Nvidia's H200 China Sales: A Door Opened Too Late

Nvidia has sold H200 accelerators to customers in China for the first time, but the volume is small: less than 1% of its $89.0 billion in second-quarter data center revenue, according to CFO Colette Kress. Washington opened the door in December 2025. Beijing has kept it mostly shut, and in the months since, Huawei's Ascend line has become the leading AI chip supplier inside China, by analysts' estimates. The H200 now arrives as a previous-generation product in a market that has learned to buy domestic.

The policy was meant to keep Chinese AI labs dependent on American silicon while the U.S. Treasury collected a share of every sale. As of September 2026, there is little public evidence that it has done either. What it has produced is a clearer picture of who controls this market now. That is Beijing, not Washington, and increasingly Huawei rather than Nvidia.

What did Nvidia actually disclose about H200 sales to China?

Nvidia reported results for the second quarter of fiscal 2027, which ended July 26, 2026, on August 26, 2026. Revenue was $96.2 billion, up 106% year over year, and data center revenue was $89.0 billion. The company guided third-quarter revenue to $108.0 billion, plus or minus 2%, and stated plainly: "NVIDIA is not assuming any Data Center compute revenue from China in its outlook."

On the earnings call, Kress gave the only hard figure on China: "In Q2, we shipped less than 1% of our total data center revenue in Hopper 200 products to customers based in China in accordance with the U.S. government licenses." She added that "current Hopper shipments are dilutive to corporate gross margins," according to the call transcript published by The Motley Fool.

Less than 1% of $89.0 billion puts the quarter's China H200 revenue somewhere under roughly $890 million. That is our arithmetic; Nvidia did not give a dollar figure. Yahoo Finance also reported that Nvidia took a $400 million charge for excess H200 inventory, a figure that does not appear in Nvidia's press release or on the call. For context, the company had previously told investors its China market share had collapsed. In an April 30 interview reported by Yahoo Finance, CEO Jensen Huang said Nvidia once had "90-some odd percent" of the market and that "today, in China, we have now dropped to zero."

How does the U.S. revenue-share arrangement work?

The arrangement was built in three steps, and the details matter because no public contract between Nvidia and the government sets out the "revenue share."

1. The announcement (December 2025). President Trump posted on Truth Social that "the United States will allow NVIDIA to ship its H200 products to approved customers in China," with 25% "paid to the United States of America," The Register reported. Blackwell and the coming Rubin generation stayed off the table.
2. The license policy (January 13, 2026). The Commerce Department's Bureau of Industry and Security moved H200-class exports to China to case-by-case review under three conditions: exports must not reduce global semiconductor production capacity available to U.S. customers, Chinese buyers must adopt compliance and customer-screening procedures, and the products must pass independent third-party testing in the United States.
3. The payment mechanism (January 14, 2026). A presidential proclamation imposed a 25% Section 232 import duty on certain advanced computing chips from January 15, with exemptions that include chips used in U.S. data centers and U.S. research and development, and no duty drawback. The proclamation does not name the H200. Trade coverage has described the duty as the mechanism for the government's 25% cut: China-bound H200s must be brought into the United States for testing, and they don't qualify for the data center exemption.

Licenses followed. According to Reuters, Commerce cleared around 10 Chinese firms to buy the H200 in May, including Alibaba, Tencent and ByteDance, though no deliveries had been made at that point. In March, Bloomberg reported (via the South China Morning Post) that the administration was weighing a cap of 75,000 H200s per Chinese customer, with AMD's MI325 counting toward the same limit. Later press reports have cited different limits, and we could not confirm the final per-customer cap from a primary source.

Why has so little actually shipped?

Because the binding constraint moved to Beijing. On July 14, Jeffrey Kessler, the Commerce Under Secretary for Industry and Security, told the House Foreign Affairs Committee there had been "minimal exports of any H200s to China so far" and described the number as "very few," Reuters reported.

