According to a Financial Times (FT) report on November 3, NVIDIA is asking cloud providers to sell at least 51% of the services they offer under future contracts to companies outside China. The report is based on five people familiar with the matter, and NVIDIA has not officially confirmed the policy. In the United States, a bill that would bring remote access to advanced AI chips through overseas data centers under export controls has stalled, so NVIDIA may be trying to limit Chinese companies' access to computing resources through its own sales terms.
How much impact the policy has, however, will depend heavily on how the 51% ratio is calculated. Providers whose business centers on leasing dedicated GPU clusters to a single Chinese company could be forced to rethink their business structure if a cap of 49% on China-bound services is applied. Two large usage contracts disclosed by Japan's Data Section (TSE Growth: 3905) total about $1.297 billion in usage fees over three years alone. The question is how such large contracts would be affected by the new sales terms.
What is NVIDIA's "51% rule"?
On November 3, 2026, the FT reported that NVIDIA is asking cloud providers to sell at least 51% of the services they offer to companies outside China. The requirement reportedly applies to contracts signed from now on, not to existing ones. It is also said to be communicated verbally through NVIDIA sales representatives and regional teams rather than in writing. According to two people familiar with the matter, it covers customers around the world, not only data centers in Asia. The FT says this is the first time NVIDIA has imposed this kind of sales restriction.
NVIDIA neither confirmed nor denied the report in response to AFP, instead stressing the importance of U.S. AI infrastructure remaining competitive in global markets. The company said that if data centers and clouds built on U.S. chips do not win globally, the United States will not win the AI race either. China's Foreign Ministry spokeswoman Mao Ning said she was not aware of the report and expressed hope that companies from all countries can operate in a fair, transparent and non-discriminatory business environment.
The biggest unknown right now is the scope over which the 51% ratio is calculated. Whether it is judged contract by contract or against a cloud provider's overall revenue or service volume changes which providers would be affected. It is also unclear what the ratio is based on: revenue, contract value, compute usage, or something else. Because no formal written notice has been confirmed, how enforceable the condition is, and what happens in case of violation, is also unknown.
According to NVIDIA's Form 10-Q for the second quarter of fiscal 2027, revenue from China (including Hong Kong) for the three months ended July 26, 2026 was $7.88 billion, about 8.2% of total revenue of $96.221 billion. However, this regional revenue is classified by the headquarters location of the direct customer and does not necessarily match where the product is ultimately used. For example, if a Chinese company rents GPU compute from a cloud provider outside China, it may not be counted as China revenue in NVIDIA's statistics.
In its earnings, NVIDIA guided to third-quarter fiscal 2027 revenue of $108 billion (plus or minus 2%) and said it does not assume any China data center compute revenue. But that outlook does not show the scale of Chinese companies' use of compute through overseas cloud providers either. In other words, NVIDIA's regional revenue alone cannot reveal how many transactions would fall under the reported 51% rule.
Why remote access from overseas is hard to catch under U.S. export controls
Since November 17, 2023, the United States has in principle required licenses to export advanced computing items to companies headquartered in countries in export-control Country Group D:5, which includes China, or in Macau. The Commerce Department's Bureau of Industry and Security (BIS) made clear in its May 31, 2026 guidance that this requirement remains in effect. Even if a covered company is located in a third country, the license requirement applies if its headquarters or ultimate parent is in a D:5 country or Macau.
The BIS guidance also explains that data center operators legitimately doing business under the Export Administration Regulations (EAR) need not stop using, storing, disposing of or maintaining existing equipment because of the guidance, until further notice. In other words, the guidance mainly addresses to whom regulated advanced computing items may be exported. It does not explicitly name as a direct target the arrangement in which a non-Chinese company holds GPUs in a third country and a Chinese company rents that compute over a network.
That said, using this structure does not guarantee export controls can be avoided. Legality depends on the parties, the end use, license conditions and other factors. NVIDIA itself explains in its Form 10-Q that the U.S. government has set conditions on export licenses for certain countries and has already restricted the ability of foreign companies to build large GPU clusters and offer them as services. Existing export controls can also affect transactions routed through overseas clouds.
To clarify rules on such remote use, the Remote Access Security Act is under consideration in Congress. H.R.2683 would, under certain conditions, bring within export controls the act of foreign persons accessing items under U.S. jurisdiction from outside their physical location via network connections such as the internet or cloud services. It contemplates cases where the Commerce Secretary determines there could be a serious risk to national security or foreign policy, and it does not impose a blanket ban on all remote access.
The bill passed the U.S. House 369 to 22 on January 12, 2026, but it has remained referred to the Senate Committee on Banking, Housing, and Urban Affairs since the following day. As of November 3, the U.S. Government Publishing Office (govinfo) record shows no further progress. Until a legal framework directly regulating remote access is in place, a situation has emerged in which NVIDIA is limiting Chinese companies' use through the terms on which it sells to customers.
What matters here is that NVIDIA is not merely selling GPUs but is deeply involved in building out AI cloud providers' infrastructure. In the Form 10-Q, it refers to a new business model with some AI cloud providers, and its maximum exposure on debt guarantees related to land, power and buildings for AI clouds reaches $3.5 billion. NVIDIA, which is involved not only in GPU supply but also in financing data center businesses, is in a position to influence business operations through its transaction terms with customers. If the 51% rule is actually applied, it could narrow the routes by which Chinese companies use overseas compute through such relationships.
How would the 51% rule affect large contracts?
