Beyond the Frontier: Deliberations on AI Safety, Infrastructure and Pacing

Dario Amodei’s We Must Pace the Frontier argues that slowing the development of the most advanced AI models could create time to address unresolved safety and control problems. This approach assumes that risk can be contained by slowing a relatively small number of frontier training runs.[1]

AI capability is not confined to the largest models or to the training stage. Capabilities can, as discussed further below, augment in various ways. Countries too, do not follow a unified trajectory. Governments have different economic, technological and strategic priorities, while companies make their own decisions based on market conditions, policy and access to infrastructure.

Safety is not ensured by pacing frontier development or restricting the physical infrastructure used to train frontier models. It develops alongside development with policy commensurate with AI.

The proposed U.S. Artificial Intelligence Data Centre Moratorium Bill demonstrates the risks of treating infrastructure restrictions as a substitute for AI safety. Restricting data-centre construction may constrain the physical capacity available for AI development, but it does not necessarily control what existing or smaller models can do. It can also affect ordinary computing infrastructure, reinforce the position of companies that already have large data-centre fleets, and create strategic difficulties if other countries continue expanding their own computing capacity.

The question is therefore not simply how fast AI should be developed. It is where AI risk emerges, whether that risk can be controlled at the frontier and across other types of models, and where regulation can intervene most effectively once AI systems are released.

The Limits of the Pacing Framework

Amodei’s proposal assumes that slowing frontier training would create valuable time for safety research. Safety cannot necessarily be separated from capability development in this way. Understanding whether an advanced system is controllable requires researchers to study the capabilities the system actually develops, how it behaves in realistic environments, and how failures emerge during use. Safety research therefore needs to develop alongside AI capabilities rather than existing entirely outside the development process.

This creates a feedback loop. Development produces new capabilities, which create new safety challenges. Research explores vulnerabilities and controls and influences subsequent system development and deployment. Safety is a strategic priority in interpreting and managing increasingly capable systems.

Restricting new frontier training runs may result in an uneven playing field for newer companies to compete with firms that already possess substantial computing resources, infrastructure and advanced models. Companies that have already reached the frontier would retain their existing advantages, while prospective competitors would face greater barriers to entry. A risk mitigation could inhibit fair competition.

Innovation is shaped by constraints throughout the technology sector. Companies routinely delay or accelerate development and deployment according to chip availability, financing, data access, commercial readiness, regulatory uncertainty and expected returns. There is therefore a distinction between deliberately slowing frontier AI as a safety intervention and recognising that technological development is already constrained by organisational realities.

Amodei’s framework also concentrates heavily on frontier models, which creates a potential blind spot. Dangerous capabilities do not necessarily remain confined to the largest models or emerge only after a particular level of training compute has been reached. Capabilities can spread beyond the frontier and become more sophisticated through the way models are used after training.

A question arises question of institutional power. If leading companies genuinely believe that their unreleased systems pose sufficiently serious risks to justify a slowdown, those companies can already choose to slow their own development. Voluntary corporate restraint is not a substitute for public regulation, but it raises the question of whether companies seeking broader regulatory intervention should also demonstrate how they apply their own safety principles to development decisions.

Infrastructure Policy Is Not Tantamount to Ensuring AI Safety

The proposed U.S. Artificial Intelligence Data Centre Moratorium Bill attempts to translate AI safety concerns into infrastructure policy. The proposal would impose a freeze on certain new data-centre development until specified federal standards addressing the impacts of AI data centres are achieved.[2]

Data centres are general-purpose infrastructure. They do not exist exclusively to develop potentially dangerous AI systems; they also support medical research, climate modelling, enterprise systems, scientific computing and other activities with little connection to frontier AI safety.

There is a mismatch between the object being regulated and the risk being addressed. Restricting physical infrastructure may reduce the availability of computing capacity, but it does not necessarily address what an AI system does once it exists.

The framework described in the proposed bill seeks to prevent increases in household water costs and negative effects on local electricity grids. Data centres are significant consumers of electricity and water. Requiring genuinely zero impact could be difficult to counterbalance with the physical realities of large-scale computing. Unrealistic prerequisites risk making a temporary infrastructure freeze  an indefinite constraint on domestic technology development. Environmental and energy impacts can be addressed through infrastructure regulation, while AI safety requires controls directed at the systems themselves and the consequences of their use.

The Risk of Entrenching Established Market Players

Infrastructure restrictions can also have an uneven effect within the domestic market. Large technology companies already have extensive operational data-centre capacity. A restriction on new construction does not remove that advantage. Instead, it can make it harder for smaller infrastructure companies and emerging competitors to build comparable capacity.

