Nvidia Tells Cloud Providers to Keep Chinese Customers Below Half of New Business, FT Reports

Nvidia has told data center and cloud operators that run its chips to keep Chinese customers to at most 49% of their business on new contracts, the Financial Times reported on November 3, 2026, according to Tom's Hardware. The guidance applies to operators worldwide, not only in Asia, and was passed on by Nvidia's sales force and regional teams rather than in a written policy (FT via Tom's Hardware).

The reported cap is Nvidia's latest move to control where its AI chips and data center capacity end up, this time by limiting who its customers may serve rather than who may buy. Nvidia has not confirmed the 49% figure. It responded with a general statement about US policy, and the account rests on unnamed people familiar with the matter.

What exactly is Nvidia reported to have told cloud providers?

Nvidia told data center operators that sell computing services to Chinese companies to track what share of their business those clients make up and to keep it at 49% or less, so that at least 51% comes from non Chinese customers, according to the Financial Times (FT via Tom's Hardware). The rule covers new and future contracts only, not hardware leasing agreements already in place.

The Information reported a softer version on the same day. Nvidia told some cloud customers, including neoclouds it has backed, to take on less business from Chinese companies so that their customer mix becomes more diverse, citing three people familiar with the matter (The Information via The Decoder). A neocloud is a cloud company that mainly rents out GPU capacity for AI work.

The two accounts differ on one point. The FT headline says Nvidia "orders" its data center clients to reduce China exposure, while The Information reported that the guidance is not mandatory (The Information via The Decoder). The Information also reported that cloud customers are being asked to disclose the identity of the end users who access the chips.

Why would Nvidia limit its customers' customers?

The reported cap aims at a gap in US export controls: the rules restrict shipments of advanced chips to China but generally do not cover remote access through cloud services, The Decoder explains. That lets Chinese AI developers rent capacity on advanced Nvidia chips in data centers abroad without importing any hardware.

That route has drawn official attention in 2026. On July 23, 2026, White House science and technology adviser Michael Kratsios said in a post on X that Moonshot AI, the Chinese developer of the Kimi models, had acquired servers with Nvidia GB300 chips and accessed GB300s in Thailand, without publishing evidence, according to The Next Web. US authorities had been looking into Chinese AI companies reaching such chips through foreign data centers after those allegations, according to The Decoder.

Nvidia has tightened its own checks before. In July 2026, the FT reported that Nvidia had more than halved the number of Asian companies allowed to buy its AI chips directly, after tougher compliance reviews in Singapore, Malaysia and Japan that hit neoclouds hardest (FT via Techzine). The July list decided who may buy; the November guidance, as reported, reaches one step further down, to whom those buyers rent the chips.

Regional clouds are under scrutiny for other reasons too. The New York Times reported on October 29, 2026 that US intelligence watched an executive of Nvidia cloud customer Megaspeed, a company that has faced chip diversion allegations, which Nvidia says were investigated and debunked.

How did Nvidia respond?

Nvidia did not confirm or deny the 49% figure in the statement quoted by Tom's Hardware, and instead described US policy. The company said that "American chips are the foundation that American model and application makers need to compete worldwide," and that current US policy "encourages" cloud providers to build on American infrastructure and to accept business from "everywhere they can" (Tom's Hardware).

That second phrase sits awkwardly next to the reported guidance, which asks the same providers to turn some business away. Neither open report says whether the cap came at the request of the US government.

Why is rented compute hard to measure?

Most public estimates of China's AI compute count chips that physically arrive, which is the part export rules already police. Epoch AI's chip imports explorer, published on October 27, 2026, estimates from customs records that about 10% of AI compute delivered from 2022 through June 2026 went to China, against 72% for the United States, and notes that cloud companies rarely disclose where their chips are deployed (Epoch AI).

Put next to The Decoder's point that export rules generally do not cover remote access, that means Chinese use of Nvidia chips rented abroad would not show up in import based counts like Epoch's. The reported 49% cap is aimed at exactly that share, which neither Nvidia nor any of these sources has quantified.

What we don't know yet

• Whether the 49% limit is a firm condition or a recommendation, since the FT and The Information describe it differently.
• How Nvidia would check compliance, and what happens to a provider that goes above the cap.
• Whether the US Commerce Department asked for the guidance or plans its own rules on remote access.
• Which providers received it, and how much of their current business comes from Chinese customers.