neutrality. The record therefore offers little basis to conclude that a reasonable consumer would expect either.

AI providers compete to offer chatbots that users regard as accurate, reliable, and useful. Advertising those products while clearly disclosing that their outputs may contain errors does not, without more, constitute deception. The Commission may investigate whether providers honor their concrete representations. Mere disagreement with the accuracy of a particular chatbot response would not ordinarily establish a Section 5 violation.

An enforcement theory based on the Commission’s own assessment of whether individual chatbot outputs are sufficiently accurate would require the agency to determine the truth of potentially vast numbers of expressive responses. That approach would recreate the First Amendment problems discussed in Part I.

III. The Statement’s Preemption Analysis Offers Little Practical Guidance

The AI Policy Statement’s treatment of federal preemption does not provide useful guidance on a difficult question of federalism and constitutional law. With little explanation and no directly applicable precedent, the Statement declares that “[a] state law that requires an AI firm to deceive its customers obviously conflicts with Section 5’s express purpose of protecting consumers from such conduct.”[82] That broad assertion may mislead regulated parties about the circumstances in which the FTC Act preempts state law.

Federal law, including the FTC Act, may preempt conflicting state laws and regulations.[83] Courts generally disfavor implied preemption, however, and usually analyze it under one of two doctrines. Field preemption applies when federal law occupies an entire area of regulation. Conflict preemption applies when compliance with both federal and state law is impossible or when state law obstructs Congress’ purposes. Neither doctrine fits the Statement’s analysis. The Statement also does not explain which state laws might be preempted, which would remain valid, or why.

We recognize the difficulty of the Commission’s assignment. The Statement responds to Executive Order 14365, Ensuring a National Policy Framework for Artificial Intelligence, which directs the FTC chair, in consultation with the special adviser for AI and crypto, to explain “how state laws requiring alterations to the accurate outputs of AI models can conflict with the requirements of the FTC Act.”[84]

Section 7 of the executive order specifically requires a policy statement explaining “the circumstances under which State laws that require alterations to the truthful outputs of AI models are preempted by the Federal Trade Commission Act’s prohibition on engaging in deceptive acts or practices affecting commerce.” Meeting that directive within the prescribed 90-day period presented a substantial challenge.

Part of the difficulty comes from the breadth of the subject. As the Commission recognizes, AI is “an umbrella term covering a universe of different tools and systems.”[85] AI products, services, models, and outputs vary widely and continue to develop. Section 5 enforcement also depends heavily on the facts of each case. Providing useful guidance in an unsettled area therefore requires careful attention to particular laws, representations, products, and applications.

The First Amendment presents another difficulty. As Part I explains, many AI models, applications, and outputs may constitute expressive speech entitled to the highest degree of constitutional protection. Both federal and state regulation must comply with those protections. Even the federal legislation contemplated by Section 8 of the executive order would face constitutional limits if it restricted protected expression.

The principal difficulty arises from preemption doctrine and the structure of the FTC Act. The Statement correctly acknowledges that “the FTC Act does not expressly preempt state law.”[86] Although implied preemption remains possible, courts generally apply a presumption against preemption in fields traditionally regulated by the states. Consumer protection is one such field.[87]

A. The FTC Act Does Not Occupy the Field of Consumer Protection

The Statement relies on Schneidewind v. ANR Pipeline Co.,[88] an implied-preemption decision that did not involve the FTC Act. In Schneidewind, the Supreme Court held that the federal Natural Gas Act of 1938 (“NGA”) preempted a Michigan law because the NGA established “a comprehensive scheme of federal regulation” and occupied the relevant field to the exclusion of state law.[89]

No comparable authority holds that Section 5 occupies the fields of advertising, marketing, or consumer protection so completely that it leaves no room for state regulation. The history and structure of consumer protection law point in the opposite direction.[90]

Courts generally recognize that state consumer protection statutes, including laws modeled on the FTC Act, operate alongside federal law when their substantive and jurisdictional requirements are satisfied. Congress has allowed the FTC Act and state consumer protection laws to coexist, subject to limited exceptions.[91]

The Supreme Court reached an analogous conclusion in California v. ARC America Corp.[92] There, the Court held that federal antitrust law, including the Sherman Act, did not preempt distinct state antitrust remedies.[93] The FTC Act likewise supplies no basis for treating federal consumer protection law as exclusive.

B. The Statement Identifies No Actual Conflict with State Law

The Statement also cites Schneidewind for the rule that state law is preempted when it actually conflicts with federal law. Such a conflict may arise when a private party cannot comply with both federal and state requirements or when state law obstructs Congress’ purposes and objectives.[94]

PLIVA, Inc. v. Mensing illustrates impossibility preemption.[95] The Supreme Court concluded that federal drug-labeling requirements prevented generic-drug manufacturers from independently making the label changes required by state tort law. Because the manufacturers could not comply with both sets of requirements, federal law preempted the state-law claims.[96]

The Statement identifies no comparable conflict between the FTC Act and any state AI law. We are unaware of a state law or regulation—regardless of its policy merits—that makes compliance with both state law and Section 5 impossible.

The Statement predicts that an AI provider might “suppress accuracy and interpose other objectives, such as so-called ‘equity,’ to avoid liability under this law, but fail to disclose these ulterior objectives in order to hide the loss of accuracy they necessitate.”[97] An AI provider could conceivably respond to a state law in that manner. But the Statement does not identify an existing law that requires such conduct or a concrete representation that would deceive consumers.

Any Section 5 analysis would depend on the provider’s actual statements, omissions, practices, and disclosures. It would also require an assessment of how reasonable consumers understood those representations and whether they were material. Speculation about how a provider might respond to an unidentified state law does not establish that simultaneous compliance is impossible.

The Statement also asserts that “[a] state law that requires an AI firm to deceive its consumers obviously conflicts with Section 5’s express purpose of protecting consumers from such conduct.”[98] An appeal to Section 5’s general purpose does not resolve the preemption question.

Section 5 broadly prohibits unfair methods of competition and unfair or deceptive acts or practices. It also defines the Commission’s enforcement authority and limits that authority. Its broad terms, case-specific application, and longstanding coexistence with state competition and consumer protection laws make obstacle preemption especially difficult to establish.

Those features also help explain courts’ growing skepticism toward preemption theories based on broad assertions about statutory purposes. As Justice Neil Gorsuch cautioned in Virginia Uranium, Inc. v. Warren, “‘pre-emptive purpose,’ whether express or implied, must therefore be ‘sought in the text and structure of the statute at issue.’”[99] The Statement does not identify language or structure in the FTC Act that supports its expansive suggestion of preemption.

AI providers can violate the FTC Act, and a particular state law could conceivably conflict with a federal law or regulation that the Commission enforces. Some existing or proposed state laws may also violate the First Amendment without regard to statutory preemption.

The Statement does not identify such a law, describe the conflict, or explain how established preemption doctrine would apply. Its general assurances therefore offer little guidance to AI providers, consumers, or federal and state lawmakers.

Conclusion

The Supreme Court has cautioned that “[o]ur constitutional tradition stands against the idea that we need Oceania’s Ministry of Truth.”[100] The Commission should heed that warning. It should not assume responsibility for deciding whether individual chatbot responses are true, false, balanced, or ideologically acceptable.

AI providers exercise editorial judgment when they develop models and generate responses, and users have a First Amendment interest in receiving those responses. Competition among providers also allows consumers to compare systems, test claims of accuracy and neutrality, and switch services when a product fails to meet their expectations.

The FTC retains authority to challenge concrete, material misrepresentations in advertising and marketing. Any enforcement action should follow the Deception Statement, distinguish commercial claims from protected expression, and account for providers’ express warnings that chatbot outputs may contain errors. Disagreement with an output, standing alone, does not establish deception.

The Commission should also revise the Statement’s preemption analysis. It should identify the state laws and factual circumstances that could create an actual conflict with Section 5 rather than rely on broad appeals to statutory purpose.

The FTC should revise the AI Policy Statement to provide clear, fact-specific guidance and to confine enforcement to conduct that falls within Section 5 and complies with the First Amendment. That approach would protect consumers without making the Commission the arbiter of truth in the marketplace of ideas.

^[1] Policy Statement Concerning the Suppression of Accuracy in Artificial Intelligence Systems, 91 Fed. Reg. 41,638 (July 7, 2026) [hereinafter AI Policy Statement].

[2] See, e.g., American Medical Ass’n v. FTC, 638 F.2d 443 (2d Cir. 1980); Iowa Chapter of Am. Physical Therapy Ass’n, 111 F.T.C. 199 (1988); Fed. Trade Comm’n Bureau of Econ., Staff Report: Effects of Restrictions on Advertising and Commercial Practice in the Professions: The Case of Optometry (1980), https://www.ftc.gov/sites/default/files/documents/reports/effects-restrictions-advertising-and-commercial-practice-professions-case-optometry/198009optometry.pdf; P. Ippolito & A. Mathios, Health Claims in Advertising and Labeling: A Study of the Cereal Market (FTC Bureau of Econ. Study, 1989); FTC Staff Comment Before the Massachusetts Department of Public Health Concerning Proposed Regulation of Limited Service Clinics (Oct. 1, 2007), https://www.ftc.gov/sites/default/files/documents/advocacy_documents/ftc-staff-comment-massachusetts-department-public-health-concerning-proposed-regulation-limited/v070015massclinic.pdf; FTC Staff Comments to the Sunset Advisory Commission of the State of Texas (Aug. 14, 1992) (addressing licensing restrictions on business practices imposed by the state boards of optometry, dentistry, and medicine).

[3] AI Policy Statement, supra note 1, at 41,639.

[4] See id. at 41,638 (“Excessive AI regulation would undermine American AI supremacy by deterring and suppressing the same ingenuity responsible for making American AI great.”).

[5] Id. at 41,641 n.44 (noting that “a law restricting truthful speech because it might lead another person to commit disparate impact discrimination would not survive First Amendment scrutiny,” but explaining that issue “is orthogonal to the FTC Act issues we address here”).

[6] Id. at 41,638.

[7] 15 U.S.C. § 45(a)(2); 15 U.S.C. § 44 (defining “corporation”).

[8] Cf. AI Policy Statement, supra note 1, at 41,639 n.11-12 and accompanying text.

[9] See AI Policy Statement, supra note 1, at 41,639 (citing Exec. Order No. 14365, Ensuring a National Policy Framework for Artificial Intelligence, 90 Fed. Reg. 58,499, 58,499 (Dec. 11, 2025)).

[10] Rice v. Santa Fe Elevator Corp., 331 U.S. 218, 230 (1947); see also Arizona v. United States, 567 U.S. 387 (2012); cf. Jones v. Google, LLC, 56 F.4th 735 (9th Cir. 2022) (holding that state-law claims were neither expressly nor impliedly preempted by the Children’s Online Privacy Protection Act).

[11] Geier v. American Honda Motor Co., 529 U.S. 861, 869-72 (2000).

[12] See infra Section III.B; see also Va. Uranium, Inc. v. Warren, 587 U.S. 761, 778 (2019) (plurality opinion) (quoting CSX Transp., Inc. v. Easterwood, 507 U.S. 658, 664 (1993)); Caleb Nelson, Preemption, 86 Va. L. Rev. 225, 231-32 (2000) (arguing against obstacle preemption).

[13] Much of this section draws from earlier ICLE work, including Int’l Ctr. for L. & Econ., ICLE Comments to FTC Regarding Technology Platform Censorship (May 21, 2025), https://laweconcenter.org/resources/icle-comments-to-ftc-regarding-technology-platform-censorship; Memorandum of Amicus Curiae International Center for Law & Economics in Support of Defendant’s Motion for Summary Judgment, State ex rel. Yost v. Google LLC, No. 21-CV-H-06-0274 (Ohio Ct. Com. Pl. Delaware Cnty. Jan. 30, 2024), https://laweconcenter.org/wp-content/uploads/2024/01/2024-1-30-ICLE-Amicus-Curiae-Brief.pdf; and Ben Sperry, Knowledge and Decisions in the Information Age: The Law & Economics of Regulating Misinformation on Social-Media Platforms, 59 Gonz. L. Rev. 319 (2024). For further discussion, see Brief of Amicus Curiae International Center for Law & Economics in Support of Appellee, State ex rel. Yost v. Google LLC, No. 25 CAE 08-0070 (Ohio Ct. App. 5th Dist. Dec. 19, 2025), https://laweconcenter.org/wp-content/uploads/2025/12/ICLE-Amicus-Yost-v.-Google-Court-of-Appeals.pdf; Brief of Amicus Curiae International Center for Law & Economics, Moody v. NetChoice, LLC, Nos. 22-277 & 22-555 (U.S. Dec. 4, 2023), https://laweconcenter.org/wp-content/uploads/2023/12/Intl-Ctr-for-Law-and-Econ-Amicus-12.4.231148722.12.pdf; Brief of Amicus Curiae International Center for Law & Economics, Murthy v. Missouri, No. 23-411 (U.S. Feb. 9, 2024), https://laweconcenter.org/wp-content/uploads/2024/02/Murthy-v.-Missouri-Intl-Center-for-Law-Econ.-Am.-Br.-2-9-24-pm-FINAL.pdf.

