Nvidia Stops Licensed AI Chip Shipments to China: What It Means

Awais Khalid

November 20, 2026

Nvidia Stops Licensed AI Chip Shipments to China: What It Means

Nvidia has stopped shipping its licensed Hopper-generation AI chips to China. On the company’s third-quarter earnings call on 17 November 2026, interim CEO Colette Kress said no further licensed Hopper shipments are planned, after $2.4 billion of such sales in the quarter, as The Neuron summarised from The Verge’s reporting. The decision closes, at least for now, the narrow channel Washington opened for Nvidia’s H200 chips in China less than a year ago. This article explains what Nvidia said, how the US-China chip fight reached this point, why the company is stepping back now and what it means for Nvidia, Chinese AI companies and cloud providers around the world.

What Nvidia said

  • No more licensed Hopper shipments: Kress said the company does not plan further shipments under existing US export licences.
  • $2.4 billion in the quarter: licensed Hopper data centre sales to China in the third quarter of fiscal 2027, according to the earnings call.
  • Guidance excludes China: Nvidia’s fourth-quarter outlook of $130 billion, plus or minus 3%, assumes no Data Center compute revenue from China. Our full report on Nvidia’s Q3 results covers the rest of the numbers.
  • A warning to cloud customers: CNBC reported on 17 November that Nvidia cautioned customers about potential legal risks from renting cloud computing capacity to Chinese companies.

The H200 in brief

The H200 is a Hopper-generation accelerator with 141GB of HBM3e memory. It sits a generation behind Nvidia’s current Blackwell and Vera Rubin products, but Tom’s Hardware notes it is still roughly six times more powerful than the H20, the cut-down chip Nvidia originally designed to comply with earlier US restrictions. For Chinese AI labs training and serving large models, it was the most capable Nvidia hardware they could legally buy.

How we got here: a timeline of Nvidia and China export controls

DateEvent
October 2022The US restricts exports of Nvidia’s top data centre chips, including the A100 and H100, to China.
October 2023Rules tighten further, catching the China-specific A800 and H800 variants.
April 2025The H20, Nvidia’s compliant China chip, is placed under a licence requirement; Nvidia takes a multibillion-dollar charge on unsold inventory.
August 2025A deal allows some H20 sales in exchange for the US government receiving 15% of the revenue.
December 2025President Trump approves H200 exports to approved Chinese customers, with the US taking a 25% cut; Blackwell remains banned.
January 2026New rules require Chinese buyers to show sufficient security procedures and guarantee no military use, and require Nvidia to certify adequate US inventory before shipping.
February 2026Nvidia secures a licence for a small number of H200 shipments, subject to inspection in the US and a 25% duty.
March 2026At GTC, Jensen Huang says Nvidia has purchase orders from Chinese customers and is restarting H200 manufacturing.
May 2026About 10 Chinese companies, including Alibaba, Tencent, ByteDance and JD.com, hold licences, but Reuters reports no chips delivered yet.
Q3 fiscal 2027 (Aug–Oct 2026)Nvidia records $2.4 billion of licensed Hopper sales to China.
12 November 2026Jensen Huang resigns as CEO; CFO Colette Kress becomes interim CEO.
16 November 2026The Commerce Department issues a rule restricting remote access to advanced chips through overseas data centres.
17 November 2026Nvidia says no further licensed Hopper shipments are planned.

The 2025 and 2026 milestones are drawn from reporting by TechCrunch, HotHardware, Bloomberg Law, Tom’s Hardware and TechRepublic.

The unusual terms of the H200 deal

The H200 arrangement was unlike any previous export licence. According to Tom’s Hardware and Bloomberg Law, each chip had to be routed through US territory for inspection and carried a 25% duty, effectively giving the US government a quarter of the revenue. Chinese buyers had to prove they had adequate security procedures and guarantee the chips would not be used by the military. Nvidia itself had to certify that it held sufficient inventory in the United States before exporting. The intent was to let Nvidia compete in China, generate revenue for the US government and keep the most advanced Blackwell chips out of reach.

Why Nvidia is stepping back now

Nvidia has not given a single reason, but four pressures converged in the same week.

1. The new remote-access rule. On 16 November the Commerce Department issued an interim final rule requiring licences for advanced chips sent to data centres anywhere that may supply large volumes of compute to companies headquartered in China, and adding licence requirements for five more countries. We explain it in our guide to the remote-access rule. It signalled that Washington’s mood had shifted from managed sales to tighter enforcement.

2. Congressional scrutiny. On 18 November, CNBC reported that a congressional commission had warned about China’s military interest in Nvidia chips. Political pressure on any company selling advanced AI hardware into China rose accordingly.

