The US government has cleared Nvidia H200 chip sales to 10 firms in China, according to Reuters, marking the first major softening of export controls that had walled off Nvidia's flagship AI accelerator from the world's second-largest chip market. The approvals land as Nvidia CEO Jensen Huang lobbies Washington for broader access, and they signal that the Trump administration is willing to use case-by-case licensing rather than a blanket ban. The names of the 10 firms have not been disclosed.
The H200 is the high-bandwidth memory upgrade to Nvidia's H100, the chip that defined the generative-AI boom. It ships with 141GB of HBM3e memory and is the most powerful Hopper-generation part Nvidia produces, sitting one tier below the new Blackwell line. Until now, Chinese buyers were restricted to the cut-down H20 variant Nvidia designed specifically to clear US export thresholds.
For Nvidia, even a narrow carve-out matters. China accounted for roughly a fifth of Nvidia's data-center revenue before the 2022 and 2023 controls tightened, and the company has spent the past two years watching that share collapse toward zero on its top-tier parts. Huang has been publicly direct that ceding China entirely hands the market to Huawei and domestic accelerator startups.
Key facts
- 01The US government has approved Nvidia H200 chip sales to 10 China-based firms, according to Reuters.
- 02The H200 is Nvidia's highest-bandwidth AI accelerator before the Blackwell generation and was previously blocked from China entirely.
- 03The clearance comes as Nvidia CEO Jensen Huang pushes for a diplomatic breakthrough on China access.
- 04The identities of the 10 firms have not been disclosed in the initial reporting.
The 10-firm green light is also a notable shift in tone from the previous policy posture, which treated any Hopper-class export to China as a national-security risk. Approving the H200 specifically — not just the export-compliant H20 — suggests the administration is now drawing a line at Blackwell and above, rather than at Hopper and above. That is a meaningfully different containment perimeter.
“The H200 carries 141GB of HBM3e memory and was previously walled off from China entirely under US export controls — until 10 named firms got the green light.”— Jaeden Schafer
Huang has framed the China question as a commercial fight Nvidia can lose if Washington overreaches. His pitch in recent months has been that US chips inside Chinese data centers extend American software and tooling dominance, while a full ban accelerates Huawei's Ascend roadmap and pushes Chinese developers off CUDA. The 10-firm clearance reads as at least a partial win for that argument.
What remains unclear is the scale. Reuters has not reported volume caps, end-use restrictions, or whether the 10 firms include hyperscalers like Alibaba, Tencent, and ByteDance — the buyers that would actually move the revenue needle — or smaller AI labs and research outfits. A licensing regime that approves names but caps units could look generous on paper while delivering modest shipments in practice.
Domestic Chinese competition is the backdrop. Huawei's Ascend 910C has been shipping to Chinese AI labs through 2025, and a handful of model developers have publicly trained on the part. Each month the H200 was unavailable in China was a month Ascend volumes compounded — which is precisely the dynamic Huang has been warning about in his Washington meetings.
On the policy side, expect pushback. China-hawk lawmakers in both parties have argued that Hopper-class chips in Chinese data centers materially advance frontier AI capability there, and the H200's memory bandwidth is a meaningful step up for inference workloads on large models. Any expansion of the 10-firm list will be litigated publicly.
Critics will also note that the 10 firms have not been named, making it impossible to assess whether the licensees have ties to the Chinese military or to entities on existing US restricted lists. Reuters' reporting is the first read, and the underlying license terms — duration, renewal, and audit requirements — have yet to surface.
The market read is straightforward. If Nvidia can sell H200s into China at scale, the 2026 data-center revenue line gets a tailwind the Street has not been modeling, and the competitive runway for Huawei Ascend narrows. If the 10-firm approval turns out to be a symbolic one-off with tight unit caps, it changes little. The next data point to watch is whether the list expands — and whether Blackwell follows Hopper through the same door.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
