China’s NVIDIA H200 Reversal Falls Short of What AI Labs Demanded

Beijing is preparing to let a small group of its top AI labs buy NVIDIA’s H200 chips again, capping the total below 200,000 units and reserving the hardware for AI training only, according to a report from The Information carried by Reuters on Wednesday.

The reversal lands more than a year after Beijing told domestic AI firms to favor Chinese-made chips, a policy meant to fund a self-reliant national compute stack. With Huawei and its peers unable to keep pace with the compute demanded by the country’s largest model-training jobs, regulators are now letting three named buyers, Alibaba, ByteDance, and DeepSeek, back in through a tightly controlled door.

Beijing Says Yes, but Only for Training

Reuters, citing two people with direct knowledge of the matter, reported that Chinese officials have told Alibaba, ByteDance, and DeepSeek in recent weeks that they may soon receive permission to buy some H200 chips. NVIDIA shares rose about 1% in Wednesday morning trade after the wire landed. Neither NVIDIA, the U.S. Commerce Department, China’s commerce ministry, nor the three named Chinese firms responded to Reuters before publication.

The cap is the binding detail. The Information reported that Beijing is still determining the exact number of chips to approve, and that the total could amount to fewer than 200,000 H200 chips, less than half of what the companies requested earlier this year. The approvals are not blanket authorizations to stock any data center: approved firms may use H200 chips only for AI training, not for inference workloads that serve answers to end users, and the chips are to handle public data rather than sensitive customer information. Beijing’s domestic-chip push is paused, not abandoned. For the full wire, see the Reuters write-up of The Information’s report.

Why the Training Crunch Finally Beat the Self-Reliance Push

The reversal is a confession as much as a concession. The Reuters wire tied the move directly to “the growing computing capacity crunch that the country’s tech companies are facing.” A year of pushing domestic alternatives has not produced enough volume at the training tier to absorb the country’s largest model runs.

The numbers explain the urgency. In 2024, Chinese firms bought around 1 million NVIDIA H20 chips against an estimated 450,000 Huawei Ascend 910B shipments, per RAND. By 2025, Chinese domestic chips had reached nearly 41% of the country’s AI accelerator market, with Huawei accounting for roughly half of that, according to Brookings.

The 2026 figures do not close the gap. Cambricon is targeting 500,000 AI chips this year, with low yields and limited high-bandwidth memory flagged as the binding risks, per Tom’s Hardware. SMIC’s leading 7nm-equivalent line is producing about 2.6 million AI chips against roughly 4.2 million in domestic demand, leaving a real shortage for the training workloads Huawei’s Ascend lineup still trails. Beijing had three paths: wait, lean harder on the local stack, or open a narrow door. It chose the third.

• 2024: About 1 million NVIDIA H20 chips sold into China.
• 2024: An estimated 450,000 Huawei Ascend 910B chips shipped.
• 2025: Chinese domestic chips near 41% of the country’s AI accelerator market.
• 2026: Cambricon targeting 500,000 AI chips, with low yields and limited HBM flagged.
• 2026: SMIC’s 7nm-equivalent line producing about 2.6 million chips against demand near 4.2 million.

That backdrop is also why NVIDIA’s re-entry is graded on a curve. For Cambricon’s production math, see Cambricon’s 500,000 chip production target for 2026.

The Caps and Conditions Beijing Wrote Into the Approval

The Information’s report lays out four conditions, not a blanket rule. They amount to a budgeted re-entry sized so the chips cannot become the default choice for any workload Beijing wants to keep local.

The conditions are written for a fixed audience: the country’s three biggest AI labs and their compute planners, plus the procurement teams at Huawei and Cambricon who will be asked to absorb everything outside this list.

1. Quantity cap. Total approvals of fewer than 200,000 H200 units, less than half of what Chinese AI labs had requested earlier in 2026.
2. Training-only use. Approved firms may use the H200 for AI training; Beijing wants inference workloads steered to Chinese-developed processors.
3. Data scope. Approved chips are to handle public data, not sensitive customer information.
4. Named buyers only. Permission is being extended to a small group, with Alibaba, ByteDance, and DeepSeek named as the firms recently told to expect clearance.

Those four lines are the terms of trade. Anything outside them still defaults to the Chinese stack.

A Win for NVIDIA, but Smaller Than the Tape Suggests

The 1% share move on Wednesday understates how much NVIDIA needs this market and how carefully the company has stopped claiming it.

We’ve really largely conceded that market to them.

The quote is Jensen Huang’s, from a May 21, 2026 interview with CNBC’s Sara Eisen. Huang was talking about Huawei and the local Chinese chip ecosystem. After telling CNBC that NVIDIA had “evacuated that market,” he framed his expectations bluntly to investors.

