Anthropic Says the US China AI Race Has a 2028 Deadline
Anthropic is sounding the alarm. In a sweeping policy paper published on May 14, 2026, the AI safety company warned that the US China AI race has entered a critical window, and Washington has until around 2028 to lock in a lasting advantage, or risk losing the moment entirely.
The paper’s central argument is that compute, specifically access to advanced AI chips, is the single most important battleground. Thanks to US export controls and the dominance of companies like Nvidia, TSMC, and ASML, democratic nations currently hold a significant hardware lead. An analysis cited in the paper projects that Huawei will reach only four percent of Nvidia’s aggregate compute capacity by 2026, falling to just two percent by 2027.
But Anthropic is not declaring victory. Chinese AI labs remain close to the frontier through two key methods: chip smuggling and access to US compute through foreign data centers, and systematic distillation attacks where Chinese labs use thousands of fake accounts to scrape and replicate the outputs of American AI models. Anthropic previously accused DeepSeek, Moonshot, and Minimax of generating more than 16 million Claude interactions through roughly 24,000 fraudulent accounts. OpenAI, Google, and the Frontier Model Forum have also condemned the practice.
The paper maps out two futures. In the optimistic scenario, the US closes export loopholes, cracks down on distillation attacks through legislation, and promotes American AI infrastructure globally, securing a 12 to 24-month lead in model intelligence and cementing democratic AI norms worldwide. In the darker scenario, those loopholes remain open, Chinese labs reach near-parity, and Huawei data centers expand globally with cheaper but capable models, giving Beijing growing influence over how AI develops and who it serves.
Safety is another casualty Anthropic flags in that second scenario. Only 3 of 13 top Chinese labs have published safety evaluations so far, and DeepSeek’s R1-0528 fulfilled 94 percent of malicious requests under common jailbreaking techniques, according to the Center for AI Standards and Innovation.
The paper was released right in the middle of President Trump’s visit to China and while Congress is actively debating AI export controls, chip licenses, and cloud access rules. The timing makes Anthropic’s message hard to miss: the US China AI race is not a slow burn, it is a timed window, and that window is closing.
