Anthropic released a major policy paper this week, and the timing was not accidental. It dropped on the same day President Trump landed in Beijing to meet Xi Jinping, with Nvidia's CEO Jensen Huang on the plane and fresh reports of the US clearing H200 chip sales to ten major Chinese tech firms.
The paper argues that the US has a narrow window, possibly closing by 2027, to lock in a 12 to 24 month lead in AI capability over China. Miss it, and the race becomes a dead heat, which Anthropic frames as its own kind of danger.
Here is what most coverage misses: the hardware gap is real but it is not the whole story.
China's domestic chip production is genuinely weak. Huawei's best AI chip will produce roughly 4 percent of what Nvidia produces this year in raw processing power. American hyperscalers like Google and Amazon are also building their own chips, meaning Nvidia is only one part of a much larger US lead. Independent analysts have estimated that if the US fully enforced its restrictions, it would hold a computing advantage 21 times larger than China for 2026-produced hardware.
And yet Chinese labs have stayed competitive. The reason is simple: they have been getting US chips anyway, through smuggling networks in Southeast Asia and by accessing export-controlled data centers remotely. Federal prosecutors have already charged one case involving $2.5 billion worth of servers diverted to China. Alibaba and ByteDance reportedly train models on restricted US chips in data centers in Malaysia and Singapore, a route current rules do not cover because US export law restricts the sale of chips, not remote access to them.
On top of that, Chinese labs have been copying the intelligence out of US models directly. Anthropic's own report earlier this year documented that DeepSeek, Moonshot AI, and MiniMax created over 24,000 fraudulent accounts and generated more than 16 million interactions with Claude to harvest its reasoning for their own training. The White House formally called this industrial espionage in April. OpenAI, Google, and Anthropic have all flagged the same pattern. DeepSeek's own paper for its V4 model openly describes using US model outputs as training teachers.
All of that matters because the gap on actual model intelligence is smaller than the chip gap would predict. DeepSeek's V4, released in late April, conceded in its own technical paper that it lags behind the current US frontier by roughly three to six months, not years.
Now add the adoption picture. Chinese open-source models went from under 2 percent of global AI usage to roughly 30 percent within about a year. In February 2026, Chinese models briefly overtook American models in weekly API call volume on the world's largest model aggregation platform, by nearly double. Countries across Southeast Asia, Latin America, and Africa are choosing Chinese models because they can download them and run them locally, without depending on a US company that might cut access due to political tensions. Malaysia announced its national AI infrastructure would run on DeepSeek. Singapore's government AI program chose Alibaba's Qwen model over Meta's Llama.
This is the dynamic Anthropic is trying to flag. The US builds the most powerful AI. China makes a version that is a few months behind, gives it away for free, and wins the global market on price. Countries building on that foundation then become part of an AI infrastructure shaped by Chinese terms, not American ones.
Anthropically's new model, Claude Mythos Preview, provides the clearest illustration of why the capability gap still matters. Released in April to a small group of partners under a program called Project Glasswing, it can autonomously find and exploit security vulnerabilities in every major operating system and web browser, including ones that had gone undetected for 27 years. Firefox fixed more security bugs in one month using Mythos than it had in all of 2025. One Chinese cybersecurity analyst reportedly wrote that China feels like it is sharpening swords while the other side has just deployed an automated weapon system. That model is not publicly available. If the gap closes and China builds something equivalent first, they would have it too, and they would not gate access the same way.
The Trump-Xi summit this week was the perfect backdrop for all of this. Chip export controls were reportedly not formally on the bilateral agenda, though Washington quietly cleared H200 sales to around ten Chinese firms the same day. That is exactly the tension Anthropic is warning about: commercial pressure to sell chips, happening at the same moment policymakers are being told this is the last window to lock in an advantage.
The three things Anthropic is asking for are tighter enforcement on chip smuggling and overseas data center access, legal clarity that makes model distillation attacks a crime, and an active push to get American AI adopted as the global infrastructure standard before Chinese models cement themselves in emerging markets.
For businesses outside the US, the practical implication is this: whatever AI tools your company or suppliers are building on top of, the governance, access terms, and long-term reliability of those tools will be shaped by whoever wins this competition. That is not an abstract geopolitical concern. It is a vendor risk question.