Nvidia orders data centre clients to reduce China exposure
Nvidia has told data centre and cloud customers in recent months that they must sell at least 51 per cent of their services to non-Chinese companies, according to five people with knowledge of the matter; there were no such restrictions before. The instructions were given verbally by Nvidia's sales force and regional teams rather than in documents, apply to clients globally and not just in Asia (per two of the people), and cover only future contracts, not existing leases. The move tightens the squeeze on Chinese tech groups that have relied on leasing overseas cloud capacity to train models, as Washington tries to clamp down on both chip-smuggling intermediaries and offshore leasing. Earlier this year Nvidia created a 'white list' and more than halved the number of Asian customers authorised to buy its AI chips. Scarcity shows in prices: a B300 rack sells for about $600,000 internationally but fetched US$2.54M in China last month, per a person with direct knowledge. A Singapore-based group said it is 'difficult to find enough non-Chinese clients outside the US' and 'if strictly implemented, it will definitely hurt business'; data centres that cannot find non-Chinese customers could be forced to shrink. Nvidia did not confirm or deny the rule, saying in a statement that data centres and clouds built on American chips are the foundation American model and application makers need to compete worldwide. Context: Nvidia shares are down more than 13% in the past month (up 7% this year) and Apple has overtaken it as the most valuable listed company. Caveats: the story rests on anonymous sources, and it does not say how the 51% is measured (revenue, capacity or customer count), which providers are affected, or whether the US government asked for it.
Why it matters
Nvidia is effectively acting as a private export regulator for the global neocloud market. A 51% non-Chinese customer floor changes the business model of Asian and other non-US GPU clouds, many of them venture- or PE-backed, and cuts off a channel Chinese AI labs used to train models offshore.
Executive takeaways
What happened (five sources): Nvidia verbally told cloud and data centre customers that at least 51% of their services must go to non-Chinese companies; new contracts only, applies globally. | • Policy backdrop: US clampdown on chip smuggling and offshore leasing; Nvidia already more than halved its authorised Asian buyers via a 2026 'white list'. | • Scarcity signal: B300 rack ~$600,000 abroad vs Rmb18.5mn ($2.8mn) in China last month. | • Who is hit: GPU clouds outside the US, especially in Asia, that rely on Chinese demand; one Singapore operator says it 'will definitely hurt business' and some may have to shrink. | • Nvidia's line: neither confirms nor denies; frames it as building American-chip infrastructure worldwide. | • Caveat: anonymous sourcing, verbal instructions, no definition of how 51% is measured or enforced.
What Financial Times may be missing
The piece does not explain how the 51% is calculated (revenue, compute hours or customer count) or how Nvidia would audit it, nor whether Chinese-owned subsidiaries abroad or resellers count as Chinese. It does not say whether the US Commerce Department requested the rule or whether Nvidia is pre-empting regulation, and gives no estimate of how much Nvidia revenue or which named neoclouds are exposed. Chinese labs that lease overseas capacity are not named and there is no response from them or from Beijing. It also leaves out the second-order effects: whether capacity freed up abroad lowers GPU rental prices for non-Chinese startups, and whether the rule pushes Chinese demand further towards Huawei and domestic chips or towards smuggling.