3. Beijing’s reluctance. Even with licences in hand, sales had been slow to materialise. TechRepublic, citing Reuters, reported that officials in Beijing were uneasy about possible tampering or hidden vulnerabilities while the hardware sat on US soil for inspection. China has also been encouraging companies to buy domestic accelerators.

4. A leadership transition. Jensen Huang, who personally championed access to the Chinese market, resigned on 12 November. An interim leadership team has less reason to defend a politically contentious, relatively small revenue line.

What it means for Nvidia financially

In the short term, the impact is small. The $2.4 billion of Hopper sales was about 2% of Nvidia’s $115.8 billion quarterly revenue, and the company’s outlook already assumed nothing from China. Demand from US hyperscalers, AI labs and enterprise buyers is growing far faster than China ever did for Nvidia’s data centre business.

The longer-term cost is strategic. China was once a double-digit share of Nvidia’s revenue. Every quarter that Chinese companies cannot buy Nvidia hardware, they invest more in domestic alternatives and in software tuned for those chips. If restrictions ease in future, Nvidia may find it harder to win back customers who have rebuilt their stacks around other hardware.

What it means for Chinese AI companies

For Chinese labs, the H200 was a stopgap: older than the chips their US rivals use, but well ahead of what was otherwise available legally. Losing it pushes them further toward domestic accelerators and toward efficiency, training and serving capable models with less compute. Chinese labs have already shown they can produce competitive open-weight models under these constraints, and that pressure is likely to intensify.

What it means for cloud providers and developers outside China

The clearest message from both Washington and Nvidia this week is the same: know who is using your GPUs. Cloud and colocation providers in Southeast Asia, the Middle East and elsewhere that rent capacity to companies headquartered in China face licence requirements under the new rule and, according to Nvidia’s warning, potential legal exposure. Expect stricter know-your-customer checks, longer onboarding for new customers and less spare top-tier capacity in regional hubs such as Singapore and Malaysia. Developers in those markets may see higher prices for advanced GPU time.

Market reaction

Semiconductor stocks were already under pressure. On 19 November, the day after the comments circulated, Nvidia closed down about 3.3% and AMD fell about 7.7% as the wider US market slid, according to INDmoney’s market summary. The Philadelphia Semiconductor Index had fallen more than 5% on 16 November, the day the remote-access rule was published.

What happens next

  • US-China talks: President Trump’s meeting with Xi Jinping could change the outlook in either direction; chips are a likely bargaining item.
  • Rule implementation: the remote-access rule’s provisions phase in over the coming months, and its practical effect depends on how strictly licences are granted and audited.
  • Nvidia’s leadership: a permanent CEO may take a different view of the China market.
  • Chinese domestic chips: watch for announcements from Chinese chipmakers and labs about capacity and performance gains.

How US export controls on AI chips work

US export controls on AI chips are administered by the Commerce Department’s Bureau of Industry and Security. Rather than naming individual products, the rules set performance thresholds: any chip above a certain level of computing performance and memory bandwidth needs a licence before it can be exported to restricted destinations such as China. Companies on the Entity List, a register of foreign businesses judged to pose national security risks, face even stricter limits. When Nvidia designed lower-performance chips such as the A800, H800 and H20 to fall under those thresholds, Washington responded by lowering the thresholds or adding licence requirements, which is why each generation of China-specific chips eventually became restricted.

The licence for the H200 was different because it was a deliberate exception: a chip well above the thresholds, allowed through on condition of inspection, a 25% duty and end-use guarantees. The 16 November remote-access rule extends the same logic to compute that is rented rather than bought, by making data centres themselves subject to licensing when they serve restricted customers at scale.

The cloud loophole, explained

For years, the gap in export controls was simple. A company in China that could not buy Nvidia’s best chips could rent time on them from a data centre in another country. The chips never crossed into China; only the computing results did. Because cloud services cross borders as data, not as hardware, this was very hard to police. Reports of chips moving through third countries, particularly in Southeast Asia, added to concerns about physical smuggling as well.

The new rule tries to close both routes. Data centres that may supply large volumes of advanced compute to companies headquartered in China, Russia or Iran need licences and must check who their customers are, and exports of advanced chips to companies headquartered in Indonesia, Malaysia, Singapore, Thailand and Brazil now require a licence. Nvidia’s warning to its own cloud customers suggests the company expects enforcement to be real.

China’s push for domestic AI chips

China’s response to US controls has been to accelerate its own semiconductor industry. Chinese technology companies are designing their own AI accelerators, and the country’s largest foundry, SMIC, has grown quickly even under restrictions; it crossed $3 billion in quarterly revenue for the first time this year, as we reported in our coverage of SMIC’s results. Domestic chips still trail Nvidia’s best on performance and on the maturity of their software ecosystems, but each restriction gives Chinese buyers a stronger reason to invest in alternatives rather than wait for US licences.

