← Back to feed

Anthropic published a paper on the US-China AI competition on May 14, laying out two scenarios for 2028 and a concrete set of interventions the US and its allies would need to execute to lock in a durable frontier lead. The paper represents Anthropic’s first formal public position on the geopolitics of AI capability.

The Core Argument

Anthropic’s position rests on a specific claim about the compute bottleneck: advanced chips are not merely one input into AI development — they are the gatekeeper for training, deployment, revenue, experimentation, and future model improvement. Everything downstream of capability depends on compute access.

The paper’s sharpest hardware claim: Huawei’s domestic semiconductor capacity will produce approximately 4% of NVIDIA’s aggregate compute in 2026, declining to roughly 2% in 2027. That gap, if maintained through export controls and allied semiconductor policy, is Anthropic’s basis for the 12-to-24-month lead scenario.

The Distillation Problem

Anthropic frames distillation — the practice of using a frontier model’s outputs to train a smaller, cheaper model — as systematic industrial espionage rather than legitimate research. The argument: Chinese labs can copy frontier capabilities without paying the full training cost, effectively extracting the value of US investment without the associated compute constraint.

This framing has direct policy implications. Restricting chip exports addresses the training compute gap. It does not address the distillation channel, which operates through API access and output observation. Anthropic’s paper argues both vectors need to be closed simultaneously.

”Country of Geniuses in a Data Center”

The paper introduces a framing that appears designed for policy audiences: future frontier models may function as a “country of geniuses in a data center” — a single model cluster acting as a massive expert workforce capable of autonomous contributions to cyber operations, scientific research, engineering, and military applications.

That framing is strategic. It shifts the policy question from “how do we regulate AI products” to “how do we prevent a peer adversary from deploying what amounts to a large-scale autonomous expert workforce.” The latter maps more cleanly onto existing national security frameworks.

What a Chinese Lead Would Enable

The paper does not treat a Chinese AI lead as a commercial inconvenience. The listed consequences include automated domestic repression at scale, stronger offensive cyber operations, faster military AI deployment, and widespread cheap AI infrastructure exports that could extend authoritarian influence through dependency.

Whether those outcomes would materialize is debated. What the paper establishes is Anthropic’s institutional view: the stakes of losing the frontier race are not primarily economic.

The Two Scenarios

The paper structures its argument around two 2028 outcomes. In the favorable scenario, coordinated US and allied action on compute controls, distillation restrictions, and export enforcement creates a self-sustaining frontier lead that Chinese labs cannot close in the near term. In the adverse scenario, piecemeal policy, enforcement gaps, and continued distillation access allow Chinese labs to stay close enough to the frontier that the lead narrows to single-digit months — effectively competitive parity on capability for any state-level actor.

Anthropic does not claim the favorable scenario is inevitable. It claims it is achievable if the policy window — which the paper implies narrows significantly past 2027 as Huawei’s domestic compute scales — is used.

The Context

The paper lands as Trump met Xi in Beijing on May 14-15, with both governments agreeing to launch dedicated AI safety talks. Whether those talks and Anthropic’s policy recommendations are compatible positions has not been addressed by either party.

Stanford’s 2026 AI Index, published in April, found the performance gap between leading US and Chinese models at 2.7 percentage points — its narrowest recorded level. Anthropic’s paper accepts that near-parity as the premise and argues the relevant question is not today’s capability gap but the trajectory over the next 24 months.