The US leads China in frontier AI development today. Whether that's still true in 2028 depends on a handful of chip factories, some export regulations, and decisions being made right now - and Anthropic just published a paper making exactly that case.
Released May 14, the paper maps two possible futures for global AI leadership by 2028. It reads less like a traditional AI safety document and more like a geopolitical risk briefing, which is notable for a company that built its reputation on technical safety research.
The Compute Argument
The core claim is straightforward: America's current lead over China in frontier AI is almost entirely a compute story. NVIDIA designs the GPUs that train the world's most powerful models. TSMC in Taiwan manufactures them. ASML in the Netherlands builds the extreme ultraviolet lithography machines that TSMC depends on to produce advanced chips. China currently can't replicate any of these three links in the chain.
Export controls - US government restrictions on selling advanced chips and chipmaking equipment to Chinese companies - have maintained this gap. Without those controls, Anthropic argues, Chinese AI labs would be training frontier models on the same hardware American labs use today. The paper treats this as a fragile advantage, not a permanent one.
The two scenarios Anthropic lays out aren't presented as definitive predictions. They're a framework for thinking about what happens depending on whether export controls hold or erode over the next two years. The implication is that the window to act is still open, but not indefinitely.
Why This Paper Reads Differently from Anthropic's Usual Research
Publishing a paper that argues for specific US trade and export control policies is a departure from Anthropic's usual output. Most of the company's public research focuses on model evaluation, interpretability, and constitutional AI - technical work about making models safer and more predictable. This paper is geopolitical advocacy dressed in research framing.
That doesn't make it wrong. But it does mean it deserves to be read as a policy document, not a neutral scientific analysis. Anthropic has financial and strategic interests in the outcome it's describing.
For people who use AI tools daily, none of this is abstract. The models behind Claude, ChatGPT, Gemini, and every other frontier product were trained on hardware that flows through a specific supply chain. If that chain's competitive advantage narrows or disappears, the product landscape shifts - pricing, capability gaps, which companies can afford to train what.
The most practically useful thing about Anthropic's paper isn't the specific scenarios. It's the reminder that the biggest near-term risks to the AI tools people rely on aren't about hallucinations or philosophical misalignment. They're about semiconductor factories in Taiwan and export enforcement in Washington.