A paradoxical outcome may therefore follow: a policy introduced to reduce the concentration of AI power could strengthen the position of companies that already control the largest computing resources.

Any infrastructure or compute-related intervention should consequently be assessed not only by whether it reduces computing capacity, but also by who retains access to that capacity and how the intervention affects future competition.

Countries Do Not Follow the Same AI Trajectory

Democracies cannot merely align on AI development. They may share broad values while pursuing different economic, technological and strategic priorities.

International initiatives can coordinate standards, supply chains and policy approaches, but they cannot necessarily override sovereign decision-making. Alliances engage but seldom entrap states.[3] Since states are sovereign actors in an anarchic state system, the “Alliance dilemma”- the tension between abandonment (being left behind by counterparts) and entrapment (being compelled down another country’s path) necessitates that states constantly renegotiate their alignments, preventing an entirely unified path.  This does not necessarily contradict international cooperation. Instead, it demonstrates that international coordination operates alongside national priorities.

Beyond Frontier Models: Other Pathways to Greater AI Capability

Amodei limits his pacing proposal to frontier AI models and their training. However, AI capabilities can evolve beyond both the frontier models and their training.

One important issue is the diffusion of capabilities from frontier models into smaller models. Amodei identifies this concern in his discussion of China: if capabilities developed by frontier systems spread, other countries may be able to acquire or reproduce advanced capabilities without leading the frontier themselves. The effects of frontier development can extend well beyond the companies and countries that initially develop the most capable systems.

Such diffusion creates another safety concern. Once capabilities become available beyond frontier systems, slowing further frontier development does not necessarily prevent those capabilities from spreading across the wider AI ecosystem. Smaller or more accessible models may therefore possess capabilities that originated in more advanced systems, even if they were not themselves trained at the frontier.

Another pathway is inference-time compute. A model can be given additional computational resources when answering a particular prompt, allowing it to spend longer working through a problem, checking its reasoning and correcting errors before responding. Capability can therefore increase through how a model operates rather than through another training run.[4]

A further pathway is autonomous interaction with external tools. An AI system can interact with software and other systems in its environment, observe the consequences of its actions, identify failures and retry with a modified approach until it succeeds. The resulting increase in capability does not necessarily come from changing the model itself, but from the model operating within an environment.

Open-weight models create another challenge. An open-weight model is one whose trained model weights—the parameters that determine how the model processes information and generates outputs—are made publicly available. Once the weights are released, users can download and run the model independently, modify it, and potentially remove safety restrictions. Centralised corporate controls, such as restricting access to the model or shutting it down, become much less effective once the model is distributed.

These pathways make it difficult to use training compute as the primary measure of AI risk. Regulation can instead focus more directly on what systems are capable of doing and the risks created by those capabilities.

A model may pass an evaluation before deployment but behave differently once it is placed in a more complex environment or used by people who adapt it for purposes that were not anticipated during testing. Real-time operational monitoring is essential- existing policies globally also consider this.

Strengthening Liability

Monitoring still leaves the question of responsibility for technological harms and breaches. Since entities are best placed to know the operational intricacies of their AI systems, and are increasingly being held liable for how ther platforms function, liability for continuing harm needs to be considered.

Harm from a breach may continue long thereafter. Personal information can remain exposed, identities can be misused and commercial consequences can persist after the initial incident has ended. Allocating restitution for continuing harm would focus on the consequences experienced by affected individuals rather than treating the breach as the only tangible manifestation reminiscent of the failures that caused it. This could require organisations to secure their deployments more resiliently.

Pacing should not be interpreted as a proposed AI governance suggestion, but rather much more stringently as an inherent expectation which entities are meant to be fulfilling without needing policy or regulation to mandate it.[5] AI develops in many ways beyond the training stage. The pacing proposal is a strategic attempt to push for regulations favouring larger entities, and to further entrench structural competitive imbalances through policy.

[1] Dario Amodei. We Must Pace the Frontier. September 2026. https://darioamodei.com/post/we-must-pace-the-frontier

[2] S.4214. Artificial Intelligence Data Centre Moratorium Bill, 2026. https://www.congress.gov/119/bills/s4214/BILLS-119s4214is.pdf

[3] Kim, T. (2011). Why Alliances Entangle but Seldom Entrap States. Security Studies, 20(3), 350–377. https://doi.org/10.1080/09636412.2011.599201

[4] Parashar, S. et al (2025). Inference-Time Computations for LLM Reasoning and Planning: A Benchmark and Insights. https://arxiv.org/pdf/2502.12521

[5] https://www.linkedin.com/posts/rassermartijn_david-sackss-response-to-dario-amodeis-share-7504888242542931968-0hFT/