Anthropic maps US-China AI race to 2028: compute and chip access as deciding factors

Anthropic released a forecast mapping how the US-China AI race could play out through 2028, framing frontier AI leadership not as a pure model-quality contest but as a competition for compute, chip supply, and infrastructure scale. The company argues that the US holds an edge today, but China remains close to the frontier — and the gap could widen or collapse depending on export policy and datacenter buildout over the next three years.

The report sketches two scenarios. In the first, US export controls on advanced chips hold and American companies scale training infrastructure fast enough to maintain a frontier lead. In the second, China catches up to frontier models despite restrictions, and global AI standards begin to form outside Western influence. Anthropic emphasizes that access to advanced silicon and the ability to build massive training clusters will determine which nation leads by 2028. Throughout the document, Anthropic treats frontier AI as dual-use technology — relevant not just for chatbots and productivity tools but for cybersecurity, military systems, and accelerated scientific research. That framing reflects a broader shift in how leading AI labs talk about their work: a year ago, most public commentary centered on consumer applications and enterprise automation; now the conversation includes national security, critical infrastructure, and the geopolitical stakes of who controls the most capable models. For open-source practitioners, the forecast underscores a tension already visible in model releases. Western labs have tightened weights distribution and added more restrictive licenses over the past eighteen months, citing dual-use risk, while Chinese research groups continue to release competitive open-weight models — Qwen, DeepSeek, and others — that often match or exceed Western counterparts on standard benchmarks. If Anthropic's second scenario plays out and China reaches frontier parity, the center of gravity for unrestricted model development could shift east.