[14] See, e.g., John Stuart Mill, On Liberty ch. 2 (1859); John Milton, Areopagitica (1644).

[15] See Exec. Order No. 14365, supra note 9; see also Kevin Schaul, Are ChatGPT and Other AI Chatbots Politically Biased? We Tested Them., Wash. Post (June 29, 2026), https://www.washingtonpost.com/technology/interactive/2026/06/24/are-ai-chatbots-like-chatgpt-politically-biased-we-tested-them.

[16] See Mill, supra note 14, ch. 2.

[17] See Thomas Jefferson, First Inaugural Address (Mar. 4, 1801), https://avalon.law.yale.edu/19th_century/jefinau1.asp.

[18] See Abrams v. United States, 250 U.S. 616, 630 (1919) (Holmes, J., dissenting) (“[T]ime has upset many fighting faiths,” and “the ultimate good desired is better reached by free trade in ideas—that the best test of truth is the power of the thought to get itself accepted in the competition of the market…. That at any rate is the theory of our Constitution.”)

[19] See David Schultz, Marketplace of Ideas, Free Speech Ctr., https://firstamendment.mtsu.edu/article/marketplace-of-ideas (last updated July 9, 2024).

[20] Manhattan Cmty. Access Corp. v. Halleck, 587 U.S. 802, 804 (2019).

[21] Moody v. NetChoice, LLC, 603 U.S. 707 (2024).

^[22] Id. at 717.

[23] Id. at 709-10.

[24] Id. at 737-38.

[25] Id. at 733.

[26] AI Policy Statement, supra note 1, at 41,641.

[27] Moody v. NetChoice, LLC, 603 U.S. 707, 745 (2024) (Barrett, J., concurring).

[28] Cf. id. at 716 (“To the extent that social-media platforms create expressive products, they receive the First Amendment’s protection.”).

[29] Zhang v. Baidu.com Inc., 10 F. Supp. 3d 433 (S.D.N.Y. 2014).

[30] Id. at 438.

[31] Id.

[32] See e-ventures Worldwide, LLC v. Google, Inc., No. 2:14-cv-646-FtM-PAM-CM, 2017 WL 2210029, at *4 (M.D. Fla. Feb. 8, 2017); Langdon v. Google, Inc., 474 F. Supp. 2d 622, 629-30 (D. Del. 2007).

[33] Search King, Inc. v. Google Tech., Inc., No. CIV-02-1457-M, 2003 WL 21464568 (W.D. Okla. May 27, 2003)..

[34] Id. at *4.

[35] Cf. Moody, 603 U.S. at 732 (“[T]he government cannot get its way just by asserting an interest in improving, or better balancing, the marketplace of ideas.”).

[36] Martin v. City of Struthers, 319 U.S. 141, 143 (1943).

[37] Thomas v. Collins, 323 U.S. 516, 534 (1945) (“That there was restriction upon Thomas’ right to speak and the rights of the workers to hear what he had to say, there can be no doubt.”)

[38] Martin, 319 U.S. at 145–47 (“Freedom to distribute information to every citizen wherever he desires to receive it is so clearly vital to the preservation of a free society that, putting aside reasonable police and health regulations of time and manner of distribution, it must be fully preserved.”) (praising pamphlets as a means of “espousing various causes”).

[39] Stanley v. Georgia, 394 U.S. 557, 564 (1969) (“It is now well established that the Constitution protects the right to receive information and ideas.”); Smith v. California, 361 U.S. 147, 153 (1959) (finding that the ordinance “tends to impose a severe limitation on the public’s access to constitutionally protected matter”).

[40] Va. State Bd. of Pharmacy v. Va. Citizens Consumer Council, Inc., 425 U.S. 748, 763–64 (1976) (explaining that a consumer’s interest in the free flow of commercial information “may be as keen, if not keener by far, than his interest in the day’s most urgent political debate,” and that society may also have a strong interest in such information).

[41] Brown v. Ent. Merchs. Ass’n, 564 U.S. 786, 794 (2011) (holding that minors receive substantial First Amendment protection and that the government may bar their access to protected material only in “relatively narrow and well-defined circumstances,” including when regulating violent video games).

[42] Packingham v. North Carolina, 582 U.S. 98, 108 (2017) (“[T]o foreclose access to social media altogether is to prevent the user from engaging in the legitimate exercise of First Amendment rights,” including access to “the world of ideas.”).

[43] See AI Policy Statement, supra note 1, at 41,641 n.45 (“The Commission at this time takes no position on whether the practices discussed in this statement may also be unfair under the FTC Act.”).

[44] FTC Policy Statement on Unfairness, appended to International Harvester Co., 104 F.T.C. 949, 1070 (1984) (Dec. 17, 1980), https://www.ftc.gov/legal-library/browse/ftc-policy-statement-unfairness.

[45] Cf. AI Policy Statement, supra note 1, at 41,641 (“A company could be tempted, for example, to abuse consumer trust by training a model surreptitiously to produce ideologically motivated distortions in a response to a factual question, such as to correct what the developer believes are ‘historical injustices’ in the facts.”).

^[46] United States v. Alvarez, 567 U.S. 709 (2012).

^[47] Id. at 717 (quoting United States v. Stevens, 559 U.S. 460, 468 (2010)).

^[48] Id. at 718.

[49] See Chaplinsky v. New Hampshire, 315 U.S. 568, 571–72 (1942) (“There are certain well-defined and narrowly limited classes of speech, the prevention and punishment of which has never been thought to raise any Constitutional problem.”).

^[50] See Alvarez, 567 U.S. at 718–22; see also Pittsburgh Press Co. v. Pittsburgh Commission on Human Relations, 413 U.S. 376, 389 (1973).

[51] Alvarez, 567 U.S. at 713.

[52] Id. at 715.

[53] See, e.g., Catherine Thorbecke & Clare Duffy, Google Halts AI Tool’s Ability to Produce Images of People After Backlash, CNN (Feb. 22, 2024), https://www.cnn.com/2024/02/22/tech/google-gemini-ai-image-generator.

[54] Alvarez, 567 U.S. at 727–28.

[55] See AI Policy Statement, supra note 1, at 41,640.

[56] For more on the interaction among Section 5 of the Federal Trade Commission Act, antitrust law, and content moderation, see Comments of the Program on Economics & Privacy, George Mason University Antonin Scalia Law School, Re: Request for Public Comment Regarding Technology Platform Censorship (May 21, 2025), https://masonlec.org/wp-content/uploads/2025/05/PEP-Comment-FTC-Censorship_Final-Filed-5.21.2025.pdf; Int’l Ctr. for L. & Econ., ICLE Comments to FTC Regarding Technology Platform Censorship, supra note 13; Daniel J. Gilman & Ben Sperry, Is There an Empty Set at the Intersection of Antitrust and Content Moderation?, Concurrences (Nov. 3, 2025), https://laweconcenter.org/resources/is-there-an-empty-set-at-the-intersection-of-antitrust-and-content-moderation.

[57] See AI Policy Statement, supra note 1, at 41,639 n.11 (citing White House, Fact Sheet: President Donald J. Trump Ensures a National Policy Framework for Artificial Intelligence (Dec. 11, 2025), https://www.whitehouse.gov/fact-sheets/2025/12/fact-sheet-president-donald-j-trump-ensures-a-national-policy-framework-for-artificial-intelligence).

[58] Id. at 41,639 n.12.

[59] See, e.g., Bolger v. Youngs Drug Prods. Corp., 463 U.S. 60 (1983); Cent. Hudson Gas & Elec. Corp. v. Pub. Serv. Comm’n, 447 U.S. 557 (1980).

[60] FTC Policy Statement on Deception, appended to Cliffdale Assocs., Inc., 103 F.T.C. 110, 174 (1984) (Oct. 14, 1983), https://www.ftc.gov/system/files/documents/public_statements/410531/831014deceptionstmt.pdf.

[61] See, e.g., Donaldson v. Read Magazine, Inc., 333 U.S. 178, 190 (1948) (recognizing that the government’s power “to protect people against fraud” has “always been recognized in this country and is firmly established”).

[62] At common law, fraud generally requires (1) a material misrepresentation or omission where there is a duty to disclose, (2) intent to induce reliance, (3) knowledge of the statement’s falsity or misleading nature, (4) justifiable reliance, and (5) resulting injury. See, e.g., Mandarin Trading Ltd. v. Wildenstein, 919 N.Y.S.2d 465, 469 (N.Y. 2011); Kostryckyj v. Pentron Lab. Techs., LLC, 52 A.3d 333, 338-39 (Pa. Super. Ct. 2012); Masingill v. EMC Corp., 870 N.E.2d 81, 88 (Mass. 2007). Likewise, restrictions on deceptive or misleading commercial speech have long been held consistent with the First Amendment. See Va. State Bd. of Pharmacy v. Va. Citizens Consumer Council, Inc., 425 U.S. 748, 771-72 (1976) (“Obviously, much commercial speech is not provably false, or even wholly false, but only deceptive or misleading. We foresee no obstacle to a State’s dealing effectively with this problem. The First Amendment, as we construe it today does not prohibit the State from insuring that the stream of commercial information flow cleanly as well as freely.”).

^[63] See Alvarez, 567 U.S. at 719 (“Even when considering some instances of defamation and fraud . . . the Court has been careful to instruct that falsity alone may not suffice to bring the speech outside the First Amendment. The statement must be a knowing or reckless falsehood.”). Thus, even in areas such as defamation and fraud, the First Amendment limits liability for false speech absent the requisite level of fault.

[64] Illinois ex rel. Madigan v. Telemarketing Assocs., Inc., 538 U.S. 600, 617 (2003).

[65] See, e.g., Schaumburg v. Citizens for a Better Env’t, 444 U.S. 620 (1980); Sec’y of State of Md. v. Joseph H. Munson Co., 467 U.S. 947 (1984); Riley v. Nat’l Fed’n of the Blind of N.C., Inc., 487 U.S. 781 (1988).

[66] Madigan, 538 U.S. at 620.

[67] FTC Policy Statement on Deception, supra note 60.

[68] Id.

[69] Id.

[70] Id.

[71] See supra note 2 and accompanying text.

[72] See AI Policy Statement, supra note 1, at 41,640 nn.35–36 and accompanying text.

[73] OpenAI, Why Language Models Hallucinate (Sept. 5, 2025), https://openai.com/index/why-language-models-hallucinate.

[74] OpenAI, Terms of Use (Jan. 1, 2026), https://openai.com/policies/row-terms-of-use.

[75] Anthropic, Consumer Terms of Service (Oct. 8, 2025), https://www.anthropic.com/legal/consumer-terms.

[76] DeepSeek, Terms of Use (Mar. 27, 2026), https://cdn.deepseek.com/policies/en-US/deepseek-terms-of-use.html.

[77] Gab AI, GAB AI Service Terms of Service (Apr. 8, 2025), https://gab.ai/terms-of-service.

[78] xAI, Terms of Service—Consumer (June 26, 2026), https://x.ai/legal/terms-of-service.

[79] Google, Generative AI Additional Terms of Service (Aug. 9, 2023), https://policies.google.com/terms/generative-ai.

[80] AI Policy Statement, supra note 1, at 41,640 n.35.

[81] Cf. FTC Policy Statement on Deception, supra note 60 (“The Commission generally will not pursue cases involving obviously exaggerated or puffing representations, i.e., those that ordinary consumers do not take seriously.”).

[82] AI Policy Statement, supra note 1, at 41,641.

[83] See, e.g., California v. ARC America Corp., 490 U.S. 93, 101 (1989); Rice, 331 U.S. at 230; Arizona, 567 U.S. 387. The Supreme Court has refined its preemption doctrine—grounded principally in the Supremacy Clause, U.S. Const. art. VI, cl. 2, and, to some extent, the Necessary and Proper Clause, id. art. I, § 8, cl. 18—since McCulloch v. Maryland, 17 U.S. (4 Wheat.) 316 (1819).

[84] See AI Policy Statement, supra note 1, at 41,639 (citing Exec. Order No. 14365, Ensuring a National Policy Framework for Artificial Intelligence, 90 Fed. Reg. 58,499, 58,499 (Dec. 11, 2025)).

[85] Id. at 41,638.

[86] Id. at 41,641.

[87] ARC America, 490 U.S. at 101 (citing Hillsborough County v. Automated Medical Laboratories, Inc., 471 U.S. 707, 716 (1985)); see also Wyeth v. Levine, 555 U.S. 555 (2009); Rice, 331 U.S. at 230.

[88] Schneidewind v. ANR Pipeline Co., 485 U.S. 293 (1988).

[89] Id. at 300; see also, e.g., Rice, 331 U.S. at 230; Arizona v. United States, 567 U.S. 387 (2012).

[90] Rice, 331 U.S. at 230; cf. Jones v. Google, LLC, 56 F.4th 735 (9th Cir. 2022) (holding that the Children’s Online Privacy Protection Act neither expressly nor impliedly preempted the state-law claims).