That dynamic is central to the policy debate. If Chinese companies build competitive chips and software stacks, Nvidia loses a market permanently and US leverage weakens. If they cannot, controls slow China’s AI progress. Both outcomes are possible, and the evidence so far is mixed: Chinese labs have released strong open-weight models while using less advanced hardware, suggesting software efficiency can partly offset hardware limits.

The arguments on both sides

The case for allowing limited sales. Supporters, including Nvidia, argued that selling older chips to vetted commercial customers keeps Chinese developers reliant on US technology, funds American research and manufacturing, and generates revenue for the government through the 25% cut. When the H200 approval was announced, an Nvidia spokesperson told TechCrunch the arrangement “strikes a thoughtful balance that is great for America.”

The case against. China hawks in Washington reject the idea that limited sales will stop Chinese firms from narrowing the gap, according to TechRepublic. They argue that any advanced compute can be diverted to military or surveillance uses, that vetting and end-use guarantees are hard to verify, and that helping Chinese labs today strengthens a strategic competitor tomorrow. The congressional commission’s warning about China’s military interest in Nvidia chips reflects this view.

Both camps agree on one point: rules that are announced and then reversed create uncertainty for companies on both sides of the Pacific. Chinese buyers who received licences in May but no chips illustrate how hard it is to plan around shifting policy.

What it means for AMD and other chipmakers

Nvidia is not the only company affected. AMD also sold China-specific accelerators under the 2025 revenue-sharing arrangement, and its shares fell about 7.7% on 19 November amid the broader semiconductor sell-off. Memory makers that supply high-bandwidth memory for AI accelerators, and equipment makers that sell into Chinese fabs, are exposed to the same tightening trend. For investors, the lesson of this week is that China revenue in AI hardware should be treated as optional upside rather than a dependable stream.

What Nvidia investors should watch

  • Fourth-quarter results: whether Nvidia hits its $130 billion outlook without China.
  • Comments from the next CEO: any shift in strategy toward China once a permanent leader is appointed.
  • Inventory: whether unsold H200 inventory leads to write-downs, as the H20 restrictions did in 2025.
  • Cloud demand outside China: whether stricter customer checks slow GPU purchases by regional cloud providers.

Key takeaways

  • The channel is closed for now. Nvidia’s licensed H200 route into China, approved in December 2025 and slow to start, produced $2.4 billion of sales in one quarter before Nvidia said no more shipments are planned.
  • The financial hit is small, the strategic one is not. China is now about 2% of Nvidia’s quarterly revenue, but every quarter without Nvidia hardware pushes Chinese buyers further toward domestic chips.
  • Enforcement is moving to the data centre. The 16 November rule means cloud providers, not just chipmakers, now carry export-control obligations when they serve Chinese customers at scale.
  • Policy can change quickly. US-China talks, a new Nvidia CEO and congressional pressure could all shift the picture again in 2027.
  • For AI users outside China, the practical effect is stricter customer checks and potentially tighter, pricier GPU capacity in regional cloud hubs.

Taken together, the week’s events mark a turning point in the export-control debate. The H200 experiment tested whether managed, taxed sales of older chips could satisfy both Washington and Beijing. Neither side fully embraced it: US hawks saw it as too generous, and Chinese officials were wary of hardware that had been inspected on American soil. With Nvidia stepping back and Commerce tightening the cloud route, the next phase of the US-China AI rivalry looks likely to be fought less over which chips can be sold and more over who can access compute at all.

FAQ

Has Nvidia stopped all sales to China?

Nvidia says it plans no further licensed Hopper data centre shipments, and its guidance assumes no Data Center compute revenue from China.

How much did Nvidia sell to China last quarter?

About $2.4 billion of licensed Hopper data centre chips in the third quarter of fiscal 2027.

What is the H200?

A Hopper-generation Nvidia AI accelerator with 141GB of HBM3e memory, a generation behind Blackwell but far more powerful than the China-specific H20.

Why did the US allow H200 sales in the first place?

The administration argued limited sales would let a US company compete in China while generating revenue for the government, which took a 25% cut, and keeping the newest Blackwell chips out of reach.

Does this affect AI users outside China?

Indirectly. Cloud providers outside the US face stricter customer checks under the new remote-access rule, which may tighten capacity and raise prices for advanced GPU time in some regions.

Sources

Stay Ahead of AI

Get the latest AI news delivered to your inbox.

We don’t spam! Read our privacy policy for more info.

Published

Spotted an error? Read our corrections policy or email editorial@perplexityaimagazine.com.