Washington Lets Nvidia Sell H200 Chips to China, Taking a 25% Cut
The December decision reopened the China AI chip market on new terms, but Blackwell stays banned and by August Nvidia was guiding as if China did not exist.
Key takeaways
- Trump announced on Dec. 8, 2025 that H200 sales to approved Chinese customers could proceed with a 25% US government cut, up from 15% floated earlier.
- Blackwell and Rubin chips stay barred; export licences reportedly followed in February 2026.
- By Nvidia's August 2026 quarter, China Hopper shipments were under 1% of data center revenue and guidance assumed none.
On December 8, 2025, President Trump said the US would allow Nvidia to sell its H200 AI accelerators to approved customers in China, with the government taking 25% of the proceeds. It reversed a posture that had barred the company from selling anything near its leading edge into the world's second-largest AI market, and it did so on unusual commercial terms.
What was decided
The H200 is a Hopper-generation GPU, more capable than the cut-down H20 that Nvidia had been permitted to sell, but a generation behind the Blackwell parts that now dominate US data centers. Trump announced the policy on social media, saying Washington would receive a 25% share of sales, up from a 15% arrangement reported earlier. Coverage differs on mechanics: one account says the fee is collected when chips enter the US for a security review before being re-exported, which lawyers have argued resembles a tariff more than a revenue share. That reading is unverified and should be treated cautiously.
Blackwell and the upcoming Rubin platform remained excluded. Reporting later said export licences were granted in February 2026 after a December understanding between the US and Chinese presidents, with conditions such as shipment caps and third-party testing. Those details come from secondary summaries, and the licence terms themselves are not public in full.
Beijing's role
Washington's approval was only half of the equation. Chinese regulators have been pushing domestic buyers toward homegrown accelerators, and Reuters reported that shipment timing would depend heavily on Beijing's sign-off. Nvidia's chief executive said at one point that no new orders had been placed while China decided. Early reporting showed the company moving quickly to prepare supply, but orders and deliveries are separate matters, and the pace has been uneven.
That uncertainty explains the subsequent financial picture. In Nvidia's quarter ended July 26, 2026, Hopper shipments to China were below 1% of data center revenue, and the company's guidance for the following quarter explicitly assumed no China data center compute sales. A market that once represented a meaningful share of sales has effectively dropped out of the plan.
Why it matters
The policy tests two competing theories of export control. One holds that selling slightly older chips keeps Chinese developers dependent on the American software stack and funds US research. The other holds that any high-end compute strengthens a strategic rival's AI capabilities. The 25% arrangement adds a third idea, that the government can monetise access, which critics say turns a security tool into a revenue stream. And the soft demand since suggests that, at least for now, the market has not delivered the windfall either side expected.
For Nvidia, the lesson is that policy volatility is a risk it can no longer fully control, so it has chosen to forecast without the market altogether. Any approval beyond that is upside.
The domestic angle
China has spent years building alternatives, and the decision lands in the middle of that effort. Huawei and other domestic vendors supply accelerators of their own, and Beijing has encouraged state-linked buyers to favor them, partly to reduce exposure to Washington's shifting rules. A reprieve for the H200 gives Chinese labs a faster option for training and serving large models, but it also gives policymakers in Beijing a reason to ask whether depending on American supply, even legally, is a risk worth taking.
Inside the US, the reaction was divided along familiar lines. Supporters of loosening argue that the H200 is no longer the frontier, that Chinese firms will otherwise buy from domestic suppliers and that revenue from sales funds American chip development. Opponents counter that compute is the scarce input to advanced AI and that handing over a capable part, even an older one, narrows the gap Washington is trying to preserve. The inclusion of a fee did little to resolve the argument; to some critics it reinforced the sense that national-security policy was being traded for revenue.
Whatever the merits, the episode shows how export controls have become an active bargaining tool, adjusted in response to diplomacy, rather than a fixed wall. That makes planning difficult for chipmakers and for customers, who must now price in the chance that rules change again in either direction.
What to watch
- Whether Beijing permits large-scale H200 purchases or keeps steering buyers to domestic chips.
- Any attempt to extend licences to Blackwell-class parts, which would be a much larger change.
- How the fee is legally structured and whether it is challenged.
- China-related line items in Nvidia's next quarterly report.