[91] Walsh v. Ford Motor Co., 807 F.2d 1000, 1013–14 (D.C. Cir. 1986) (“Congress intended the application of state law, except as expressly modified by Magnuson-Moss, in section 110(d) breach of warranty actions.”); Deadwyler v. Volkswagen of America, Inc., 748 F. Supp. 1146, 1150 (W.D.N.C. 1990) (explaining that, in enacting the Magnuson-Moss amendments to the Federal Trade Commission Act, “Congress did not intend to replace state law except in those instances[] where it explicitly said so”).

[92] California v. ARC America Corp., 490 U.S. 93 (1989).

[93] See id. at 101.

[94] Schneidewind, 485 U.S. at 300; see also Florida Lime & Avocado Growers, Inc. v. Paul, 373 U.S. 132, 142–43 (1963); Freightliner Corp. v. Myrick, 514 U.S. 280, 287 (1995).

[95] PLIVA, Inc. v. Mensing, 564 U.S. 604 (2011).

[96] Id. at 618 (quoting Freightliner, 514 U.S. at 287).

[97] AI Policy Statement, supra note 1, at 41,641.

[98] Id.

[99] Va. Uranium, 587 U.S. at 778 (plurality opinion) (quoting CSX Transportation, 507 U.S. at 664); see also Nelson, supra note 12, at 231–32 (arguing against obstacle preemption).

[100] Alvarez, 567 U.S. at 728.

Executive Summary

Radio spectrum is essential infrastructure for artificial intelligence, augmented reality, virtual reality, and other next-generation wireless applications. Yet spectrum is not a conventional natural resource. Its scarcity arises from limits on bandwidth, interference, network design, and the legal rights that determine who may transmit, where, and under what conditions.

No single frequency band or access model can meet every need. Licensed spectrum supports wide-area networks and reliable service. Unlicensed spectrum powers Wi-Fi, short-range devices, and much indoor immersive use. Dynamic sharing can open bands that cannot be fully cleared, while satellite networks extend coverage beyond terrestrial systems. Policymakers should preserve this balance rather than judge allocations by auction revenue alone.

The current U.S. system creates avoidable delay and uncertainty. Divided authority between the Federal Communications Commission and the National Telecommunications and Information Administration complicates federal reallocation. The Spectrum Relocation Fund remains cumbersome, and worst-case interference analysis can block productive uses based on highly improbable scenarios.

The United States should strengthen interagency coordination, streamline relocation funding, adopt risk-informed interference analysis, expand advanced coordination tools, and present coherent positions internationally. These reforms would protect incumbents where necessary while making more spectrum available for investment, competition, and innovation.

I. Introduction

In the opening issue brief in this series, International Center for Law & Economics Director of Innovation Policy Kristian Stout examined how infrastructure shapes where innovation takes root.[1] Technology clusters flourish where roads, power, communications networks, and other shared resources support investment and growth. The returns may take time, but the United States’ past investments in these enabling conditions helped make it the world’s technological leader. This approach amounts to horizontal industrial policy, which creates broad conditions for innovation without choosing favored technologies or firms.[2]

Stout identifies several forms of infrastructure that the United States will need to sustain that lead. Radio spectrum—the frequencies used to transmit wireless signals—rarely receives equal attention. It tends to make news when Congress treats auction revenue as a way to pay for must-pass legislation or when harmful interference threatens to disrupt air travel.[3] Yet every wireless service depends on predictable access to frequencies where competing signals will not overwhelm or interrupt one another. Spectrum therefore forms a basic layer of modern communications infrastructure.

Spectrum is harder to picture than a highway, power line, or pipe. It consists of a finite range of electromagnetic frequencies that can carry signals through the air. Those signals support mobile calls, Wi-Fi, GPS navigation, emergency communications, satellite services, and countless other applications. Different frequencies have different physical properties, and several operators may want to transmit on the same frequency, in the same place, at the same time. Without rules governing those uses, interference can prevent every affected service from working as intended.

The United States has therefore created operating rights that divide access to spectrum among users. Federal agencies and international bodies allocate those rights to limit harmful interference and allow different services to coexist.[4] These decisions determine who may use particular frequencies, under what conditions, and for which purposes. Spectrum management is consequently an economic and political exercise as well as a technical one.

Poor spectrum management can constrain the digital economy much as highway congestion impedes commerce. Artificial intelligence (AI), autonomous vehicles, smart manufacturing, precision agriculture, and augmented- and virtual-reality applications require substantial wireless capacity, very low latency, and reliable connections. Meeting those demands requires a balanced strategy that combines high-powered exclusive licenses, shared-access regimes, and unlicensed spectrum.

Exclusive licenses give mobile carriers predictable protection from interference, allowing them to invest in dense 5G and future 6G networks. These networks can support AI-enabled and other data-intensive applications when users lack access to a dedicated local network. Unlicensed spectrum offers easier access for Wi-Fi, internet-of-things sensors, experimental technologies, and services operating over shorter distances. Shared-access regimes allow several users to operate within the same band under rules that coordinate their transmissions.

Each model serves different technical and economic purposes. Overreliance on any one of them can restrict entry, weaken investment certainty, or leave valuable frequencies underused. Exclusive licensing, shared access, and unlicensed use should therefore operate as complementary tools within a broader spectrum strategy.

The federal government must also move more quickly to reallocate underused spectrum held by incumbent users, including federal agencies. Commercial networks, scientific research, and emerging technologies will require access to additional frequencies. Delayed reallocation carries real costs by slowing deployment, investment, and experimentation.

Spectrum policy is infrastructure policy. The choices federal officials make today will shape whether the United States can build and deploy the next generation of wireless technologies.

II. What Spectrum Is, Why It Is Scarce, and How It Creates Value

Sound spectrum policy begins with a clear account of the resource being managed. Spectrum refers to the frequencies used to carry wireless signals, but the policy term often blends physical capacity, interference constraints, and government-created operating rights.

Those distinctions shape the analysis that follows. Technical scarcity reflects limits on bandwidth, receiver performance, and the number of systems that can operate without harmful interference. Legal scarcity arises from the licenses, band plans, and operating rules that determine who may transmit. Because those rights are indispensable and their value varies across commercial, public, and unlicensed uses, regulators must decide which allocation method will produce the greatest economic and public benefit.

A. What Is Spectrum?

Any discussion of spectrum policy should begin with the technology itself. Radios transmit information by modifying an electromagnetic wave, usually by changing its amplitude, frequency, or another feature of the signal. Most modern systems use digital transmission, which encodes information as binary data. Digital signals can carry large amounts of information and can be copied or regenerated with less degradation than analog signals.

An electromagnetic wave moves through repeating cycles. Its frequency measures how many cycles occur during a given period, usually one second. Frequency is measured in hertz. A wave that completes 1 million cycles per second operates at 1 megahertz, while one that completes 1 billion cycles per second operates at 1 gigahertz.

A radio receiver detects the electric and magnetic fields carried by an incoming wave. Those fields generate a small alternating current that corresponds to the wave’s frequency. Because the surrounding environment contains many electromagnetic signals, the receiver must isolate the one it is designed to use.

A tuner performs that task by adjusting the receiver’s resonant frequency. The receiver responds most strongly to the selected carrier frequency while filtering out much of the surrounding noise. It then demodulates the signal, extracting the information carried by the wave. Traditional amplitude modulation (AM) stores information in changes to the wave’s strength. Frequency modulation (FM) stores it in changes to the wave’s frequency. The receiver then amplifies and converts the extracted information into the desired output, such as sound from an AM radio broadcast.

This distinction matters for policymakers because “spectrum” can refer to different things. In physical terms, the radio spectrum is the range of frequencies used for wireless communications. In policy debates, officials often speak of “freeing up” spectrum or making more of it available. That language treats spectrum as a scarce, property-like resource. Frequency, though, differs from ordinary property in important ways. The underlying frequencies do not run out. Scarcity arises because multiple users may seek to transmit at the same frequency, time, and place, creating interference that prevents their systems from working as intended.

B. What Spectrum Scarcity Really Means

Spectrum is often described as a scarce natural resource. The analogy is useful, but incomplete. Radio frequencies are not consumed through use. Scarcity arises because communications systems have limited capacity and may interfere with one another when they operate at the same frequency, time, and place.[5]

Spectrum scarcity therefore has two related forms. Technical scarcity reflects limits on bandwidth, receiver performance, network design, and the number of systems that can operate without harmful interference. Legal scarcity results from the licenses, band plans, and operating rules that determine who may transmit and under what conditions. Together, these technical constraints and legal rights determine how much spectrum capacity users can put to productive use.

1.        Technical Scarcity

From a technical perspective, spectrum scarcity can describe two different constraints. The first concerns bandwidth, which is the range of frequencies available to carry a signal.[6] Wider channels can transmit more data per second than narrower ones.

Channels sit within larger frequency bands, each defined by an upper and lower boundary and often allocated for a particular use. A channel only a few megahertz wide can carry limited information at any one time. A much wider channel can carry far more data. Higher-frequency bands often accommodate wider channels,[7] but they also tend to have shorter range and weaker ability to penetrate walls and other obstacles. Spectrum policy therefore requires tradeoffs between channel width and the physical characteristics of the frequencies involved.[8]

Radio frequencies themselves are not depleted through use. A transmission does not consume a channel or leave less of it for the next user. The relevant constraint is the amount of information a channel can carry.

Claude Shannon formalized that limit. Shannon’s law establishes the maximum amount of information that a communications channel can transmit given its bandwidth and signal-to-noise ratio.[9] Better antennas, coding, and signal processing can move a system closer to that ceiling, but they cannot exceed it.

Policymakers usually have a second form of scarcity in mind. Several radios may seek to operate at the same frequency, time, and place. Electromagnetic waves can pass through one another, but a receiver may fail to distinguish its intended signal from surrounding transmissions or noise. The desired information may then become distorted, drowned out, or impossible to recover. That condition is interference.

Every radio system encounters some interference. Harmful interference occurs when interference prevents a service from functioning as intended.[10] The Federal Communications Commission (FCC) licenses many radio operations and adopts technical rules designed to limit that risk. This process is known as spectrum management.

Interference does not always result from too many users competing for too few frequencies. It may instead reflect poor receiver design, inadequate filtering, insufficient network density, or other engineering choices within a particular system. Coordination can also allow several users to share a channel while each operates near the channel’s technical capacity.

Technical spectrum scarcity therefore does not mean that one user consumes a finite resource at another’s expense. It describes limits created by channel capacity, receiver performance, network design, and the interaction among multiple radio systems.

2.        Legal Scarcity

Technical scarcity depends on both the bandwidth available to a system and its ability to operate without harmful interference. Spectrum management seeks to preserve that capacity by defining who may transmit, where, when, and under what technical conditions. Those legal rules create a second form of scarcity.

Before a radio service may operate in the United States, it generally needs authorization from either the FCC or the National Telecommunications and Information Administration (NTIA). The FCC oversees nonfederal uses, while NTIA manages spectrum used by federal agencies. Depending on the service, the FCC may issue an individual license or permit unlicensed operation under technical rules that apply within designated bands.[11]

A license gives its holder the legal right to transmit on specified frequencies, in specified locations, and subject to stated conditions. Band plans divide larger frequency ranges into channels and establish how many licenses may be issued. Because those rights are limited, several applicants may seek the same authorization. The FCC initially relied on first-come, first-served assignments and later used comparative hearings when applicants competed for the same license.[12]

The FCC and NTIA manage spectrum by separating users across frequency, geography, time, or some combination of the three.[13] Clear operating rights and technical limits can reduce harmful interference and allow more systems to function at once. Those limits can also restrict performance. Some applications therefore require several channels or exclusive rights within a band to obtain sufficient capacity and reliability.

Regulators must also distinguish harmful interference from ordinary increases in background noise. A new entrant should not lose access merely because its operations add some interference. The relevant question is whether that interference prevents another authorized service from functioning as intended.

Legal and technical scarcity together define each operator’s bundle of rights and restrictions. Those rules aim to provide enough bandwidth and protection from interference for the service to work. Because many users seek the same rights, allocating licenses and bands has become increasingly difficult and politically contentious.

For most operators, the legal authorization to transmit is the scarce input that matters most. Any proposal to reassign frequencies or revise the rights attached to a license can change the value and viability of existing systems. That is why even narrow technical proceedings before the FCC and NTIA often produce fierce disputes.

C. Spectrum Rights as an Economic Input

Access to spectrum is an indispensable input for any radio operation. Unlike many production inputs, spectrum rights cannot be replaced, manufactured, or avoided. A manufacturer that cannot obtain a particular mineral may redesign a product around another material. A logistics company facing a fuel shortage may reroute shipments or use another mode of transportation. A radio operator without legal authority to transmit cannot operate.

This lack of substitutes gives spectrum rights a distinctive economic character. In most markets, scarcity raises prices, which encourages conservation, increased production, or the development of alternatives. Higher license prices cannot create additional frequencies. Engineering can increase the amount of information carried over a band or allow more users to share it, but those gains remain subject to physical limits.

