Anthropic published a policy paper this week titled "2028: Two Scenarios for Global AI Leadership," arguing that the United States has roughly a two-year window to harden its chip-export regime before transformative AI systems begin arriving in earnest. The paper, released on May 14, frames 2026 as a "breakaway" year — one in which decisions made in Washington and allied capitals will compound into a durable lead or collapse into parity with China.
Two scenarios for 2028
The paper sketches two outcomes. In the first, the US successfully defends its compute advantage: policymakers tighten export controls, disrupt distillation attacks in which Chinese labs extract capabilities from American models, and accelerate AI adoption across democracies. American systems remain 12 to 24 months ahead of Chinese rivals, and democracies set the global norms for how AI is built and deployed.
In the second scenario, Washington does not act. Smuggling routes for advanced chips remain open, offshore data centers give Chinese labs back-door access to restricted compute, and Beijing's models close the capability gap. In that world, Anthropic argues, AI norms are shaped by authoritarian regimes, and the best frontier systems enable automated repression at scale.
Compute as the strategic chokepoint
Anthropic places advanced semiconductors at the center of the analysis. Powerful AI systems still depend on chips designed by American and allied firms — and on the equipment used to manufacture them. The paper credits bipartisan US export controls with creating today's lead, but warns that the controls are leaking: chip smuggling into China, offshore data center access that bypasses geographic restrictions, and distillation attacks that let Chinese labs piggyback on American training runs.
The policy prescriptions track closely with positions Anthropic has pushed in Washington throughout 2026. The company recommends closing smuggling and foreign data-center loopholes, enforcing semiconductor equipment controls, restricting unauthorized model access, and actively promoting global adoption of American AI infrastructure so that allied governments default to US-built systems.
Industry and policy implications
The paper lands at a politically charged moment. Anthropic itself is locked in a court fight over the Pentagon's blacklisting of the company — even as the NSA is reportedly using Anthropic's Mythos model despite that ban — and Washington has just cleared Nvidia H200 sales to roughly 10 Chinese firms, with Beijing reportedly pressuring those buyers to hold off on deliveries. Anthropic's framing puts those skirmishes inside a larger narrative: the chip-control regime is the load-bearing wall of US AI leadership, and patching its cracks is more urgent than any single licensing fight.
For enterprise buyers and allied governments, the message is implicit but pointed. If Anthropic's first scenario holds, American models will remain the default for regulated industries and sovereign deployments. If the second prevails, cheaper Chinese systems become the global standard for cost-sensitive markets — and the norms around safety, surveillance, and content moderation get written elsewhere.