Deliveries picked up over the summer. The South China Morning Post reported that Beijing began letting selected Chinese AI companies buy limited quantities in July. On August 19, the Financial Times reported that ByteDance and Tencent had each received about 10,000 H200 processors over the previous few weeks, as summarized by Engadget and other outlets. The same reporting said Chinese authorities had told companies to ship their H200 orders to Hong Kong instead of the mainland. Treat the Hong Kong detail as single-source reporting: it came from the FT and has not been confirmed by Beijing or Nvidia.

Nvidia's own framing is consistent with this. Yahoo Finance reported that the company described its China sales as "restricted by the PRC government." The Associated Press, citing analysts, reported in June that Beijing had switched to encouraging companies to use domestically designed chips from local suppliers led by Huawei.

Date	Event	Who controlled the outcome
Dec 2025	Trump announces H200 sales to approved Chinese customers, with 25% to the U.S.	Washington
Jan 13–15, 2026	BIS case-by-case license policy; 25% Section 232 duty takes effect	Washington
Apr 30, 2026	Huang says Nvidia's China share has "dropped to zero"	Beijing (purchases not approved)
May 2026	Around 10 Chinese firms licensed by Commerce, no deliveries yet (Reuters)	Beijing
Jul 14, 2026	Commerce official tells Congress shipments are "very few"	Beijing
Aug 19, 2026	FT reports ByteDance and Tencent have received about 10,000 H200s each	Beijing (selective approval)
Aug 26, 2026	Nvidia: China H200 sales under 1% of data center revenue; no China compute revenue in outlook	Beijing

The last column is the story. Every lever Washington pulled after January was about permission to sell. None of them could create permission to buy, and that is the half of the transaction China kept.

How much ground has Huawei gained?

According to a Bernstein report cited by the Associated Press in June, Nvidia held about 40% of China's AI chip market in 2025, roughly matched by Huawei. Bernstein projected that Nvidia's share would shrink to around 8% in 2026, while Huawei's grows to about 50%. These are one firm's estimates, not measured shipments, but the direction is consistent with Huang's own "zero" comment.

"Nvidia has definitely lost significant ground to Huawei, which (now) leads domestically," Antonia Hmaidi of the Mercator Institute for China Studies told the AP. He Hui, director of semiconductor research at Omdia, said: "China now believes in its own self-sufficiency and supply capabilities."

The product gap has narrowed as well. Huawei's September 2025 roadmap, presented by Eric Xu at Huawei Connect, scheduled the Ascend 950PR for the first quarter of 2026 and the Ascend 950DT, with 144 GB of in-house HBM and 4 TB/s of memory bandwidth, for the fourth quarter. The Atlas 950 SuperPoD, which links up to 8,192 Ascend chips, is also slated for Q4 2026, and Huawei gave it an international debut at MWC Barcelona in March. For comparison, Nvidia lists the H200, a Hopper-generation part, at 141 GB of HBM3e and 4.8 TB/s. Industry analysts quoted by the AP describe the Ascend 950 series as "roughly comparable" to the H200.

Is Huawei a global threat to Nvidia, or only a domestic one?

Mostly domestic, for now, and that distinction matters for how you read the policy debate. The Wire China reported in July that Nvidia told the U.S. government that "if Nvidia is kept out of the China market, Huawei production and sales will soar." The same report cited Epoch AI research estimating that Huawei accounted for just 6% of worldwide AI chip sales in the first quarter of 2026, and about half that once adjusted for performance. Chris McGuire of the Council on Foreign Relations told The Wire China that "Huawei simply does not pose a competitive threat to Nvidia globally right now."

Both things can be true. Huawei does not need to beat Nvidia globally to win the argument that matters in Beijing. It needs to be good enough, available at volume, and politically preferred inside China. On those three tests it is ahead of an export-licensed H200 that arrives in capped batches, requires U.S. testing, and carries the 25% U.S. levy.

Did U.S. export policy end up strengthening Nvidia's rival?

This is our analysis, not a finding from any single source: the sequence of U.S. policy handed Huawei the one thing it could not build on its own, a protected home market with committed demand.