Dynamic spectrum sharing, improved interference mitigation, denser networks, and more advanced modulation can increase the productive use of existing frequencies. These techniques allow communications systems to approach the capacity limit described by Shannon’s law. They do not eliminate that limit.

The value of spectrum rights also varies sharply across users and applications. A license or unlicensed-access framework defines a bundle of rights and restrictions, but its economic value depends on the service involved, the capacity the operator can use, and the revenue or operational benefit that capacity can produce.

For a commercial mobile carrier, spectrum is a core input into a network serving millions of subscribers. A license’s value may depend on the population within the licensed area, expected data demand, the band’s technical characteristics, and the competitive consequences of allowing a rival to acquire it. The high prices paid in FCC auctions reflect carriers’ assessments of the long-term commercial value of those rights.[14] Nationwide mobile carriers have historically dominated such auctions, but cable companies, fixed-wireless providers, and other network operators also treat spectrum as one component of broader capital investment in communications infrastructure.

Other users value spectrum differently. A television broadcaster may care most about a signal’s geographic reach and the advertising revenue associated with that audience.[15] A federal agency operating radar, military communications, or scientific instruments may measure value through mission performance rather than revenue. Traditional market measures offer only a partial account of those uses.

Unlicensed spectrum presents another valuation problem. Wi-Fi routers, Bluetooth devices, industrial sensors, and other connected technologies rely on access that no single firm owns. Businesses, cable operators, universities, hospitals, manufacturers, municipalities, and households invest heavily in equipment and networks that use those frequencies. No auction bid captures the full value of that activity because the benefits are dispersed across billions of devices and transactions.[16]

These different uses complicate efforts to allocate spectrum efficiently. Competitive auctions generally provide the best mechanism for discovering value when a limited number of operators seek exclusive licenses. Bids reveal how much firms expect to earn from the rights offered and direct licenses toward users willing to pay the most.[17]

Auction prices do not, however, capture every form of economic or public value. Users that generate direct revenue can usually express their demand through a bid more readily than users whose benefits are indirect, diffuse, or tied to public missions. A mobile carrier can estimate subscriber revenue. A weather satellite, military radar, hospital network, or open Wi-Fi system may create substantial value that never appears on a licensee’s balance sheet.

Spectrum policy should seek to maximize the value created by an allocation and assignment, rather than auction receipts alone. In some cases, that goal will support an auction for exclusive licenses. In others, it may favor unlicensed access, shared use, or a private transaction that places frequencies into service more quickly.[18]

Regulators therefore face a recurring institutional choice. Markets can reveal which parties place the highest monetary value on defined rights. But government agencies must still define those rights, address uses whose benefits are difficult to monetize, and decide when exclusive licensing, sharing, or unlicensed access will produce the greatest value.

III. How Spectrum Rights Are Allocated and Reallocated

Radio operating rights are indispensable inputs for wireless services, making their allocation and assignment among the most consequential decisions in spectrum management. The quality, certainty, and flexibility of those rights shape whether firms will invest in networks and devices capable of supporting artificial intelligence, augmented reality, virtual reality, and other data-intensive applications.

The current system divides authority between the FCC and the NTIA, whose differing mandates can slow reallocation and create uncertainty. Within bands under FCC control, regulators must choose among exclusive licenses, unlicensed access, and dynamic-sharing arrangements, each of which offers different levels of interference protection, flexibility, and investment certainty. Secondary markets can later move licensed rights toward more productive uses, though transaction costs and limits on federal participation constrain their reach.

A. Divided Authority and Regulatory Delay

The United States is unusual among major economies in dividing spectrum-management authority between two federal agencies with different mandates and constituencies.[19] The Federal Communications Commission (FCC) governs nonfederal radio operations. Its authority traces to the Federal Radio Commission, which Congress created in response to disorder in early radio broadcasting.[20] Congress directed the FCC to regulate radio communications in the public interest, a mandate that has often pushed the agency toward productive use, flexible allocation, and market-based assignment.[21]

The National Telecommunications and Information Administration (NTIA) manages spectrum used by federal agencies. Its role grew out of the Interdepartment Radio Advisory Committee (IRAC), a cooperative body that agencies formed before Congress created a formal regulatory system for civilian users.[22] NTIA also seeks efficient spectrum use, but its institutional structure gives incumbent agencies substantial influence over decisions affecting their operations. Its approach therefore tends to place greater weight on protecting federal systems from harmful interference, though priorities vary by administration.[23]

The FCC and NTIA must make joint decisions about a shared resource, even though their institutional incentives often diverge. Neither agency has final authority over every decision modern spectrum management requires. The FCC can regulate nonfederal operators and reallocate bands occupied solely by commercial users. It cannot compel a federal agency to relocate, nor can it determine whether a federal assignment remains necessary without NTIA’s cooperation.

NTIA can use presidential authority over the executive branch to direct agencies to relinquish or share spectrum. In practice, the IRAC gives incumbent agencies considerable influence, and aggressive action depends heavily on White House direction. The resulting structure lacks a clear final arbiter.

That uncertainty carries economic costs. Delays reduce the value of spectrum rights by postponing deployment and revenue.[24] Auction bidders and equipment manufacturers also need confidence that allocation decisions will remain settled. When later challenges may reopen those decisions, firms have less reason to invest in networks, devices, and services tied to the affected bands.

The consequences become clearest during spectrum reallocation. The 3.1-3.55 GHz band supported federal defense radar systems, and the FCC could not repurpose those frequencies without agreement from the Department of Defense (DOD) and NTIA.[25] Congress ultimately directed further study and sharing through the MOBILE NOW Act.[26] Without that mandate, even the limited arrangement covering the upper 100 MHz of the band might not have emerged because the DOD had little institutional reason to surrender access voluntarily.

The 24 GHz proceeding showed how neighboring federal operations can create uncertainty even when no federal incumbent occupies the band itself.[27] The National Oceanic and Atmospheric Administration (NOAA) raised interference objections late in the process, after auction planning was underway. The agency appeared to bypass the established IRAC process and took its concerns directly to Congress and the public. That dispute weakened confidence that the FCC’s allocation decision was final and threatened to reduce the licenses’ value.

The 5.9 GHz proceeding revealed a related problem. An agency with no operating rights in or adjacent to the band, but with a strong programmatic interest in its use, exerted substantial pressure to reopen FCC decisions.[28] These episodes show how divided authority allows federal and other incumbent interests to resist reallocation, prolong proceedings, and reduce the value of spectrum rights even after the technical record supports a change.

Successful spectrum management therefore depends on close coordination between the FCC and NTIA, backed by clear White House direction. The next generation of wireless applications will require faster decisions involving difficult technical and economic tradeoffs. An administration that seeks continued U.S. leadership in wireless technology must align the agencies around common priorities, shared technical standards, and a consistent account of the national interest. Without that coordination, the federal government will spend too much time reopening old disputes while new technologies wait for access.

B. Licensed, Unlicensed, and Secondary-Market Access

Even when a band falls entirely under FCC control, allocating access presents difficult choices. The FCC uses two primary models. It may grant licenses to individual operators, or it may permit unlicensed operation under generally applicable technical rules.

A license defines an operator’s rights and restrictions. These may include the frequencies, power levels, geographic areas, and times in which the licensee may transmit. The precise terms depend on the band plan, the service involved, and neighboring operations.[29]

Many licenses provide exclusive access to a frequency range and protection against harmful interference. Those rights give operators greater confidence that later entrants will not jeopardize investments in networks and equipment. Other licenses grant several operators similar or overlapping rights, particularly in bands used for satellite services.[30]

Licensing can also support sharing across time or geography. The 3.5 GHz Citizens Broadband Radio Service (CBRS) band, for example, gives naval radar systems priority access while allowing Priority Access Licensees to operate when they will not interfere with those federal systems.[31]

The FCC’s second model permits unlicensed operation. Unlicensed bands initially offered inventors a low-cost way to develop and test devices without obtaining an individual license.[32] Consumers encounter this model through Wi-Fi routers, Bluetooth headphones, and other devices that create short-range personal or local networks.

Requiring every consumer to obtain FCC approval before setting up a home network would make those technologies commercially impractical. Device manufacturers instead design products to comply with the FCC’s Part 15 rules and obtain equipment authorization. Consumers may then operate approved devices in designated unlicensed bands without further agency permission.[33]

Unlicensed access supports a large share of internet traffic, much of which travels over Wi-Fi.[34] Part 15 imposes power and technical limits intended to accommodate many nearby users. Unlicensed devices must also accept interference from other lawful operations. A user generally has no regulatory remedy when a neighboring device causes interference while complying with FCC rules.

Licensed and unlicensed access serve different purposes. A license offers greater certainty, stronger interference protection, and often higher permitted power. Those features support wide-area networks (WANs) and reduce some of the risks associated with capital-intensive broadband deployment.

Unlicensed access works best for localized uses that can tolerate variable interference conditions. In a large apartment building, residents can establish separate wireless networks without seeking individual licenses and usually without causing harmful interference to their neighbors. Part 15 also allows firms to develop and deploy short-range communications for internet-of-things devices without the cost and delay of securing operating licenses.

These two models allocate initial operating rights, but only licenses generally create transferable rights. An unlicensed user has no individual property-like interest to sell or assign. A licensee, by contrast, may transfer all or part of its authorization to another party, subject to FCC rules and approval.

Secondary-market transactions can move spectrum rights toward users that value them more highly, but transaction costs often impede those deals. A licensee may need to divide a geographic license into smaller areas, separate frequencies within a band, negotiate operating protections, and obtain FCC consent. Recent policy changes have reduced some of these costs, but they have not removed them.[35] Secondary markets remain a viable yet underused means of reallocating spectrum rights.

C. Dynamic Spectrum Sharing

“Spectrum sharing” has become a prominent theme in recent policy debates, but the term can obscure more than it clarifies. Every spectrum allocation involves sharing. The FCC defines how users divide access across frequency, geography, and time.

An exclusive license is one form of sharing. It gives one operator priority rights within a defined frequency range and geographic area while requiring others to operate in adjacent channels, different locations, or at different times. A mobile carrier that wins a C-band license receives priority within its licensed area, but it still shares the broader band with neighboring licensees and other services. Users within the carrier’s network also share the same frequencies through network coordination.

Unlicensed systems rely on a different arrangement. Many short-range networks may operate on the same frequencies at the same time because their limited power and geographic reach allow nearby users to coexist. FCC technical rules govern how those devices share access without granting any one user priority rights.

Current policy debates often use “spectrum sharing” to describe a narrower model. Dynamic sharing uses technical and regulatory systems to permit multiple users to access the same frequencies in the same area, sometimes at nearly the same time. Rather than assigning static rights years in advance, these systems coordinate access as operating conditions change.[36]

Dynamic sharing has attracted more attention as regulators face fewer bands that can be cleared for exclusive commercial licenses. The 3.5 GHz CBRS band offers the most ambitious U.S. example. CBRS uses a three-tier hierarchy that assigns priority among several classes of users operating within the same frequencies.[37]

Federal incumbents, primarily naval radar systems along the coasts, occupy the highest tier. Their operations receive the strongest protection and may preempt lower-tier users when necessary. Priority Access Licensees occupy the second tier. They obtain geographically defined rights to specific channels and receive protection from other nonfederal users, but they must yield when incumbent federal systems need access.

General Authorized Access users occupy the third tier. They may operate without an individual license when federal incumbents and Priority Access Licensees are not using the relevant frequencies in the area. This structure incorporates unlicensed access into a coordinated sharing system rather than treating it as a residual use.

An automated frequency coordinator known as the Spectrum Access System (SAS) manages these priorities. The cloud-based system tracks incumbent activity, assigns channels, revokes access when higher-priority users need the band, and manages interference among the three tiers.[38] It translates the FCC’s regulatory hierarchy into operating decisions that respond to current conditions rather than relying entirely on a fixed band plan.

CBRS shows that dynamic sharing can accommodate changing demand while giving commercial operators some basis for investment. It also exposes the model’s costs. Priority Access Licensees face greater uncertainty than holders of traditional exclusive licenses because they may lose access when federal incumbents return. The system’s geographic and technical complexity may also exceed the capabilities of some operators.[39]

Those constraints may limit the services that shared bands can support. Applications involving artificial intelligence, augmented reality, or virtual reality may require consistent capacity and low latency. A licensee that must vacate a channel on short notice may struggle to provide that quality of service.

Dynamic sharing nevertheless expands the available set of allocation tools. Spectrum management need not rely exclusively on either traditional licenses or unlicensed access. Regulators can design sharing systems around the technical characteristics, incumbent uses, and commercial demands of a particular band.

The next challenge is to determine where dynamic coordination can improve spectrum use without creating unacceptable uncertainty. That question will grow more difficult as terrestrial, satellite, and airborne systems increasingly seek access to the same frequencies. Future sharing frameworks must provide clear priority rules, reliable coordination, and enough operating certainty to support investment.

D. Secondary Markets for Spectrum Rights

Operators may also sell or lease spectrum rights, either in full or in part. Secondary markets allow licensees to exchange those rights after the initial assignment without requiring the FCC to conduct a new allocation proceeding whenever market conditions change.[40]

The economic case for secondary markets is straightforward. The value of spectrum rights varies among users and over time. One licensee may lack the infrastructure, customers, or business model needed to use a license productively, while another could generate greater value from the same rights. A voluntary transaction can benefit both parties and move spectrum toward a more productive use with limited regulatory involvement.

