President Donald Trump announced Monday that the United States will permit Nvidia to sell its more powerful H200 AI chips to certain customers in China — but not the top-tier Blackwell family — under a deal that, unusually, directs about 25% of the sales revenue to the U.S. government.
The move marks a striking pivot in Washington’s approach to restricting advanced AI hardware: after months of trade negotiations and close talks between Nvidia’s CEO Jensen Huang and the White House, the administration has chosen a controlled path that couples market access with revenue-sharing and customer vetting.
What was approved — and what wasn’t
According to the announcement, Nvidia may ship Hopper H200 accelerators to “approved customers” in China and elsewhere. The H200 is a step up from the H20 chips that had previously been allowed for export. But the administration drew a hard line on the newer Blackwell architecture — the chips widely expected to power the most demanding generative AI workloads — which remain off limits.
The administration’s reported conditions include an approval process for recipients and a requirement that a quarter of the sales proceeds be remitted to the U.S. government. Chinese President Xi Jinping is said to have “responded positively” to the arrangement, according to the president’s social media post, signaling at least a diplomatic opening for the deal to be operationalized.
Markets and motive
Investors reacted calmly. Nvidia’s stock ticked up after hours and semiconductor peers nudged higher in sessions following the announcement. Traders are parsing what the concession means for supply chains, competitive advantage and corporate profits. For Nvidia, access to China — the world’s largest market for datacenter compute — could unlock substantial sales even after the government’s cut.
Behind the scenes, this is about more than one company’s bottom line. The U.S. administration appears to be attempting a middle path: tempering the export ban’s blunt disruption to global tech supply chains while preserving a degree of leverage over how advanced hardware flows into strategic adversaries’ AI ecosystems.
Enforcement headaches and technical fence‑lines
The practicalities will be tricky. Defining an “approved customer” in a way that prevents diversion to military or surveillance programs is technically and legally thorny. Chip sales often sit inside complex supply chains — servers, OEMs, system integrators and cloud providers — making post‑shipment controls hard to police.
Moreover, relying on a revenue share as a form of influence is novel. Twenty‑five percent of gross sales is a blunt instrument: it raises revenue but does little by itself to prevent sensitive technology from being used in ways Washington deters. Critics will ask whether the arrangement merely monetizes risk without materially reducing it.
Strategic ripples
Allowing H200 exports under conditions will reshape how companies plan product road maps and partnerships. Cloud providers and AI developers in China who can qualify as approved buyers will gain access to higher performance hardware, potentially accelerating model training and inference capabilities.
That matters not only for commercial applications but for national competitiveness. The demand for such accelerators is being driven by increasingly large and sophisticated AI models; Microsoft’s recent AI efforts, for example, underscore a broader arms race for compute resources — see Microsoft’s MAI work and other in‑house models like Microsoft’s MAI‑Image‑1. Meanwhile, new approaches to locating compute — from terrestrial hyperscale centers to far‑out ideas — show why hardware policy resonates well beyond chip fabs. Projects exploring novel infrastructure, such as Google’s Project Suncatcher, underscore how data‑center strategy and chip policy are now intertwined.
Political optics and precedent
Domestically, the deal will draw scrutiny from lawmakers across the aisle. Some will view it as a pragmatic compromise that preserves jobs and commercial ties; others will see it as a risky concession that could undercut long‑term security goals. For trade policy, the arrangement creates a precedent: Washington’s willingness to trade limited market access for direct financial share and recipient controls could be replicated in other sectors or with other countries.
Internationally, Beijing’s positive reception is notable but not definitive. Agreement in principle must translate into contracts, compliance mechanisms and inspections — or into a pragmatic accounting exercise where money flows but controls are porous.
A fragile middle ground
This decision carves out a delicate middle ground between full export bans and unrestricted trade. It acknowledges the economic realities — the scale of AI demand, Nvidia’s dominant position and the clout of Chinese buyers — while trying to retain levers of influence. Whether the approach will slow the diffusion of frontier AI capability or simply channel it through a taxed pipeline remains to be seen.
For now, companies, investors and policymakers will be watching how the administration defines approved recipients, enforces the revenue arrangement and prevents technology diversion. The tradeoffs are immediate: more compute for some Chinese users, a revenue stream and a say for the U.S. government — and, potentially, new headaches in monitoring and enforcement.
If the past two years taught the tech world anything, it’s that compute scarcity shapes who wins in AI. This policy changes who can get access — and how much Washington is willing to charge for the privilege.