Southeast Asian data centers, whose biggest clients are Chinese tech groups, must win non-Chinese customers or risk shrinking, and Chinese labs lose a main route to Nvidia training capacity.
Nvidia has told its cloud computing customers that they must sell at least 51% of their services to companies outside China, the Financial Times reported, citing five people with knowledge of the matter. Nvidia had set no such limit before. Its sales force and regional teams passed the instruction on verbally rather than in writing, and it applies to Nvidia’s clients globally, not only to data centers in Asia, the FT reported. It covers future contracts only, not existing leasing agreements.
Nvidia brought in the requirement over the summer, two people familiar with the matter told the FT, as a pre-emptive step to blunt the criticism it routinely receives from US lawmakers over its approach to China. The rule caps Chinese customers’ share of each partner’s business, which in practice reduces how much Nvidia-based capacity Chinese companies can rent from overseas providers.
Nvidia neither confirmed nor denied the report to AFP. Several details are missing from every account, the FT’s included: how the 51% is measured, whether by revenue, capacity or number of customers; how Nvidia defines a Chinese company; and what happens to a partner that falls below the threshold.
The rule moves Nvidia’s controls from buyers to the buyers’ customers
Many Chinese groups have leased overseas cloud capacity for advanced AI work, including model training, and Chinese tech groups have been the biggest clients of data centers in markets such as Southeast Asia, the FT reported. US export controls bar the sale of Nvidia’s most capable chips to China, so renting time on servers in other countries has been one of the few legal ways for Chinese developers to reach them. EastFrontier has reported on this pattern, including a report that Chinese AI firms tap Southeast Asian Nvidia clouds.
The leasing route grew partly because of Nvidia’s earlier checks. This year the company more than halved the number of Asian customers authorized to buy its AI chips, keeping a whitelist of those that passed tougher compliance checks. Fewer advanced chips then reached China through smuggling, and more Chinese groups turned to overseas cloud providers instead, according to the FT. EastFrontier reported on that whitelist in July, when H200 shipments to China were confirmed as the US halved Nvidia’s Asian buyer whitelist. The new instruction goes one layer further down the chain. Instead of vetting the data center that buys the chips, it sets a condition on the customers that data center serves.
Scarcity already shows in prices. In China, a rack of Nvidia’s B300 chips fetched RMB18.5 million (US$2.8 million) last month, against about US$600,000 internationally, the FT reported. The figure came from a person with direct knowledge of the deal, and the B300 cannot legally be sold in China under US export controls. Smuggling remains part of the picture, as Yahoo Finance noted. In March, US prosecutors charged three people tied to Super Micro, the assembler of servers built on Nvidia chips, among them a company co-founder, over the alleged diversion of about US$2.5 billion of US AI technology to China. Super Micro itself was not charged, and a contractor in the case pleaded guilty in October, Reuters reported.
Nvidia framed the rule as building on American chips, and Beijing called for fair treatment
Nvidia’s statement did not address the 51% figure. Clouds built on American chips are "the foundation that American model and application makers need" to compete worldwide, the company said. It added that US policy has long encouraged clouds everywhere to build on American infrastructure, and that building it out abroad brings prosperity, influence and jobs back to the US.
China’s response was measured. Foreign ministry spokeswoman Mao Ning said she was not aware of the report. Speaking in principle, she said China hopes all parties will provide "a fair, transparent and non-discriminatory business environment" for companies from all countries, including Chinese companies, according to AFP via Malay Mail.
The statement places the decision inside Nvidia’s commercial argument that American chips should power clouds globally. A threshold enforced through sales teams also leaves Nvidia room to apply it partner by partner, without the formal process that a government export rule would require.
Cloud operators’ next contracts will show how strictly the threshold bites
The immediate pressure falls on operators whose customer lists lean heavily toward Chinese clients, and some are already struggling. A Singapore-based cloud group told the FT that enough non-Chinese customers are hard to find outside the US, and that strict enforcement would hurt its business. Data centers that cannot win non-Chinese clients could be forced to shrink their operations, the FT reported.
For Chinese AI developers, the effect depends on how quickly they can replace rented overseas capacity with domestic chips or with Nvidia products cleared for sale in China. Model training is directly exposed, since it is one of the main uses Chinese firms have rented overseas capacity for, according to Investing.com.
The first concrete evidence will come from the cloud providers themselves. Watch for operators in Southeast Asia to disclose changes to their customer base, renegotiated leases or slower capacity additions in their next quarterly reports, and for any written version of the rule that defines how the 51% is counted.