The FCC embraced this logic in the early 2000s by establishing a framework for license transfers and leases. A licensee may assign an entire license to another holder, partition it by selling rights in part of the geographic service area, or disaggregate it by transferring part of the assigned frequency range. A licensee may also lease some or all of its operating rights while retaining the underlying license.[41]

Together, these mechanisms support a wide range of commercial arrangements. They allow parties to transfer only the rights relevant to a transaction rather than requiring the FCC to auction a newly configured license.

Secondary-market transactions have helped carriers assemble large, contiguous blocks of spectrum suitable for nationwide 5G networks. National carriers have acquired licenses from regional operators, cable companies, and other holders that could not use the rights as intensively.[42] The C-band reallocation also incorporated market mechanisms that allowed satellite operators to negotiate relocation terms and share in some of the value created by clearing the band for flexible use.[43]

Secondary markets support other arrangements as well, including private enterprise networks, utility communications, fixed-wireless services, neutral-host systems, and mobile services offered by cable companies or mobile virtual-network operators (MVNOs). Most transactions have involved permanent transfers or longer-term leases. More advanced automated coordination could eventually permit licensees to sell unused capacity for shorter periods, allowing another operator to access the band under privately negotiated terms.[44]

Secondary markets remain an incomplete substitute for FCC action. Spectrum transfers can involve substantial transaction costs, including negotiations over geography, frequency, interference protection, and agency approval. Concerns about spectrum warehousing also persist when licensees retain underused rights in anticipation of future gains rather than deploying service.[45]

Secondary markets offer even less help with underused federal spectrum. Federal agencies generally cannot sell or lease their assignments for direct financial benefit. Moving those frequencies toward higher-valued uses still requires political direction, interagency coordination, and formal reallocation.

IV. Spectrum for AI, AR, and VR Applications

Artificial intelligence (AI), augmented reality (AR), and virtual reality (VR) will all increase wireless demand, but they impose different technical requirements. AR and VR need sustained capacity, low latency, and reliable connections. AI generates additional traffic, including greater upload demand, while also offering tools to improve spectrum coordination and sharing.

No single frequency band or access model can satisfy all of these needs. Lower frequencies provide coverage, mid-band frequencies balance reach and capacity, and higher frequencies support dense, short-range connections. Licensed, unlicensed, and shared-access regimes each serve different applications. Satellite networks add another layer by extending coverage and backhaul beyond terrestrial systems. Spectrum policy must combine these bands, access models, and network layers into a coherent strategy that supports both investment and experimentation.

A. A Layered Spectrum Strategy

No single frequency band can meet every wireless-network requirement. Lower frequencies travel farther and penetrate buildings more effectively, while higher frequencies can support wider channels and greater capacity over shorter distances. Coverage, capacity, propagation, and reliability therefore involve unavoidable tradeoffs.

The next generation of wireless applications will require a layered spectrum strategy. Different frequency ranges must supply different capabilities within an integrated network, and each layer needs operating rights, interference protection, and regulatory certainty suited to its role.

1.        Low Band Below 1 GHz

Low-band frequencies travel long distances and penetrate buildings well, but their relatively narrow channels limit data capacity. They are poorly suited to immersive media and other data-intensive applications. They remain valuable for widespread, low-rate connectivity supporting distributed artificial intelligence (AI) and internet-of-things devices.

Low band works especially well when most processing occurs on the device and the application needs only a persistent connection. Its propagation characteristics allow users to maintain basic connectivity across large geographic areas and inside buildings.

2.        Lower Mid-Band at Roughly 1–6 GHz

Lower mid-band spectrum is central to current 5G deployments because it balances capacity and coverage. Reallocating it can be difficult, however, because many bands already support federal and commercial incumbents.

The lower 3 GHz band at 3.1–3.45 GHz, the 3.5 GHz CBRS band, and the C-band at 3.7–4.2 GHz can support wide-area mobile augmented reality (AR) and cloud-rendered virtual reality (VR).[46] Signals in these ranges travel far enough to serve suburban and exurban areas without the dense small-cell networks required at higher frequencies. They also penetrate buildings well enough to provide indoor service without always requiring a separate indoor network.

These bands can accommodate channels wide enough to support cloud-rendered AR, AI inference, and other data-intensive applications used by pedestrians, drivers, and workers on industrial floors.

3.        The 6 GHz Unlicensed Band at 5.925–7.125 GHz

The 6 GHz band provides 1,200 MHz of unlicensed spectrum and supports Wi-Fi 6E and Wi-Fi 7. It will play a major role in AR, VR, and many consumer-facing AI applications.

Regulators first authorized low-power indoor devices and very-low-power portable operations in the band. They later added higher-power, geofenced device categories aimed at AR and VR headsets, wearables, and similar products used indoors and outdoors.[47]

Most immersive applications operate indoors and across short distances. Those conditions make 6 GHz a natural fit for a substantial share of AR and VR traffic.

4.        Upper Mid-Band at Roughly 7–15 GHz

Upper mid-band frequencies offer wide channels without the severe range limits associated with millimeter wave. The 7.125–8.4 GHz segment, for example, can provide large contiguous blocks of spectrum while retaining propagation characteristics suitable for some wide-area coverage.

Standards organizations and industry groups have identified several extended mid-band ranges as candidates for future wireless use, including 7.125–8.5 GHz, 10.7–13.25 GHz, and 14–15.35 GHz.[48] North American studies have modeled extended-reality and other high-demand applications in the 3.1–3.45 GHz, 7.125–8.5 GHz, and 12.7–13.25 GHz bands.[49]

The shorter propagation range of upper mid-band frequencies may also make some portions suitable for unlicensed operation. Their principal advantage is the ability to support wide channels for extended reality without imposing the coverage limitations of millimeter wave.

5.        Millimeter Wave at Roughly 24–47 GHz

Millimeter-wave bands provide enormous capacity over short distances. They are well suited to dense environments such as venues, campuses, factories, and indoor networks, where fixed or tethered AR and VR systems can use very wide channels.[50]

These frequencies travel shorter distances and are easily blocked by walls, foliage, and other obstacles. They therefore complement mid-band coverage rather than replace it. As networks become denser and equipment improves, millimeter wave will likely assume a larger role in high-capacity wireless infrastructure.

6.        Sub-Terahertz and Terahertz Bands Above Roughly 100 GHz

Researchers are studying frequencies around 100–140 GHz and 275–325 GHz for use in the 2030s and beyond. These bands could eventually support terabit-per-second capacity for holographic communications, digital twins, and the most demanding extended-reality applications.[51]

Severe signal loss, strict line-of-sight requirements, and immature hardware will constrain these frequencies for the foreseeable future. Their most likely early uses involve short-range indoor communications and sensing. Mid-band frequencies will continue to carry most wide-area traffic.

B. Matching Allocation Models to Emerging Technologies

Making additional bands available for commercial use solves only part of the problem. The FCC must also decide whether to assign those bands through exclusive licenses, unlicensed access, or dynamic sharing. AI, AR, and VR depend on different combinations of coverage, capacity, reliability, and ease of access, so the choice among these models can shape which applications develop.

Exclusive licenses give operators defined rights to exclude other users, protect against harmful interference, and operate at higher power. That certainty supports the large investments required to build wide-area networks. Auctions can also generate substantial public revenue. Exclusive rights may limit entry, though, and can leave capacity underused when licensees do not fully employ their assigned frequencies.

Unlicensed bands take a different approach. Any device that complies with the FCC’s technical rules may operate without an individual license. The benefits extend across Wi-Fi, Bluetooth, and other short-range technologies, making them difficult to capture through a single auction bid or license valuation. Estimates credit unlicensed spectrum with generating trillions of dollars in annual economic value.

Dynamic-sharing systems occupy the middle ground. The 3.5 GHz CBRS band uses a three-tier coordination system that permits commercial users to operate when incumbent federal radar systems do not need the frequencies. Automated sensing and database tools manage access and enforce priority rights.

These allocation models support different uses. Indoor, tethered, and short-range immersive applications align closely with unlicensed spectrum, particularly the 6 GHz band. Outdoor and mobile AR applications, including heads-up navigation and mobile mixed reality, depend more heavily on the coverage and quality-of-service protections available through licensed mid-band networks.

AI applications use both. Latency-sensitive edge inference may rely on licensed networks when mobility and reliability matter. Large model updates, data synchronization, and stationary applications can shift substantial traffic to unlicensed networks. No single allocation model can support the full range of immersive and AI-enabled services.

The One Big Beautiful Bill Act, signed in July 2025, restored the FCC’s general spectrum-auction authority after it expired in March 2023. The law directed the federal government to identify at least 800 MHz within the 1.3–10.5 GHz range for auction and required an auction of upper C-band spectrum.[52]

Congress excluded the heavily contested 3.1–3.45 GHz and 7.4–8.4 GHz federal bands from that mandate. It did not provide similar statutory protection for the 5.925–7.125 GHz unlicensed band or CBRS. That omission prompted concern among Wi-Fi and shared-access advocates that frequencies supporting AR, VR, and edge-AI applications could later be converted to exclusive licensed use.[53] The Trump administration has continued to support preserving the full 6 GHz band for Wi-Fi.[54]

The spectrum needs of AI, AR, and VR therefore call for a layered portfolio of both frequency bands and access models. Licensed mid-band spectrum supplies wide-area coverage and predictable service. The 6 GHz unlicensed band and millimeter-wave frequencies provide dense, short-range capacity for many immersive applications. Dynamic sharing can expand access where full clearing or exclusive licensing would prove impractical.

The allocation balance matters as much as the amount of spectrum available. A policy that favors exclusive licensing too heavily could sacrifice the dispersed benefits of Wi-Fi and shared systems. A policy that relies too heavily on unlicensed or shared access could weaken the certainty needed for large network investments and applications requiring guaranteed service. The FCC should preserve a deliberate mix of licensed, unlicensed, and shared access suited to the technical demands of each band and application.

C. Satellite Networks as a Third Connectivity Layer

A full account of spectrum demand must include low-Earth-orbit (LEO) satellite constellations, which are becoming a third connectivity layer alongside fixed and mobile networks. Systems such as Starlink and Project Kuiper operate roughly 300 to 2,000 kilometers above Earth, far below geostationary satellites at about 36,000 kilometers.[55] Their lower altitude reduces latency, often to less than 100 milliseconds, which can support many interactive applications.[56]

Connectivity-focused LEO systems rely primarily on the Ku band at 12–18 GHz for user-terminal links and the Ka band at 26.5–40 GHz for gateway and feeder links. Higher-capacity Q/V bands at 37.5–51 GHz are emerging for next-generation feeder links, though they are more vulnerable to signal loss.[57] Optical intersatellite links can also carry traffic between satellites, reducing pressure on radio frequencies and avoiding some ground-based routing delays.

These systems operate through coordinated, licensed allocations governed internationally by the International Telecommunication Union’s (ITU) nongeostationary fixed-satellite-service framework. Regulators do not generally assign these frequencies through national auctions. Instead, they establish interference limits and coordination rules that allow multiple constellations to use the same bands.[58]

Direct-to-device connectivity and Supplemental Coverage from Space are developing even faster. Rather than relying only on dedicated satellite bands and specialized terminals, these services allow ordinary mobile devices to connect directly to satellites when no terrestrial cell site is available. They may use terrestrial mobile frequencies or L- and S-band mobile-satellite-service spectrum. Since the FCC adopted an initial regulatory framework in 2024, the model has entered commercial service and prompted multibillion-dollar transactions involving mobile-satellite spectrum.[59]

For AI, AR, and VR, LEO service will complement terrestrial networks rather than replace them. Its clearest role is extending coverage to rural, maritime, aviation, and disaster-affected areas where fiber and dense cellular networks are unavailable. It can also provide resilient backhaul to terrestrial edge nodes that perform AI inference and cache content closer to users.

Latency will limit the most demanding immersive applications. A LEO round trip of roughly 20 to 50 milliseconds can support cloud-rendered or social VR, video, telemetry, AI inference, and model synchronization. It remains too slow for tightly coupled AR and VR experiences that require motion-to-photon latency below about 15 milliseconds. Those applications will continue to depend on on-device processing or nearby terrestrial edge infrastructure.[60]

Direct-to-device service may reach lightweight AR wearables and always-on AI assistants sooner. It could support off-grid notifications, positioning, and low-rate queries well before it can deliver full immersive video.

