Nvidia has told the cloud and data center operators that run its AI chips that at least 51% of their sales must go to non-Chinese companies, the Financial Times reported on November 3, citing five people with knowledge of the matter. The requirement applies to future contracts only, and was passed on verbally by Nvidia's sales and regional teams rather than written into agreements, according to accounts of the FT report by Techzine and Analytics Insight.
Why it matters: Nvidia's most advanced chips can't be shipped to China under US export controls, so Chinese AI developers have increasingly rented capacity from data centers abroad, particularly in Southeast Asia, to train models. A customer-mix rule from the chip supplier itself would squeeze that route without any new government regulation. Before this, Nvidia had no rule about where its cloud partners' customers came from, Techzine reports.
What the reported rule says
| Detail | As reported |
|---|---|
| Threshold | At least 51% of a provider's services sold to non-Chinese customers |
| Contracts covered | New contracts only; existing leases are not affected |
| Geography | All Nvidia cloud customers worldwide, not just Asia (per two of the FT's sources, via Techzine) |
| How it was communicated | Verbally, through Nvidia sales and regional teams |
| Penalties | None described in the reports we read |
Nvidia neither confirmed nor denied the report to AFP, according to The Edge Malaysia. In a statement, it said clouds built on American chips are what US model and application makers need to compete, and that long-standing US policy encourages clouds everywhere to build on American infrastructure. Its line: "America cannot win the AI race unless America's AI infrastructure wins worldwide."
China's foreign ministry spokeswoman Mao Ning said she was not aware of the report, and that Beijing hopes all parties offer a fair and non-discriminatory business environment to companies from every country, The Edge Malaysia reports via AFP.
Why Southeast Asian clouds are worried
Chinese tech companies are among the largest buyers of AI capacity in Southeast Asia, so operators there would need to find many more non-Chinese customers to stay compliant. A Singapore-based group told the FT it is hard to find enough non-Chinese demand outside the US, and a Singapore provider said strict enforcement could hurt its business, according to Techzine and Analytics Insight.
The incentive to route around the controls is large. One insider told the FT that a rack of Nvidia B300 chips, worth about $600,000 on the international market, recently sold in China for 18.5 million yuan, roughly $2.8 million, according to Analytics Insight. That is more than four times the price.
Part of a wider tightening
The reported rule follows other moves by Nvidia this year to control where its hardware ends up. It has cut its list of authorized Asian AI chip buyers by more than half and introduced a stricter approval list meant to stop advanced processors reaching China through unauthorized channels, Analytics Insight reports, citing the FT.
There are also limits to what a 51% rule can do. It leaves existing contracts alone, and it would still let up to 49% of a provider's new business go to Chinese customers. The reports we read don't say how Nvidia would check compliance or what happens to a provider that misses the threshold.
Nvidia shares had fallen more than 13% over the past month but were up about 7% for the year, according to Analytics Insight.
What to watch
- Confirmation in writing. A verbal policy is easy to soften. Watch for it to show up in contract terms, Nvidia's next earnings call, or partner disclosures.
- Enforcement. Whether Nvidia audits customer mix, and whether it cuts allocations to providers that miss the bar.
- Where Chinese demand goes next. If rented compute abroad gets harder to buy, domestic accelerators such as Huawei's Ascend line become more important to Chinese labs.
- Who's exposed. Cloud operators in Singapore and elsewhere in Southeast Asia with large Chinese customer bases are the obvious pressure points.
This story is based on secondary reports of a Financial Times article. Nvidia has not confirmed the rule.