An example of a business that could be affected if the policy is applied is the AI data center business run by Japan's Data Section. On July 10, 2025, the company announced that it had procured GPU servers with 5,000 NVIDIA B200 chips for an AI data center to be opened in Osaka Prefecture, and had indirectly signed a usage contract with one of the world's largest cloud service providers through its business partner NowNow Japan. The contract runs three years, with annual usage fees of $135.342 million and a three-year total of $406.026 million. Converted at the exchange rate at the time (¥144.01 per dollar), the company put the amount at about ¥58.471 billion, and the contract can be extended by two years at the company's discretion. The customer's name has not been disclosed, citing the customer's wishes and confidentiality obligations.
On October 3, 2025, Data Section further announced that it had won the order for its second project, in Sydney, Australia, from the same customer as the first Osaka project through a similar commercial flow. The original plan was to use 10,000 NVIDIA B200 chips, but on December 19 the GPUs were changed to B300 and the contract was re-signed. The revised usage fees are $297 million per year and $891 million over three years, equivalent to about ¥139.115 billion at the exchange rate at the time (¥156.15 per dollar). Separately, the contract provides for a $74 million setup fee for building the customer's dedicated GPU cluster.
Combining the three-year usage fees for the two projects gives $1.297026 billion. The breakdown is $406.026 million for the first project in Osaka and $891 million for the second in Sydney, excluding setup fees and fees for the extension period. The counterparty for both contracts is the same undisclosed single company.
In December 2025, the FT reported that China's Tencent was using Data Section's compute through a third party. However, Data Section has not disclosed the name of its counterparty, and it has not been independently confirmed that the customer in the disclosed contracts is the same as Tencent. It is also unclear whether Data Section itself counts as a "cloud customer" in NVIDIA's terms.
It therefore cannot be said that the 51% rule applies directly to Data Section's business. Still, for the two disclosed contracts alone, the counterparty is a single company, accounting for 100% of the customer mix by contract value. If that customer is a Chinese company and the 51% rule is applied contract by contract, similar new contracts could run afoul of a condition holding China-bound services to 49% or less. If the ratio is calculated across a provider's entire business, deals with other customers would also have to be taken into account.
The information Data Section has published does not reveal the share of its overall business that comes from Chinese companies. For this reason, the existing large contracts cannot simply be run through the 51% rule to conclude that they violate the condition. Even so, the more a provider depends on contracts leasing dedicated clusters to a single Chinese company, the more it would need to increase contracts with other customers to bring its share of non-Chinese sales to 51% or more. If the test is applied per contract, the constraint could be even greater.
Since the FT report says existing contracts are exempt from the new policy, the two contracts Data Section signed in 2025 would not be immediately affected. However, it is unclear whether a two-year extension or an expansion of a GPU cluster would be treated as a "future contract." Assessing the policy's impact requires confirming the scope of covered contracts and the basis for calculating the ratio.
Can a verbal sales condition be effective? The November 17 earnings are the focus
The report leaves an important question about the policy's effectiveness. According to the FT, NVIDIA's request does not apply to existing contracts, and notice to customers has been only verbal. It is also unclear whether NVIDIA would stop GPU supply or change contract terms if a cloud provider failed to meet the requested ratio. Moreover, NVIDIA itself has not officially confirmed the reported 51% rule, and its statement to AFP stressed the commercial need to keep U.S.-made AI infrastructure internationally competitive.
Another question is whether the providers hit hardest would be those that lease dedicated GPU clusters to specific Chinese companies. According to the FT, a Singapore-based cloud provider explained that it is difficult to secure enough non-Chinese customers outside the United States. This suggests that not only providers dependent on large contracts with specific Chinese companies but also cloud providers with many customers could see a high share of Chinese usage. The policy's impact therefore cannot be limited to dedicated-cluster providers.
Still, there is some basis for the view that NVIDIA's sales terms would work to restrict Chinese companies' access to overseas compute. If the 51% rule is applied to future contracts, cloud providers would have to consider their contract ratio with non-Chinese companies when increasing business with Chinese firms. However, it is impossible at this point to estimate how much compute would actually become unavailable to Chinese companies. The size of the impact will depend on how the ratio is calculated, the structure of contracts, and how NVIDIA applies the condition.
There are three developments to watch in testing this view. The first is whether BIS publishes a new export control rule on remote access in the Federal Register, or whether the Remote Access Security Act becomes law. If legal regulation of remote access is put in place, the current arrangement, in which NVIDIA's own sales terms supplement regulation, could change. However, enactment of the bill alone would not automatically ban all remote access, and the specific targets and conditions of application would need to be checked.
The second is whether providers such as Data Section announce new or additional large contracts with Chinese companies after November 3. In particular, a new contract leasing a dedicated GPU cluster to a specific Chinese company would offer a clue to how the 51% rule is actually being applied. But the mere announcement of such a contract would not be enough to conclude that the rule is not being applied. Without clarity on which providers are covered and how the ratio is calculated, consistency with any contract cannot be judged.
The third is an official explanation from NVIDIA itself. The company will hold its conference call for third-quarter fiscal 2027 results at 2 p.m. Pacific time on November 17, with a Form 10-Q filing to follow. The question is whether it will deny the report in the call or disclosures, or formally explain the policy as a sales condition for cloud customers. Even if the policy's existence is confirmed, it will be hard to judge its actual impact unless its enforceability and specific scope are shown.
U.S. export controls have been built mainly around which companies may receive advanced AI chips, while how to handle remote use through overseas data centers is still under debate. If NVIDIA has begun restricting Chinese companies' use through its sales terms in the meantime, it can be seen as a new move in which a GPU maker effectively supplements export controls. But how well such a mechanism works is unconfirmed, and what NVIDIA says on its November 17 earnings call will be an important input for judging it.