Satellite spectrum presents management problems that differ from those affecting terrestrial networks. Terrestrial licenses often divide frequencies into exclusive blocks. Satellite systems instead depend on coordination among operators.[61]

ITU rules assign priority largely by filing date. In the United States, the FCC uses processing rounds that give earlier-authorized systems priority over later applicants. A later system must coordinate with earlier constellations or comply with default protection standards.[62]

Equivalent power-flux-density (EPFD) limits add another constraint. These rules cap the combined power that nongeostationary systems may direct toward geostationary satellites, even when no particular geostationary link faces a demonstrated threat. The limits can restrict how many satellites a constellation may direct toward one area and reduce total network capacity.[63]

In 2026, the FCC moved toward a performance-based framework for important Ku- and Ka-band frequencies. The new approach relies on a default 3% throughput-degradation threshold and good-faith bilateral coordination. The agency also granted interim waivers to leading constellations before finalizing the rules.[64]

Coordination still depends heavily on negotiation among operators. As more constellations enter service and orbital density rises, the number of overlapping systems and potential interference relationships grows quickly. U.S. rules are also moving ahead of the international framework, which continues to rely on EPFD limits.

The priority system favors incumbents. Earlier filings receive stronger protection, which rewards operators that apply and deploy first. It also encourages speculative filings because an operator can secure priority before putting spectrum to productive use.

Deployment milestones seek to limit that behavior by requiring operators to launch a specified share of an authorized constellation within fixed periods. Those requirements also raise the capital needed to enter the market and may favor well-financed incumbents over smaller or later challengers.

Rules designed to prevent interference can therefore restrict entry when they grant durable advantages to early filers or impose high upfront deployment costs. Satellite policy faces the same basic tradeoff as terrestrial spectrum policy. Regulators must protect existing operations without allowing coordination rules to preserve underused rights or exclude more productive entrants.

D. AI as Both Spectrum User and Management Tool

Artificial intelligence complicates the usual choice among licensed, unlicensed, and shared access because it affects both spectrum demand and spectrum management. Generative-AI and edge-AI applications can produce substantial bursts of traffic across mid- and upper-mid-band frequencies. They may also require far more upload capacity than current consumer applications, placing pressure on networks designed primarily for downloads.

AI may also help networks use existing frequencies more efficiently. Machine-learning systems can identify unused capacity, predict interference conditions, and adjust channel access in real time. These tools could allow secondary users to operate without causing harmful interference to incumbents.

Research on reinforcement learning and hybrid optimization for 6G cognitive-radio networks reports gains in spectral efficiency and reductions in interference.[65] If these methods mature and regulators accept them, spectrum policy will increasingly focus on how efficiently users can share a band, alongside decisions about whether to license it exclusively or permit unlicensed access.

AI therefore occupies an unusual role. It increases demand for wireless capacity while offering tools that may ease some of the resulting congestion. Its contribution to dynamic coordination could become especially valuable as immersive applications place greater demands on scarce frequencies.

V. A Spectrum Policy Agenda for Emerging Technologies

Realizing the benefits of artificial intelligence, augmented reality, and virtual reality will require more than adding spectrum to a federal pipeline. Policymakers must preserve a balanced mix of licensed, unlicensed, and shared access, move underused federal bands into commercial use, improve coordination between the FCC and the NTIA, and replace reflexive worst-case interference analysis with risk-informed methods.

The United States must also engage more coherently in international spectrum policy. Decisions made through the International Telecommunication Union and the World Radiocommunication Conference shape global equipment markets, national-security interests, and competition with China. Each band presents distinct technical and economic questions, but these principles should guide the broader policy agenda.

A. Preserve a Balanced Mix of Access Models

AI, AR, and VR rely on different wireless capabilities at different times. Regulators should therefore preserve a mix of frequency bands and access models rather than favoring one regime across the board.

Exclusive licenses support high-powered, wide-area networks and provide the interference protection needed for capital-intensive deployment. Those features matter for mobile AR and latency-sensitive AI inference. Unlicensed spectrum supplies low-cost, permissionless capacity for indoor and short-range immersive applications, as well as AI workloads that generate substantial traffic in both directions. Dynamically shared bands can accommodate incumbents while opening additional capacity to commercial users. Each model serves a distinct purpose.

The political process does not value these benefits evenly. Auctions produce large, visible, and immediate federal revenue. The gains from unlicensed and shared spectrum appear more diffusely through consumer surplus, lower entry barriers, and new devices and services. They are also harder to estimate in advance.[66] When Congress uses auction receipts to offset federal spending, that asymmetry creates pressure to convert more bands to exclusive licenses.

That pressure bears directly on the 6 GHz unlicensed band and the 3.5 GHz CBRS framework. Both support immersive and edge-computing applications, yet both remain exposed to proposals for licensed reallocation.

An excessive tilt toward unlicensed or shared use would create different costs. Without enough exclusive-use spectrum and reliable interference protection, operators may lack the certainty needed to invest billions of dollars in wide-area networks.[67] Mobile immersive services cannot become widely available without that infrastructure.

Allocation decisions should therefore account for consumer welfare and total economic value rather than auction revenue alone. Regulators should compare the full benefits and costs of exclusive licensing, unlicensed access, and dynamic sharing in each band.

The same approach should govern satellite and direct-to-device systems. Rules for nonterrestrial networks should promote competition and economic value while ensuring that the satellite layer complements terrestrial networks rather than dividing spectrum policy into separate and inconsistent regimes.

B. Reform Federal Spectrum Reallocation

Much of the prime mid-band spectrum needed for AI, AR, and VR remains assigned to federal users, especially the Department of Defense (DOD). Reallocating those frequencies for commercial use requires coordination between the NTIA and the FCC. The One Big Beautiful Bill Act was designed to accelerate that process, but its structure shows why federal reallocation remains the weakest part of the spectrum pipeline. If that process stalls, pressure may shift toward nonfederal bands.

The law’s 800 MHz pipeline draws on two sources.[68] It directs NTIA to identify at least 500 MHz of federal spectrum between 1.3 and 10.5 GHz, while excluding the 3.1–3.45 GHz and 7.4–8.4 GHz bands reserved for the DOD. NTIA must identify at least 200 MHz within two years, with FCC auctions scheduled on staggered deadlines extending into the early 2030s.

The remaining 300 MHz must come from nonfederal spectrum under FCC control. That tranche includes a requirement to auction at least 100 MHz of upper C-band spectrum between 3.98 and 4.2 GHz.

The statute does not require the auction of any particular federal band. It establishes a total target and directs NTIA to identify qualifying frequencies that the FCC must later auction. Meeting the federal target therefore depends on identifying suitable bands and relocating or accommodating incumbent systems.

That process requires cooperation from the agencies using those frequencies. Clearing a federal band may require moving systems into adjacent frequencies, modifying military equipment, or redesigning operations. The executive branch can direct NTIA to pursue relocation, but it cannot easily dismiss operational and national-security objections from agencies that must move.

If the DOD refuses to accommodate a relocation, NTIA may struggle to identify the full 500 MHz. A statutory target cannot make frequencies available when the responsible agency will not relinquish them. Congress’s exclusion of the lower 3 GHz and 7.4–8.4 GHz bands illustrates the problem.[69] Those exemptions responded to legitimate defense concerns, but similar objections may arise as NTIA examines other bands.

The FCC can deliver only the 300 MHz drawn from nonfederal spectrum without relying on another agency. If the federal portion proves difficult to clear, auction deadlines, projected revenue, and political pressure may encourage the FCC to draw more heavily on bands within its own jurisdiction.

That pressure could fall on frequencies already supporting immersive and edge-AI applications, including the 3.55–3.7 GHz CBRS band and the 5.925–7.125 GHz unlicensed band. Congress declined to protect either band in the final legislation. A process intended to reclaim underused federal spectrum could therefore consume commercial, shared, or unlicensed spectrum when federal incumbents resist.

Durable reallocation requires federal-agency cooperation, adequate funding for relocation, and clear White House direction. Regulators must also reduce the procedural barriers that delay transitions once agencies agree to move.

Congress created the Spectrum Relocation Fund (SRF) through the Commercial Spectrum Enhancement Act of 2004 to reimburse federal agencies for relocation costs.[70] The fund directs a portion of auction proceeds toward expenses associated with moving federal systems. It was meant to remove a recurring obstacle by ensuring that agencies would not resist reallocation merely because their budgets could not absorb the cost.

The fund has not operated smoothly.[71] Before receiving money even to study whether relocation is feasible, an agency must prepare a detailed pipeline plan, obtain review from the Technical Panel, secure approval from the director of the Office of Management and Budget, and complete a mandatory congressional-notification period. These requirements can delay the preliminary studies needed to determine whether a band can be cleared or shared.[72]

The Trump administration has taken steps to improve the process, but several of its most burdensome requirements, including the 60-day congressional-notification period, are statutory. Administrative reforms alone cannot remove them.[73]

Congress should streamline the review and notification process while preserving meaningful oversight. Faster access to relocation funding would allow NTIA to study bands sooner, shorten auction timelines, increase license values by giving bidders greater certainty, and bring additional mid-band spectrum into commercial use more quickly.

C. Strengthen FCC-NTIA Coordination

The aviation-altimeter dispute surrounding the lower C-band exposed serious weaknesses in coordination between the FCC and the NTIA. The Biden administration responded with several reforms. In February 2022, the agencies launched a Spectrum Coordination Initiative. Six months later, they signed an updated memorandum of understanding, the first revision in nearly two decades.[74]

The agreement formalized coordination at both the leadership and staff levels. It required the FCC chair and NTIA administrator to meet at least quarterly and their staffs to meet monthly. It also created procedures for sharing plans up to a year in advance and extended the notice period for proposals that could cause interference from 15 to 20 business days. These measures sought to identify conflicts early rather than allow them to erupt near the end of a proceeding.

The 2023 National Spectrum Strategy and its 2024 implementation plan expanded that effort. They called for longer-term interagency planning, peer-reviewed models for evaluating reallocation and coexistence, and a pipeline of bands for detailed study. They also proposed a national testbed for dynamic spectrum sharing, including the use of artificial intelligence to support real-time coordination.[75]

Coordination remains difficult, and further improvement will be necessary to carry out the One Big Beautiful Bill Act. When a band can be cleared, effective planning determines whether federal systems relocate on time and at a reasonable cost. When clearing is impractical, commercial access may depend on dynamic sharing. That approach works only when the FCC, NTIA, and incumbent agencies jointly develop, test, and trust the technical framework.

Both reallocation and sharing depend on durable cooperation grounded in a common technical record. The FCC and NTIA should continue refining their procedures, disclose concerns early, and resolve disputes before agencies and private firms commit substantial resources. Without that discipline, the new statutory pipeline may reproduce the delays and uncertainty that the coordination reforms were designed to prevent.

D. Adopt Risk-Informed Interference Analysis

Every spectrum reallocation and sharing arrangement depends on a threshold question. How much harmful interference would a proposed use cause to incumbents operating in or near the band? The method used to answer that question often determines whether regulators permit new entry at all.

For decades, agencies have relied heavily on deterministic, worst-case analysis.[76] This method assumes that transmitters operate at maximum power, at the least favorable distance and geometry, under the most adverse propagation conditions, and at the same time. Such caution imposed little cost when spectrum was lightly used. In today’s congested bands, it can block entry or impose severe operating restrictions even when the probability of harm is remote.

Worst-case analysis treats every conceivable event as equally relevant, regardless of likelihood. A harmful-interference scenario may require several unlikely conditions to occur simultaneously, including peak power, minimum separation, and unusually adverse propagation. Evaluating only the most severe possible outcome systematically overstates risk and gives incumbents a potent means of resisting new services. An alarming hypothetical can derail entry even when expected harm is negligible.

Risk-informed interference assessment asks a broader set of questions.[77] What harmful events could occur? How likely are they? What consequences would follow? Aviation, nuclear, and environmental regulators have used quantitative risk assessment for decades to evaluate hazards whose probabilities and consequences vary.

The process generally has four steps:

1. Identify significant interference hazards.
2. Define a measure for the consequences of each hazard.
3. Estimate each hazard’s likelihood and consequences using realistic deployment densities, equipment-use patterns, and statistical propagation models.
4. Combine those results into a probabilistic assessment of risk.[78]

This method does not guarantee approval for a new entrant. It does, however, give regulators a fuller account of likely outcomes rather than allowing one extreme scenario to control the decision.

Risk-informed analysis is especially valuable for dynamic-sharing systems. Automated coordinators such as the CBRS Spectrum Access System and the 6 GHz automated frequency-coordination system must assign channels and manage interference in real time. They cannot resolve every potential conflict through an individual proceeding. They need objective criteria that account for both the probability and severity of interference.

Clear thresholds give incumbents and entrants greater certainty because the rules governing coexistence are established before deployment. Worst-case assumptions instead produce unnecessarily large exclusion zones and protection margins, wasting much of the capacity that sharing is intended to make available.

The FCC used probabilistic interference analysis when it opened the 6 GHz band to unlicensed devices despite incumbents’ worst-case objections. The U.S. Circuit Court of Appeals for the D.C. Circuit upheld that approach as reasonable, confirming that quantitative risk assessment can provide a legally defensible basis for spectrum decisions.[79] The National Spectrum Strategy likewise calls for data-driven, risk-informed compatibility and coexistence studies.[80]

Regulators should make risk-informed analysis the default for reallocation and sharing proceedings. Agencies should develop internal expertise, publish their models and assumptions, and allow independent researchers and affected parties to test the results. Greater transparency would improve both the quality and credibility of interference assessments.

