SANTA CLARA (Realist English). Nvidia has stopped shipping Hopper data-center chips approved by the US authorities to China, even though the export license remains in force. This was stated by the company’s interim CEO Colette Kress during the call on the results of the third quarter of fiscal year 2026.
Background: How the License Came About
In December 2025, the Trump administration allowed Nvidia to sell H200 chips to certain Chinese clients on condition that the chips pass through US territory and Washington collects a 25 percent export fee. According to reports, this arrangement became possible thanks to the personal relationship between then-CEO Jensen Huang and Trump. The H200 belongs to the Hopper architecture and lags two generations behind the current Rubin lineup.
What Triggered the Halt in Shipments
The decision coincided in time with several events. On November 12, Jensen Huang stepped down as CEO amid a scandal over an extramarital affair; one of the women involved in it, according to the American side, is suspected of ties to Chinese intelligence and of possible assistance in smuggling chips to China. Beijing denies these accusations.
On November 16, the Trump administration issued new rules restricting the sale of American processors to foreign data centers that plan to provide Chinese companies with remote access to advanced AI chips. The new rules close the loophole under which “the chip does not leave the country, but computing power is exported remotely.”
At the same time, pressure from Congress intensified: House Speaker Johnson stated that a bipartisan consensus on restricting AI chip exports to China is “gradually taking shape.”
A number of bills are being advanced, including the “Chip Security Act,” which requires tracking the location of chips, and the “Remote Access Security Act,” which restricts remote access.
Kress did not name a specific reason for halting the shipments, but this decision marks a notable shift in Nvidia’s position — Jensen Huang was the most active advocate of restoring exports to China.
Financial Consequences: Chinese Revenue Was Minimal Anyway
The share of approved Hopper shipments to China was about 2% of Nvidia’s total revenue of $115.8 billion in the third quarter. Data-center segment revenue for the quarter was $107.6 billion, up 110% year on year. The company’s guidance for the next quarter — about $130 billion — explicitly assumes zero revenue from data-center computing in China.
The historical reference point is the lesson of 2025. When Washington in 2025 required a license for the export of H20 chips (a special version for China), Nvidia recorded $4.5 billion in excess losses on inventory and purchase obligations, and margin fell from 74.6% to 60.5%.
But the company’s revenue for that year still grew by 65%. This is the baseline scenario under which Nvidia remains calm about the halt in shipments: direct sales to China are small and already excluded from guidance, and policy losses can be absorbed within a single quarter.
The real risk is not Chinese revenue itself, but the “peripheral layer.” Nvidia provided guarantees under the lease agreements of partner data centers; as of July 2026, the disclosed maximum volume of such obligations was $3.5 billion, after which new guarantees of up to $105 billion were signed — a total of about $108.5 billion, alongside simultaneous supply and capacity obligations of $279 billion.
If data centers in third countries lose Chinese clients due to a license denial and default, the losses will not be limited to the operator’s report but will shift to Nvidia’s guarantee obligations.
The Logic of Policy: From “Flow of Chips” to “Flow of Computing”
The key change in the new rules is the shift in the object of regulation. Previously, control focused on the physical cross-border movement of chips; the rules of November 16 extend jurisdiction to data centers with “covered capacity” (over 12 million aggregate total processing performance), wherever they are located, if they serve clients headquartered in China.
Exceptions are provided if the recipient or end user is headquartered in the US or has an ultimate parent company headquartered in the US.
The new rule directly targets the existing model of “separation of the billing address and the actual destination.” For example, in the quarter ended July 2025, Nvidia billed $10.16 billion (22% of revenue) to Singapore clients, but the 10-Q form states that local clients “are used for centralized billing, and products are almost always shipped elsewhere,” and more than 99% of controlled data-center computing comes from orders placed by American clients. The government of Singapore told parliament that less than 1% of Nvidia’s revenue physically enters the country.
The Actual State of the Chinese Market
The volume of H200 in circulation in China is extremely small, and prices are exorbitantly high. Reuters reported that about 600,000 H200 chips were stuck in Hong Kong because Beijing did not approve large-scale imports. Due to the shortage of official supplies, a server with 8 H200 chips in mainland China cost more than $900,000 in late October — roughly three times more than a year earlier.
The position of Chinese regulators also creates obstacles. Georgetown University researcher Jacob Feldgoise notes that the Chinese government previously refused to import H200 in order to support the domestic chip industry.
However, Chinese AI labs still use a limited number of Nvidia Grace Blackwell chips in overseas data centers to train their flagship models.
Market share has shrunk significantly. Revenue in China (including Hong Kong) for fiscal year 2026 was $19.7 billion — about 9% of total revenue of $215.9 billion, down from $25 billion a year earlier. Nvidia’s share of the Chinese AI chip market fell from 55% to 34%, while Huawei is rapidly closing the gap. Data-center revenue in China before the ban took effect reached $4.6 billion per quarter, after which it approached zero.
Synchronous Steps at the Corporate Governance Level
On November 17, Nvidia announced the creation of a board-level national security and export control committee, which will be responsible for overseeing export compliance and global trade risks.
New board chairman Harvey Jones stated that the board recognizes that the company “is in a world where national security dynamics are bringing serious changes to our markets,” and that the next CEO will have to “navigate these waters very carefully.” Jones also stated directly that the board does not intend to bring Jensen Huang back to the position.
As observers note, the key feature of the situation is the gap between the formal existence of the license and the actual halt in shipments. Nvidia is stopping shipments not because it is prohibited by law, but because, after Jensen Huang’s departure and amid a sharp tightening of the regulatory environment, the company made an independent adjustment.
The financial consequences are limited, since Chinese data-center revenue is already zeroed out in guidance; however, the institutional consequences are deeper — the logic of American export control is shifting from the question “does the chip cross the border” to the question “does the computing power serve Chinese clients,” which makes data-center operators in third countries and the financing and guarantee structures behind them a new layer of risk.
For the AI chip market, the immediate effect will be visible primarily in the dynamics of shipments of Vera Rubin systems and in assessments of Nvidia’s guarantee obligations to data-center operators — and that is precisely where it is most sensitive for the balance between export control and the global expansion of American manufacturers.







