Dario Amodei's "Pacing the Frontier": Some Anti-Trust Law Limitations

Dario Amodei, CEO of Anthropic, has called for the AI industry to “pace the frontier”,  essentially, to moderate the speed at which increasingly capable AI systems are developed so that safety measures can keep pace with technological progress. Other leading figures in the AI industry have expressed support for the idea. Amongst other things, Amodei’s proposal contemplates third-party evaluators, common safety standards and coordination among frontier AI companies. In describing the proposed framework, Amodei states that:

“Frontier AI companies within democratic countries coordinate to establish common safety standards as well as limits on the rate of unchecked AI progress.”

The precise nature and scope of any future agreement or coordination between AI companies remain uncertain. However, if competing AI companies were to agree to limit or slow the development, deployment, or release of increasingly capable AI systems, such arrangements could raise significant competition law concerns.

We discuss some of the primary legal considerations below.

1. Output Restriction

Output restriction is a well-established competition-law concern. Agreements between competitors to limit or control production, supply, or output restrict competition by eliminating the competitive pressure that would otherwise force firms to bring their best offerings to market. In Nigeria, agreements to limit or control the production of goods or services, or to restrict technical development and investment, are expressly prohibited under Section 59 of the Federal Competition and Consumer Protection Act (FCCPA).

In the AI industry, the output of an AI company may not simply be the number of subscriptions sold or API calls processed. It may include the continuous development and deployment of increasingly capable models, compute allocation, the discovery of new technological capabilities, and the speed at which those capabilities are commercialised.

Accordingly, a horizontal agreement between competing AI companies to limit training scale or delay the release of frontier models could amount to an unlawful agreement to restrict output or, more broadly, to suppress technological competition.

2. Competition is Not Limited to Price

Competition law is not concerned only with price-fixing cartels. In technology-driven markets, consumer harm does not always present as higher prices and frequently manifests as diminished quality, delayed features, and artificially suppressed capabilities.

Competition occurs just as fiercely through non-price factors such as quality, innovation, research, and technological development. This is particularly critical in the AI sector, where the ability to develop and deploy higher-performing models is itself a primary dimension of competition.

The antitrust risk of joint technological restraint is illustrated by United States v. Automobile Manufacturers Association, where the U.S. government challenged an alleged agreement among major automobile manufacturers to delay the development and adoption of pollution-control technology. The case underscores a vital principle that horizontal competition concerns may arise whenever competitors coordinate to slow technological progress, even in the absence of price coordination.

The same principle may apply to frontier AI. If competing AI developers agree to collectively delay training, benchmark achievement, or public release of advanced models, the antitrust exposure may extend beyond conventional output restrictions. Such arrangements could potentially operate as innovation restraints that have the purpose or effect of restricting competition.

3. Can AI Safety Justify Coordination?

The primary commercial justification for coordination is that unchecked competition in frontier AI creates systemic risks that individual market players may not be able to manage alone.

A developer may argue that it cannot unilaterally pause or slow development without surrendering its market position to accelerating rivals. Collective action is therefore framed as a necessary measure to prevent a dangerous “race to the bottom.” However, the legitimacy of the underlying objective does not necessarily resolve the competition-law question.

Good faith intentions do not automatically remove horizontal restraints from the scope of competition law. The relevant question therefore is whether the proposed coordination falls within a recognised exemption or can otherwise be justified under the applicable legal framework.

Under Section 60 of the FCCPA, an agreement that would otherwise fall within Section 59 may be authorised where it contributes to the improvement of production or distribution or promotes technical or economic progress, while allowing consumers a fair share of the resulting benefit, imposing only restrictions that are indispensable to achieving those objectives, and not eliminating competition in respect of a substantial part of the goods or services concerned.

This creates an important distinction for AI safety initiatives. First, a narrowly defined arrangement directed at establishing safety standards, independent evaluation mechanisms or technical safeguards may raise different competition law questions from an agreement between competitors to delay the development or commercial release of more capable models. The nature of the restraint, its scope, duration, necessity and effect on competition would therefore matter.

4. The Role of Government

The role of government becomes particularly important where achieving the stated public-interest objective requires coordination between competitors. Amodei's proposal itself contemplates coordination among frontier AI companies and, ultimately, coordination involving governments. This raises a further question as to whether government involvement can provide a legal basis for cooperation that might otherwise raise competition concerns.

Nonetheless, Government participation should not generally act as an automatic immunity from competition law. In our view, the legal position would depend on the nature of the government intervention, the statutory authority under which it occurs, the scope of the coordination permitted, and whether competitors are being authorised to exchange competitively sensitive information or to restrict competition beyond what is necessary to achieve the regulatory objective.

In Nigeria, the FCCPA provides a mechanism through which restrictive agreements may be authorised by the Federal Competition and Consumer Protection Commission where the statutory conditions are satisfied. This may be particularly relevant where AI safety requires coordinated industry standards but the proposed coordination could otherwise affect competition in technological development.

5. The Difficulty of Drawing the Line

The most difficult question may therefore not be whether AI safety is a legitimate objective. It plainly is. The more difficult question is where to draw the line between legitimate safety cooperation and unlawful coordination between competitors.

For example, competitors may reasonably collaborate on technical safety standards, incident reporting, independent evaluation methodologies or other measures that improve the safety of AI systems without materially restricting competition.

The position becomes more difficult where competitors agree on the pace at which models may be trained, the capabilities they may develop, the benchmarks they may achieve, or when they may release new models. The distinction is important because a safety standard can protect competition and consumers, while an agreement to suppress technological development can potentially remove the very competitive process that competition law seeks to protect. The challenge for regulators will therefore be to distinguish genuine safety cooperation from coordination that uses safety as a justification for suppressing competition.

Key Takeaways

The debate over “pacing the frontier” presents a novel challenge for competition law. AI safety may justify cooperation between industry participants, but cooperation between competitors can also create significant competition law risks where it restricts output, technological development, innovation or other dimensions of competition. For AI developers, the issue is therefore not simply whether they should cooperate on safety but how that cooperation is structured.

This publication is provided Balogun Harold for general informational purposes only and does not constitute legal advice. Specific circumstances may require tailored legal analysis. For consultation requests, please reach out to your usual Balogun Harold contact or via support@balogunharold.com