Nvidia's 51% rule: foreign clouds must cut China ties to keep AI chips

What Nvidia's 51% rule actually says

Foreign cloud providers that want Nvidia's advanced AI processors must now show that at least 51% of their revenue comes from customers outside China.

The Financial Times, citing people familiar with the discussions, reports that Nvidia passed the requirement down verbally through sales representatives and regional management teams rather than in a public policy document. It took effect over the summer and applies to new commercial agreements, leaving existing cloud contracts untouched for now.

The scope is global. Two sources told the FT the condition covers Nvidia customers worldwide, not only data centers operating across Asia. In practice, a cloud operator in Singapore, Frankfurt or Virginia that built a big share of its business on Chinese AI companies now has to replace that revenue with international customers before it can keep ordering.

Nvidia's public statement stays on message: "Data centres and clouds built on American chips are the foundation that American model and application makers need to compete worldwide." The company did not confirm the 51% figure directly.

Why a $2.8 million chip rack explains everything

The gap between the international price and the Chinese street price is the whole story.

A complete Nvidia B300 rack sells for roughly $600,000 internationally. In China, one reportedly changed hands for about $2.8 million in October 2026. That is a markup of more than four times, and it exists for one reason: US export controls bar the direct sale of the B300 into China.

When the legal price is $600,000 and the illegal one is $2.8 million, every intermediary, foreign cloud and shell customer has a strong incentive to look the other way. That is precisely what Washington and Nvidia are trying to squeeze out. Nvidia has already cut the number of Asian companies authorised to buy its AI chips by more than half this year and added stricter compliance checks.

Chinese AI labs have leaned on overseas cloud providers to rent the compute they cannot buy. Training a frontier model needs tens of thousands of GPUs, and domestic alternatives remain constrained. Closing the rental route hits them where it hurts most.

The clouds caught in the middle

For data center operators that built their business around Chinese clients, 51% is not a rounding error.

A representative of a Singapore-based cloud provider told the FT the requirement is painful: "If strictly implemented, it will definitely hurt business." Finding enough non-Chinese, non-American customers to clear the threshold is the hard part, especially for smaller regional operators that cannot compete for US enterprise contracts.

There is an irony here. Washington wants American chips to power the world's AI infrastructure, and Nvidia says as much. But a rule that forces clouds to drop Chinese customers also makes those clouds less competitive on price and scale, which is exactly the opening domestic Chinese chipmakers need.

The policy is not a formal export control, so enforcement depends on Nvidia's own sales channel. That makes it flexible and deniable, but also uneven. A large US hyperscaler will barely notice; a 20-rack operator in Southeast Asia may not survive the transition.

What happens next

Expect three things. First, more consolidation among regional cloud providers as smaller players sell or pivot. Second, a fresh round of workarounds, from re-invoicing to joint ventures structured to keep Chinese revenue below the line. Third, louder complaints from Asian governments that US export policy is being enforced by a private company rather than through treaties or trade rules.

Nvidia has not published the rule, so its exact wording and audit process remain unclear. The FT's reporting is based on people familiar with the discussions, and Nvidia has not confirmed the 51% threshold on the record. Treat the number as well-sourced but unofficial until the company says otherwise.

What it means for US, UK and EU cloud buyers

If you run workloads on a US or European cloud, the 51% rule is unlikely to change your bill directly. Hyperscalers already draw the vast majority of revenue from non-Chinese customers. The effect is more subtle: capacity that once went to Chinese tenants may free up, and regional providers competing for the same GPUs could become more aggressive on price.

For companies sourcing AI compute through smaller regional clouds in Asia, check your provider's customer mix. A supplier that suddenly has to replace a large Chinese client may reprioritise contracts, raise prices or exit the market altogether.

Nvidia is trying to police the global cloud market with a sales rule rather than a law, and the $2.8 million Chinese price tag shows how much money is on the other side of that line. Whether the 51% threshold holds will depend less on Nvidia's paperwork than on how many regional clouds can afford to lose their Chinese customers.