Artificial intelligence can strengthen this approach. One promising tool is the spectrum digital twin, a continuously updated virtual model of a real radio environment. Regulators, operators, and researchers can use such models to simulate electromagnetic systems under many operating conditions before making allocation decisions.[81]

A digital twin can represent environments ranging from a single building or industrial campus to a national network. Rather than producing one static worst-case estimate, it can run thousands of simulations across realistic combinations of device locations, power levels, traffic patterns, and propagation conditions. The result is a probabilistic account of interference risk that can also identify ways to improve spectrum use while protecting incumbents.

Risk-informed analysis does not discount catastrophic harms. Some systems support safety-of-life or national-security functions for which even a rare failure could carry grave consequences. A sound risk assessment weighs severity alongside likelihood and can impose stringent protections when potential harm is extreme.[82]

Its advantage lies in distinguishing those cases from ordinary disputes in which incumbents invoke remote hazards to delay competition. Protection should correspond to the actual risk posed by a proposed use rather than applying the same degree of caution across every band and service.

The consequences for AI, AR, and VR are substantial. Much of the mid-band capacity these applications need will depend on reallocation and sharing. Worst-case analysis can keep usable frequencies unavailable because of improbable combinations of events. A risk-informed standard would protect critical systems while allowing regulators to make more productive use of spectrum.

E. Strengthen U.S. Leadership in International Spectrum Policy

Spectrum policy does not stop at national borders. Every four years, the International Telecommunication Union (ITU) convenes the World Radiocommunication Conference (WRC) to revise the international Radio Regulations and coordinate how countries allocate frequency bands.[83] For AI, AR, and VR, this process matters for two reasons. It creates global economies of scale, and it has become an arena for strategic competition, especially with China.

When countries harmonize a band for the same use, manufacturers can design chipsets, devices, and base stations to a common specification. Producing equipment for a global market lowers costs, improves interoperability, and speeds adoption. Those gains can make AR and VR hardware more affordable and allow AI-connected devices to reach a broader market.

China has pursued a coordinated, state-backed effort to place its firms in leadership roles within standards bodies and incorporate Chinese technical proposals into 5G and early 6G standards. Its centralized regulatory system can reallocate prime bands more quickly than the United States. Chinese firms also hold roughly 40% of declared 5G standard-essential patents.

Patent-declaration counts provide an imperfect measure of influence. Firms often declare patents that later prove nonessential, independent reviews find that only a fraction of declared 5G patents are truly essential, and Chinese portfolios contain fewer high-value radio-access-network patents. Even so, the volume of Chinese participation reflects sustained public and private investment in shaping technical standards. China has also sought satellite-interference limits that could constrain U.S. LEO deployments.[84]

The United States and its allies rely more heavily on private firms and consensus-based institutions. That model can produce strong technical standards, but domestic disputes sometimes weaken the U.S. position. Disagreements between commercial and federal spectrum users have prevented the United States from advancing unified mid-band proposals at some WRCs.[85]

Technical standards also influence which vendors supply networks around the world. Standards shaped disproportionately by Chinese proposals may favor Chinese equipment makers. That creates supply-chain, espionage, and surveillance risks of the same kind that prompted U.S. programs to remove Huawei and ZTE equipment.[86] Those risks may prove especially acute in developing countries with limited resources to assess vendors or replace compromised infrastructure.

International proceedings must also protect U.S. defense needs. Bands reserved domestically for military systems may face pressure for identification as international mobile telecommunications spectrum. Resisting such proposals when they would impair assured military access is itself a national-security objective.

The 6 GHz band illustrates the broader commercial competition. The United States authorized the band for unlicensed use in 2020, supporting Wi-Fi and a growing market for unlicensed devices and infrastructure. China has favored licensed mobile networks and seeks broader international adoption of Chinese 5G and 6G equipment. Decisions about international harmonization can therefore influence whether future markets favor licensed cellular systems or the unlicensed technologies in which U.S. firms hold important advantages.

Sustained, coordinated U.S. engagement at the ITU and WRC is necessary to secure globally harmonized bands, protect critical federal operations, and promote technical rules that support competition. The United States must resolve domestic disagreements early enough to present coherent proposals internationally. Otherwise, other countries will shape the standards and allocations on which future AI, AR, and VR networks depend.

VI. Conclusion

Radio spectrum is essential infrastructure for the AI era, but it cannot support next-generation applications alone. Wireless capacity depends on fiber backhaul, edge computing, data centers, devices, and power systems developing alongside it. Spectrum policy must ensure that this communications layer keeps pace with the rest of the infrastructure chain.

The United States will need more than additional frequencies. It needs a layered mix of low-, mid-, and high-band spectrum, paired with licensed, unlicensed, and shared-access models suited to different technical demands. Terrestrial and satellite networks must complement one another. Exclusive rights must provide enough certainty to support investment, while unlicensed and shared bands preserve access for Wi-Fi, immersive applications, new devices, and smaller entrants.

Current institutions can frustrate those goals through accumulated delay and uncertainty. Divided authority between the FCC and the NTIA can slow federal reallocation. The Spectrum Relocation Fund can delay the studies it was created to finance. Worst-case interference analysis can block productive uses based on remote hazards. Domestic disagreements can also weaken U.S. influence at the World Radiocommunication Conference, where international allocations and standards shape equipment markets for decades.

These problems do not require rebuilding the entire spectrum-management system. Policymakers can improve it through sustained FCC-NTIA coordination, faster access to relocation funding, risk-informed interference analysis, transparent technical models, spectrum digital twins, and a coherent U.S. strategy at the International Telecommunication Union. Clear White House direction will be necessary to align agencies whose institutional incentives often diverge.

The OBBBA’s 800 MHz pipeline provides a mandate, but its value will depend on implementation. Federal resistance could delay reallocation and shift pressure toward productive unlicensed and shared bands. Poorly designed sharing rules could discourage investment. Coordination frameworks that entrench early satellite entrants could restrict competition. Each decision should account for total economic value rather than auction revenue or incumbent protection alone.

Success would produce an integrated spectrum system in which licensed, shared, and unlicensed access serve complementary purposes, satellite networks extend terrestrial coverage, and U.S. proposals shape global standards. Failure would preserve underused federal assignments, expose productive bands to reallocation pressure, and leave new services waiting while agencies reopen old interference disputes. American leadership in AI, augmented reality, virtual reality, and wireless communications will depend in part on whether spectrum governance can move with the technologies it is meant to support.

[1] Kristian Stout, Infrastructure Is Destiny: The Geography of the Next Technology Race, Int’l Ctr. for L. & Econ. (May 26, 2026), https://laweconcenter.org/resources/infrastructure-is-destiny-the-geography-of-the-next-technology-race.

[2] Id.

[3] Peter Elkind, Inside the Government Fiasco That Nearly Closed the U.S. Air System, ProPublica (May 26, 2022), https://www.propublica.org/article/fcc-faa-5g-planes-trump-biden; Jeffrey Westling, Primer: A Staffer’s Guide to Spectrum Auction Reauthorization, Am. Action Forum (Feb. 20, 2025), https://www.americanactionforum.org/insight/primer-a-staffers-guide-to-spectrum-auction-reauthorization.

[4] Jonathan E. Nuechterlein & Philip J. Weiser, Digital Crossroads: Telecommunications Law and Policy in the Internet Age (2d ed. 2013).

[5] Jean Pierre De Vries & Jeffrey Westling, Not a Scarce Natural Resource: Alternatives to Spectrum-Think, TPRC45 (Oct. 2, 2017), https://ssrn.com/abstract=2943502.

[6] Id.

[7] See, e.g., Mustafa Riza Akdeniz et al., Millimeter Wave Channel Modeling and Cellular Capacity Evaluation, 32 IEEE J. on Selected Areas in Commc’ns 1164 (2014) (discussing millimeter-wave spectrum as a means to increase capacity by providing access to wider channels).

[8] Fed. Commc’ns Comm’n, Office of Eng’g & Tech., OET Bulletin No. 70, Millimeter Wave Propagation: Spectrum Management Implications (July 1997), https://transition.fcc.gov/Bureaus/Engineering_Technology/Documents/bulletins/oet70/oet70a.pdf.

[9] C.E. Shannon, A Mathematical Theory of Communication, 27 Bell Sys. Tech. J. 379 (1948).

[10] 47 C.F.R. § 15.3 (2024) (defining “harmful interference” as “[a]ny emission, radiation or induction that endangers the functioning of a radio navigation service or of other safety services or seriously degrades, obstructs or repeatedly interrupts a radiocommunications service operating in accordance with [Title 47]”).

[11] 47 U.S.C. §§ 154, 302a, 303, 304, 307, 336, 544a.

[12] Nuechterlein & Weiser, supra note 4; see also Thomas W. Hazlett, Assigning Property Rights to Radio Spectrum Users: Why Did FCC License Auctions Take 67 Years?, 41 J.L. & Econ. 529 (1998).

[13] See U.S. Gov’t Accountability Off., GAO-24-106325, Spectrum Management: Key Practices Could Help Address Challenges to Spectrum Sharing (July 2024), https://www.gao.gov/assets/gao-24-106325.pdf.

[14] See, e.g., Fed. Commc’ns Comm’n, FCC Announces Winning Bidders in C-Band Auction; Auction 107 of 3.7 GHz Service Licenses Yields Over $81 Billion in Gross Bids (Feb. 25, 2021), https://www.fcc.gov/document/fcc-announces-winning-bidders-c-band-auction.

[15] Jeffrey Westling, Eric Fruits & Kristian Stout, Comments of the Int’l Ctr. for L. & Econ., Empowering Local Broadcast TV Stations to Meet Their Public Interest Obligations: Exploring Market Dynamics Between National Programmers and Their Affiliates, MB Docket No. 25-322 (Dec. 10, 2025), https://laweconcenter.org/wp-content/uploads/2025/12/Network-Affiliation-Comments-of-the-International-Center-for-Law-Economics.pdf.

[16] NCTA—The Internet & Television Ass’n, Spectrum & Wi-Fi, https://www.ncta.com/priorities/dynamic-networks/spectrum-and-wi-fi (last modified Feb. 12, 2026).

[17] Hazlett, Assigning Property Rights to Radio Spectrum Users, supra note 12.

[18] Paul Milgrom, Jonathan Levin & Assaf Eilat, The Case for Unlicensed Spectrum 3–4 (Oct. 12, 2011), https://web.stanford.edu/~jdlevin/Papers/UnlicensedSpectrum.pdf.

[19] Jeffrey Westling, Rivalrous Regulators: Historical Analysis of the Dual Agency Approach to Spectrum Management, R Street Pol’y Study No. 231 (Oct. 4, 2021), https://www.rstreet.org/2021/10/04/rivalrous-regulators-historical-analysis-of-the-dual-agency-approach-to-spectrum-management.

[20] R.H. Coase, The Federal Communications Commission, 2 J.L. & Econ. 1, 6 (1959).

[21] 47 U.S.C. § 303 (granting the Federal Communications Commission authority to regulate radio communications “as public convenience, interest, or necessity requires”); 47 U.S.C. § 307(a) (requiring the Commission to grant broadcast licenses “if public convenience, interest, or necessity will be served thereby”); 47 U.S.C. § 309(a) (requiring the Commission to grant license applications when “public interest, convenience, and necessity will be served”).

[22] R.H. Coase, The Interdepartment Radio Advisory Committee, 5 J.L. & Econ. 17 (1962).

[23] Westling, Rivalrous Regulators, supra note 19.

[24] Id. at 1.

[25] Press Release, Sen. Deb Fischer, Defense Officials to Fischer: If DOD Is Forced to Vacate Spectrum Frequencies, U.S. Would Assume High Level of Risk for Homeland Defense (May 13, 2025), https://www.fischer.senate.gov/public/index.cfm/2025/5/defense-officials-to-fischer-if-dod-is-forced-to-vacate-spectrum-frequencies-u-s-would-assume-high-level-of-risk-for-homeland-defense.

[26] MOBILE NOW Act, S. 19, § 5, 115th Cong. (2017).

[27] Jeffrey Westling, It Takes Two to Make a Proceeding Go Slow, R Street Inst. (July 20, 2020), https://www.rstreet.org/commentary/it-takes-two-to-make-a-proceeding-go-slow.

[28] Seth L. Cooper, D.C. Circuit Upholds FCC’s 5.9 GHz Order Reallocating Spectrum for Unlicensed Use, Fed. Soc’y Blog (Sept. 12, 2022), https://fedsoc.org/commentary/fedsoc-blog/d-c-circuit-upholds-fcc-s-5-9-ghz-order-reallocating-spectrum-for-unlicensed-use.

[29] See generally Fed. Commc’ns Comm’n, Licensing, https://www.fcc.gov/licensing (last visited June 8, 2026).

[30] Eric Fruits & Kristian Stout, Comments of the Int’l Ctr. for L. & Econ., Modernizing Spectrum Sharing for Satellite Broadband, SB Docket No. 25-157 (July 28, 2025), https://laweconcenter.org/wp-content/uploads/2025/07/2025-EPFD-Comments.pdf.