“I don’t have any expectation,” Huang said, “which is the reason why we put all of our guidance, all of our numbers, all the expectations that I’ve set with all of our analysts and investors to invest nothing, to expect nothing.” Then the door. “We would be more than delighted to serve the market. We have a lot of customers there, we have a lot of partners there, and we’ve been there for 30 years.” For the full interview, see Jensen Huang’s May admission that NVIDIA has conceded China to Huawei.

The math also constrains the upside. Each H200 that crosses into China carries a 25% tariff, set by the Trump administration in January. The H200 sits a generation behind NVIDIA’s current flagship; the Vera Rubin platform entered full production this year and is rolling out from partners in the second half of 2026, per NVIDIA’s announcement that the Vera Rubin platform is now in full production. Beijing’s cap applies to the older architecture, not the new one. Reuters also reported last month that NVIDIA is pitching a companion “Vera” CPU to Chinese clients with deliveries eyed as soon as August. For coverage of that separate CPU push, see NVIDIA’s Vera CPU hitting mass production for the first time. On the levy itself, see Trump’s 25 percent tariff on H200 sales to China, set in January.

Where Huawei, Cambricon, and DeepSeek Sit After the Reversal

Huawei and Cambricon were the two names Beijing’s chip policy was built around. They get a smaller protected market, not a closed one. Huawei already holds roughly half of China’s $50 billion domestic AI chip market, a position built largely on the U.S. export ban that kept NVIDIA’s best chips out.

Cambricon, Beijing’s other flagship accelerator maker, is targeting 500,000 AI chips in 2026 and ramping production, with limited high-bandwidth memory flagged as the binding constraint. The arithmetic for both firms narrows but does not collapse. The H200s allowed in will absorb the training stage where Huawei’s Ascend lineup still trails NVIDIA by a wide margin, while inference volumes, which scale with usage rather than model size, remain steered toward Chinese processors.

DeepSeek is the most exposed third party. The startup trained earlier models on NVIDIA’s H800 and shifted much of its recent work, including its V4 release in April 2026, onto Huawei’s Ascend chips. Reuters also reported this week that DeepSeek is separately developing its own inference chip to cut its dependence on both NVIDIA and Huawei at once. Industry surveys cited in The Information’s coverage now point to Chinese firms shifting 46% of their AI-accelerator budgets to domestic suppliers within a year, up from around 30%. The H200 re-entry makes room at the margin; it does not reverse the trend.

A Year of H200 Whiplash

Seven dated turns in twelve months made Wednesday’s report possible. Read in order, they explain why Beijing’s reversal carries more conditions than usual and why NVIDIA’s 1% share move was the size it was, not a re-rating.

Each turn tightened or loosened the same chip’s status. None of them, on its own, put a single H200 in a Chinese data center.

1. Dec 8, 2025: Reuters carries a Semafor report that the U.S. will allow H200 chip shipments to China.
2. Jan 14, 2026: President Trump announces a 25% tariff on H200 sales to China.
3. Jan 28, 2026: Reuters reports China is set to approve imports of more than 400,000 H200 chips.
4. May 14, 2026: The U.S. Commerce Department clears H200 sales to roughly 10 Chinese firms, including Alibaba, Tencent, and ByteDance.
5. May 21, 2026: Jensen Huang tells CNBC NVIDIA has “largely conceded” the China AI chip market to Huawei.
6. Jun 12, 2026: NVIDIA tells Chinese clients a new “Vera” CPU for AI data centers may be available as soon as August, sources tell Reuters.
7. Jul 8, 2026: The Information reports Beijing is preparing to let Alibaba, ByteDance, and DeepSeek buy H200 chips, capped below 200,000 units.

Beijing is still determining the exact number, and the volume that lands in signed orders during the third quarter will set the practical size of the re-entry.

Frequently Asked Questions

What is the NVIDIA H200?

The H200 is a data-center GPU built on NVIDIA’s older Hopper architecture. NVIDIA specifically designed it to satisfy U.S. export rules while keeping enough performance for large-model training jobs.

Why is Beijing reopening the door now?

Regulators cited the same training-compute crunch Chinese AI labs have been raising with domestic policymakers for months, a gap domestic chip suppliers could not yet fill at the volume required.

What does “training-only” mean in Beijing’s approval?

Approved firms can run the H200 during the training stage, when raw compute matters most and a single model’s cost stacks up fast. Beijing wants inference workloads, the tasks that serve answers to end users, steered toward domestic chips.

How does this differ from the earlier self-reliance push?

The earlier push was an exclusion: keep NVIDIA out and force the local stack to grow. The current arrangement is a split: let foreign chips in for the training stage, keep inference and customer-data processing local, and cap the volume.

Could Beijing raise the cap later?

The Information described the figure as “fewer than 200,000” and as a number Beijing is still determining. A higher or lower final figure has not been confirmed by Chinese regulators or the named buyers.