If China can't buy America's best AI chips, it has been renting them instead. But Washington has now shut that door.
Jefferies estimates the move could take $44 billion off Nvidia's 2027 data centre chip revenue.
A new US Commerce Department rule, introduced this month, targets foreign data centres that lease computing capacity to Chinese companies.
Closing the remote-access route
Until now, Chinese companies and AI labs could use advanced US chips without importing them, by renting capacity at data centres overseas.
Under the new rule, overseas data centres need a US licence if they supply "significant volumes" of computing power to firms based in China, Macau, Russia, Iran, Cuba or other restricted countries.
The threshold is 12 million aggregate Total Processing Performance, a US measure of a chip's computing power.
That is roughly equivalent to 750 Nvidia H100 or H200 GPUs, the chips that power most large AI systems.
Singapore, Malaysia, Thailand, Indonesia and Brazil now face licensing requirements for the first time, because Washington sees them as higher-risk transit points.
Existing operators have 90 days to get approval.
How big is the hit?
Jefferies estimates that about 1 gigawatt of planned data centre capacity is at risk. UBS puts the figure as high as 2 gigawatts.
Jefferies also noted that the diversion of Nvidia chips to China had already fallen sharply over the past six months.
A delay rather than a loss
Analysts say Nvidia could send the affected chips to other customers instead.
That makes the rule more of a long-term question about demand and allocation than an immediate loss of sales.
Even so, the timing is awkward for a company already dealing with the sudden resignation of its chief executive, Jensen Huang, and reporting quarterly results tonight.
Read more about Chipmakers