Nvidia has asked its data center partners around the world to limit business from Chinese companies, the Financial Times reported on November 3, citing five people familiar with the matter. Chinese customers should account for no more than 49%, with at least 51% of services sold to non-Chinese firms. The requirement applies only to new and future contracts; hardware rental agreements already in force are not affected.
According to the report, Nvidia had not previously imposed any such ratio on its partners. The requirement was relayed verbally by sales and regional teams, with nothing in writing. In a statement, Nvidia did not confirm the 49% figure and spoke only about principles: U.S. policy encourages cloud providers to build on American infrastructure and to take business wherever they are able to do so.
"America cannot win the AI race unless America's AI infrastructure wins worldwide."
The gaps a verbal rule leaves behind
The outlets that picked up the FT story do not describe the rule in quite the same terms. Tom's Hardware says operators must track the makeup of their revenue, while Investing.com and The Next Web write about the share of "services." The two measures could produce very different results. As a hypothetical: a facility hands 40% of its GPUs to a large Chinese customer. If that customer's contract pays a higher unit price, the operator could be under the line by compute and over it by revenue.
How compliance would be checked, and whether going over the limit would affect an operator's chip supply, is likewise not covered in the reports. With no document, operators have no text to measure themselves against and can only gauge the threshold through conversations with Nvidia's sales staff.
The reach extends beyond Asia
The requirement applies to data centers worldwide, not just those in Asia. The route behind it is a familiar one: many Chinese AI companies do not buy chips outright but rent compute from overseas cloud providers to train their models. Washington has been tightening sales of advanced chips to China since 2022, and has long wanted to close off cloud rental as well.
Pressure on both sides has grown this year. On the U.S. side, several Super Micro employees were charged earlier this year with smuggling roughly $2.5 billion worth of Nvidia AI servers into China; Tom's Hardware also notes that five-year-old A100 servers now sell in China for about three times their original price. On the Chinese side, Beijing banned state-funded data centers from using foreign AI chips a year ago while backing domestic suppliers. Chinese foreign ministry spokesperson Mao Ning said she hoped all parties would provide a "fair, transparent and non-discriminatory" business environment.
What it means for Chinese teams
Existing leases stay as they are, so the near-term impact falls mainly on renewals and expansions, which is where Chinese teams are most likely to run into the requirement first.
The pressure does not land only on tenants. The 49% is a ratio imposed on operators: a facility with a high share of Chinese customers has to find enough non-Chinese ones before it can take on new contracts. The report quotes a Singapore-based provider as saying it is hard to find enough non-Chinese demand outside the United States. That could leave Southeast Asian providers that rely mainly on Chinese orders choosing between broadening their non-Chinese customer base and slowing down new business.
The point that would affect the scope most has no answer yet: whether the Singapore or Malaysia subsidiaries of Chinese companies count as "Chinese companies." If only the place of registration matters, the requirement bites far less; if it looks through to the parent, many more contracts would be affected. Nvidia has not publicly defined the term, and the FT's account offers no answer.
Sources: Financial Times, Tom's Hardware, CocoLoop, The Next Web, Investing.com. The 49% and 51% ratios, the scope and the manner of communication are as described by sources cited in the reports; Nvidia has not confirmed them.