• The ban created the opening. Earlier restrictions cut Chinese labs off from Nvidia's best parts, and Beijing responded by steering buyers toward domestic chips. In the April interview reported by Yahoo Finance, Huang said the approach had "largely backfired."
• The reopening came with friction. A 25% duty, mandatory U.S. testing, customer screening and reported per-customer caps make the H200 a hard sell against a domestic part that Beijing actively favors.
• The product that was cleared is aging. Blackwell and Rubin were excluded from the December decision. The H200 is Hopper-generation, and Huawei's roadmap targets it directly.
• Beijing learned it holds the veto. Selective approvals for ByteDance and Tencent show China can let in exactly as much American compute as it wants, when it wants, which is the opposite of dependence.

The counterargument deserves a fair hearing. U.S. officials can point out that Chinese labs are now getting fewer frontier-class chips than they would under an open market, and that Huawei's global output is still a fraction of Nvidia's. Export controls were never designed to preserve Nvidia's revenue. The trouble is that the December 2025 policy was explicitly sold on commercial grounds: keep Chinese buyers on the American stack and collect revenue while doing it. Judged on its own terms, the result is thin. It brought in under 1% of one quarter's data center revenue, and China's domestic champion has kept growing its share throughout.

What should operators and investors watch next?

• Nvidia's Q3 FY2027 report. Guidance assumes zero China compute revenue, so any China contribution is upside. Watch whether Kress quantifies H200 shipments again and whether the reported $400 million inventory charge grows.
• Beijing's approval list. If buyers beyond ByteDance and Tencent receive H200 allocations, that signals China is using the chips to fill specific training gaps. If not, the summer deliveries look like a one-off.
• Ascend 950DT and Atlas 950 SuperPoD shipments in Q4 2026. These are the parts meant to compete for training workloads, where Nvidia has held its strongest position.
• Congress. The July hearing showed lawmakers scrutinizing the licensing program. Tighter rules would shrink a channel that is already small.

For teams building on AI infrastructure outside China, the practical effect is limited: the H200 channel to China is too small to move global GPU supply or pricing. The strategic lesson is broader. Export policy can close a market quickly, but reopening it depends on the buyer, and the buyer has found a substitute. For a different case of export controls cutting off access to AI technology, see our report on the Anthropic Fable 5 and Mythos shutdown. For the scale Nvidia is operating at while China shrinks, see our earlier look at Nvidia's $4 trillion milestone. For the wider fight over custom silicon, see Google's TPU strategy.

Frequently Asked Questions

Has Nvidia started selling H200 chips to China?

Yes, in small volumes. Nvidia said that in its second fiscal quarter, which ended July 26, 2026, H200 shipments to customers in China made up less than 1% of its $89.0 billion data center revenue, under U.S. government licenses. The Financial Times reported that ByteDance and Tencent each received about 10,000 units. Nvidia's third-quarter outlook assumes no China data center compute revenue.

What is the 25% revenue share on Nvidia's H200 China sales?

President Trump announced in December 2025 that 25% of H200 sales to approved Chinese customers would go to the U.S. government. Trade coverage describes it as collected through a 25% Section 232 import duty proclaimed on January 14, 2026, though the proclamation does not name the H200. Chips must be tested by a third party in the United States before export, and chips destined for U.S. data centers are exempt from the duty.

Why is China limiting purchases of Nvidia H200 chips?

Beijing has encouraged companies to use domestically designed AI chips, led by Huawei's Ascend line, and has reportedly let only selected companies buy limited quantities of H200s. According to Financial Times reporting, authorities also told buyers to route H200 orders to Hong Kong rather than the mainland. Nvidia has described its China sales as restricted by the Chinese government.

How much of China's AI chip market does Huawei hold?

No official figures exist. Bernstein, as reported by the Associated Press in June 2026, estimated that Nvidia and Huawei each held about 40% of China's AI chip market in 2025, and projected about 50% for Huawei and around 8% for Nvidia in 2026. Treat these as analyst projections, not measured shipments.

Is the H200 better than Huawei's Ascend 950?

Industry analysts quoted by the Associated Press call the Ascend 950 series roughly comparable to the H200. On paper, Nvidia lists the H200 at 141 GB of HBM3e and 4.8 TB/s, while Huawei lists the Ascend 950DT, due in Q4 2026, at 144 GB and 4 TB/s. Real-world performance also depends on software, interconnect and supply.