[31] Randall Berry, Thomas W. Hazlett, Michael L. Honig & J. Nicholas Laneman, Evaluating the CBRS Experiment (Aug. 1, 2023) (unpublished manuscript), https://ssrn.com/abstract=4528763.

[32] Milgrom, Levin & Eilat, supra note 18, at 3–5 (“[U]nlicensed spectrum is an enabling resource. It provides a platform for innovation upon which innovators may face lower barriers to bringing wireless products to market because they are freed from the need to negotiate with exclusive license holders.”).

[33] 47 C.F.R. pt. 15 (2024).

[34] Opensignal, USA, Converged Experience, April 2026 (Apr. 30, 2026), https://insights.opensignal.com/2026/04/usa-converged-experience-april-2026/dt.

[35] Partitioning, Disaggregation, and Leasing of Spectrum, Report and Order and Second Further Notice of Proposed Rulemaking, WT Docket No. 19-38, FCC 22-53, 37 FCC Rcd. 8825 (2022).

[36] Nat’l Telecommc’ns & Info. Admin., Advanced Dynamic Spectrum Sharing Demonstration in the National Spectrum Strategy (June 28, 2024), https://www.ntia.gov/issues/national-spectrum-strategy/advanced-dynamic-spectrum-sharing-demonstration-in-the-national-spectrum-strategy.

[37] Amendment of the Commission’s Rules with Regard to Commercial Operations in the 3550-3650 MHz Band, Report and Order and Second Further Notice of Proposed Rulemaking, GN Docket No. 12-354, FCC 15-47, 30 FCC Rcd. 3959 (2015).

[38] Id. ¶¶ 301–378.

[39] Berry et al., supra note 31 (“The U.S. experience in the CBRS band has not evinced efficiencies in reducing delays, or in enhancing the flow of large amounts of bandwidth to higher valued uses. In terms of what rival nations have accomplished in their 3.5 GHz allocations, which uniformly rely primarily on awarding flexible, exclusive rights and which do not include the U.S.’s approximately 50-50 split between licensed and unlicensed, substantially more spectrum has been put into play [via exclusive licensing].”).

[40] Martin Cave, Chris Doyle & William Webb, Spectrum Trading: Secondary Markets, in Essentials of Modern Spectrum Management 85 (2007).

[41] Promoting Efficient Use of Spectrum Through Elimination of Barriers to the Development of Secondary Markets, Report and Order and Further Notice of Proposed Rulemaking, WT Docket No. 00-230, FCC 03-113, 18 FCC Rcd. 20604 (2003); see also Promoting Efficient Use of Spectrum Through Elimination of Barriers to the Development of Secondary Markets, Second Report and Order, Order on Reconsideration, and Second Further Notice of Proposed Rulemaking, WT Docket No. 00-230, FCC 04-167, 19 FCC Rcd. 14165 (2004).

[42] Linda Hardesty, Sprint PCS Spectrum Looks Like Boon for T-Mobile, Fierce Network (Aug. 6, 2019), https://www.fierce-network.com/5g/sprint-pcs-spectrum-looks-like-boon-for-t-mobile.

[43] Expanding Flexible Use of the 3.7 to 4.2 GHz Band, Report and Order and Order of Proposed Modification, GN Docket No. 18-122, FCC 20-22, 35 FCC Rcd. 2343, ¶ 178 (2020), https://www.fcc.gov/document/fcc-expands-flexible-use-c-band-5g-0.

[44] See, e.g., Lili Cao & Haitao Zheng, Towards Real-Time Dynamic Spectrum Auctions, 87 Computer Networks 1 (2008) (proposing a low-complexity auction framework for allocating spectrum in real time among large numbers of wireless users with dynamic traffic).

[45] See John W. Mayo & Scott Wallsten, Enabling Efficient Wireless Communications: The Role of Secondary Spectrum Markets, 22 Info. Econ. & Pol’y 61 (2010); see also U.S. Gov’t Accountability Off., GAO-14-236, Spectrum Management: FCC’s Use and Enforcement of Buildout Requirements (Feb. 2014), https://www.gao.gov/assets/gao-14-236.pdf (discussing spectrum warehousing concerns).

[46] 5G Ams., From Spectrum Scarcity to Scale: Why the Hunt for Mid-Band Spectrum Is Defining the 5G-to-6G Transition (Dec. 3, 2025), https://www.5gamericas.org/from-spectrum-scarcity-to-scale-why-the-hunt-for-mid-band-spectrum-is-defining-the-5g-to-6g-transition.

[47] Unlicensed Use of the 6 GHz Band; Expanding Flexible Use in Mid-Band Spectrum Between 3.7 and 24 GHz, Report and Order and Further Notice of Proposed Rulemaking, ET Docket No. 18-295, GN Docket No. 17-183, FCC 20-51, 35 FCC Rcd. 3852 (2020); Unlicensed Use of the 6 GHz Band; Expanding Flexible Use in Mid-Band Spectrum Between 3.7 and 24 GHz, Fourth Report and Order and Third Further Notice of Proposed Rulemaking, ET Docket No. 18-295, GN Docket No. 17-183, FCC 26-1 (Jan. 30, 2026).

[48] Nokia, Spectrum for 6G Explained, https://www.nokia.com/6g/spectrum-for-6G-explained (last visited June 8, 2026).

[49] ATIS Next G Alliance, Derivation of Spectrum Needs for 6G (June 2024), https://nextgalliance.org/white_papers/6g-spectrum-considerations.

[50] Maxim Susloparov, Artem Krasilov & Evgeny Khorov, Providing High Capacity for AR/VR Traffic in 5G Systems with Multi-Connectivity, IEEE PIMRC 2022 (2022), https://arxiv.org/pdf/2208.08277.

[51] Ian F. Akyildiz, Chong Han, Zhen Hu, Shuai Nie & Josep M. Jornet, Terahertz Band Communication: An Old Problem Revisited and Research Directions for the Next Decade, 70 IEEE Trans. Commc’ns 4250 (2022).

[52] One Big Beautiful Bill Act, Pub. L. No. 119-21, 139 Stat. 72 (2025).

[53] NCTA—The Internet & Television Ass’n, Why Protecting Wi-Fi and CBRS Is Smart Policy—and Smart Business (June 10, 2025), https://www.ncta.com/news/protecting-wi-fi-and-cbrs-is-smart-policy.

[54] Arielle Roth, Assistant Sec’y of Com. for Commc’ns & Info. & Adm’r, Nat’l Telecommc’ns & Info. Admin., Remarks at the Americas Spectrum Management Conference (Oct. 30, 2025), https://www.ntia.gov/speech/testimony/2025/remarks-assistant-secretary-arielle-roth-americas-spectrum-management-conference (“Let me be clear: OB3 does not alter U.S. policy for the 6 GHz band, which was designated for unlicensed use during the first Trump Administration. That remains our position.”).

[55] LEO Policy Working Group, Low Earth Orbit Satellites: Policies to Promote Spectrum Sharing, Foster Competition, and Close Digital Divides: A Report of the LEO Policy Working Group, New Am. & Int’l Ctr. for L. & Econ. (Oct. 30, 2025), https://www.newamerica.org/insights/leo-satellites.

[56] Id.

[57] Aizaz U. Chaudhry & Halim Yanikomeroglu, Laser Inter-Satellite Links in a Starlink Constellation: A Classification and Analysis, 16 IEEE Vehicular Tech. Mag. 48 (June 2021).

[58] Int’l Telecommc’n Union, ITU Radio Regulatory Framework for Space Networks (2020), https://www.itu.int/en/ITU-R/space/snl/Documents/ITU-Space_reg.pdf.

[59] Single Network Future: Supplemental Coverage from Space; Space Innovation, Report and Order and Further Notice of Proposed Rulemaking, IB Docket No. 22-271, FCC 24-28 (Mar. 15, 2024).

[60] Mohammed S. Elbamby, Cristina Perfecto, Mehdi Bennis & Klaus Doppler, Toward Low-Latency and Ultra-Reliable Virtual Reality, 32 IEEE Network 78 (2018).

[61] LEO Policy Working Group, supra note 55.

[62] See, e.g., Space Bureau Opens Processing Rounds for NGSO FSS Systems, Public Notice, DA 26-552 (Space Bureau June 5, 2026), https://docs.fcc.gov/public/attachments/DA-26-552A1.pdf.

[63] Int’l Telecommc’n Union, Equivalent Power-Flux Density Limits Examination: Part I—Overview (Dec. 2024), https://www.itu.int/en/ITU-R/seminars/wrs/Documents/2024/Space-Workshops/EPFD%20Part%20I%20Overview%20WRS-2024.pdf; see also Modernizing Spectrum Sharing for Satellite Broadband, Report and Order, SB Docket No. 25-157, FCC 26-26, ¶ 18 (rel. May 1, 2026), https://docs.fcc.gov/public/attachments/FCC-26-26A1.pdf.

[64] Modernizing Spectrum Sharing for Satellite Broadband, supra note 63.

[65] Nada Abdel Khalek, Deemah H. Tashman & Walaa Hamouda, Advances in Machine Learning-Driven Cognitive Radio for Wireless Networks: A Survey, 26 IEEE Commc’ns Surveys & Tutorials 1201 (2024).

[66] Westling, Primer: A Staffer’s Guide to Spectrum Auction Reauthorization, supra note 3.

[67] CTIA, The Looming Spectrum Crisis (Mar. 27, 2025), https://www.ctia.org/news/the-looming-spectrum-crisis.

[68] One Big Beautiful Bill Act, Pub. L. No. 119-21, 139 Stat. 72 (2025).

[69] Id. § 40002(b)(1).

[70] Commercial Spectrum Enhancement Act of 2004, Pub. L. No. 108-494, tit. II, 118 Stat. 3986, 3991 (2004) (codified at 47 U.S.C. §§ 923(g)–(i), 928).

[71] Arielle Roth, Assistant Sec’y of Com. for Commc’ns & Info., Remarks at the 2026 CTIA Summit (May 6, 2026), https://www.ntia.gov/speech/testimony/2026/remarks-assistant-secretary-arielle-roth-2026-ctia-summit.

[72] Id.

[73] 47 U.S.C. § 928(g)(2)(D)(ii).

[74] Press Release, Nat’l Telecommc’ns & Info. Admin., FCC, NTIA Establish Spectrum Coordination Initiative (Feb. 15, 2022), https://www.ntia.gov/press-release/2022/fcc-ntia-establish-spectrum-coordination-initiative; Memorandum of Understanding Between the Federal Communications Commission and the National Telecommunications and Information Administration (Aug. 1, 2022), https://www.ntia.gov/other-publication/2022/memorandum-understanding-between-fcc-and-ntia.

[75] Nat’l Telecommc’ns & Info. Admin., National Spectrum Strategy (Nov. 2023), https://www.ntia.gov/sites/default/files/publications/national_spectrum_strategy_final.pdf; Nat’l Telecommc’ns & Info. Admin., National Spectrum Strategy Implementation Plan (Mar. 12, 2024), https://www.ntia.gov/sites/default/files/publications/national-spectrum-strategy-implementation-plan.pdf.

[76] Spectrum & Receiver Performance Working Grp., Fed. Commc’ns Comm’n Technological Advisory Council, A Quick Introduction to Risk-Informed Interference Assessment (ver. 1.00, Apr. 1, 2015), https://transition.fcc.gov/bureaus/oet/tac/tacdocs/meeting4115/Intro-to-RIA-v100.pdf.

[77] Id. at 1.

[78] Id. at 3.

[79] AT&T Servs., Inc. v. FCC, No. 20-1190, at 15 (D.C. Cir. Dec. 28, 2021).

[80] Nat’l Telecommc’ns & Info. Admin., National Spectrum Strategy, supra note 75, at 16.

[81] Serhat Tadik et al., Digital Spectrum Twins for Enhanced Spectrum Sharing and Other Radio Applications, IEEE J. Radio Frequency Identification (2023), https://ieeexplore.ieee.org/document/10293151.

[82] Spectrum & Receiver Performance Working Grp., Fed. Commc’ns Comm’n Technological Advisory Council, supra note 76.

[83] Int’l Telecommc’n Union, World Radiocommunication Conferences (WRC), https://www.itu.int/en/ITU-R/conferences/wrc/Pages/default.aspx (last visited June 9, 2026).

[84] Evan Grey, The 200,000-Satellite Filing: When Commercial Loopholes Become State Weapons, SatNews (Jan. 23, 2026), https://satnews.com/2026/01/23/the-200000-satellite-filing-when-commercial-loopholes-become-state-weapons.

[85] Jeffrey Westling, Primer: World Radio Conference 2023, Am. Action Forum (Aug. 30, 2023), https://www.americanactionforum.org/insight/primer-world-radio-conference-2023.

[86] Secure and Trusted Communications Networks Act of 2019, Pub. L. No. 116-124, 134 Stat. 158 (2020) (codified at 47 U.S.C. §§ 1601–1609) (authorizing the Federal Communications Commission to prohibit the use of federal subsidies for covered equipment and to reimburse providers for removing Huawei and ZTE equipment).