China is opening a door it slammed a year ago, but only part way. Reporting that broke on the evening of July 8 and gained its crucial details on July 9 says Beijing will permit Alibaba, ByteDance and DeepSeek to buy Nvidia's H200 accelerators — while capping total approvals below 200,000 chips, less than half what the companies asked for.

The conditions

The chips may be used to train AI models. Inference — the far larger, recurring workload of actually serving those models — must run on domestic processors, with authorities directing firms toward Huawei silicon. Each company must justify the quantities it requests. The effect is a rationing scheme, not an opening. Alibaba and ByteDance are among roughly ten Chinese firms the US Commerce Department licensed in May to buy up to 75,000 H200s apiece — a single company's allowance that would consume much of Beijing's reported ceiling. DeepSeek was not among those licensees; its access runs through a separate, China-side approval.

Why the freeze existed

The blockage was China's, not Washington's. The Trump administration banned exports of Nvidia's China-compliant H20 in April 2025, then reversed course that July. Days later, on July 31, 2025, China's Cyberspace Administration summoned Nvidia over claims the H20 could be tracked or remotely disabled. Nvidia denied it flatly: "NVIDIA does not have 'backdoors' in our chips that would give anyone a remote way to access or control them." Beijing nonetheless discouraged domestic purchases and pushed firms toward local suppliers, part of a broader self-sufficiency drive.

What is at stake for Nvidia

Nvidia once held roughly 95% of China's high-end AI chip market. That business is now close to gone: China's share of Nvidia's revenue, as high as 13% in fiscal 2025, has fallen to the low single digits, and the company booked essentially no China data-center revenue in its most recent reported quarter. Shares rose about 3.65% to $204.12 on July 8, the session the report broke, valuing the company near $4.94 trillion, before slipping about 0.7% the next day. Neither Nvidia nor China's commerce ministry has commented on the reported caps, and the final approval numbers have not been officially announced.

The design of the concession

Read closely, the policy concedes exactly one thing: Chinese labs cannot yet train frontier models on domestic hardware at competitive speed. It concedes nothing on inference, where Huawei's chips are good enough and the volume — and the long-run market — actually lives. The asymmetry is deliberate. Training is a burst of spending that produces a model; inference is a permanent bill that produces a dependency, and Beijing has decided which of the two it is willing to pay an American company to supply.

What happens next

The final approval figures have not been published, and firms are still filing the justifications Beijing demands. That May licensing round in Washington also covered Tencent, JD.com, Lenovo and Foxconn — a reminder that two governments now hold a veto over every chip that moves, and that clearing one says nothing about clearing the other.