Policy/ nvidia · export-controls · china · cloud-computing

Nvidia Reportedly Limits Chinese Cloud GPU Rentals Abroad

Nvidia has reportedly told cloud partners that 51% of services must go to non-Chinese firms, tightening the rental route Chinese AI labs lean on.

Nvidia has reportedly told cloud partners worldwide that at least 51% of their services must go to non-Chinese customers, according to the Financial Times. Nvidia has neither confirmed nor denied the rule.

The instructions reportedly came verbally, through sales and regional teams, and apply globally rather than only in Asia. Existing deals and leases stay as they are; new ones must comply. Southeast Asian providers that want to become Nvidia Cloud Partners now need Fleet Intelligence tracking, and Nvidia's whitelist of authorized Asian buyers has more than halved.

Prices show the squeeze. A B300 rack that sells for about $600,000 internationally reportedly fetched $2.8 million in China in October. An eight-card H200 server cost over $900,000 there at the end of that month, about 3x the price a year earlier. One source told CNBC that Moonshot was asked to pay 30% to 50% more upfront than a non-Chinese customer.

Renting became the favored route because U.S. law gives regulators little authority over overseas cloud access. The Remote Access Security Act would change that. It has passed the House, but the Senate has not scheduled a vote. Neoclouds (smaller, GPU-focused cloud providers) and their financiers reportedly price in a possible ban anyway.

Buying is not much easier. Chinese customers have ordered more than 6 million export-compliant RTX Pro 5500 cards, against the 3 million Nvidia could reportedly make in 2027. Domestic chips fall short too: supply this year reaches up to 3.5 million, against demand above 5 million. Many Chinese firms still prefer CUDA, Nvidia's programming platform, to Huawei's CANN.

The pattern suggests that when every legitimate route tightens at once, demand looks for the next one. Moonshot has reportedly discussed renting clusters of 8,000 or more Blackwell GPUs in Europe for inference, meaning serving models to users. Nvidia's rule would also ease criticism from U.S. lawmakers, which may be part of the point.

The first checks come soon. Nvidia reports third-quarter results on November 17, and Jensen Huang is due at a closed-door briefing for lawmakers a day later. Huang has pushed for laxer controls and more access, so the reported 51% rule will face scrutiny from the very people he has been lobbying.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →