The United States has moved to close the route that let Chinese AI companies rent American computing power instead of buying the chips outright. Nvidia and other makers of advanced processors have long been barred from selling their best data centre chips to China without a licence, but nothing stopped a Chinese lab from renting that same capacity from a cloud operator in a third country. The Commerce Department interim final rule issued on Monday extends the control from the shipment to the use, which is the part of the trade that had stayed outside the system.
How the threshold is drawn
The trigger is scale rather than intent alone. A licence is required when an exporter knows a data centre will provide more than 12 million aggregate TPP to a firm headquartered in a restricted destination, or to a firm whose ultimate parent sits in one. Capacity counts across linked data centres, including sites joined only digitally, so splitting a cluster across buildings does not take it below the line. Intermediaries are covered too: a customer outside China that trains models for, or supplies models to, a Chinese company can bring a transaction into scope, and data centres are expected to check customers for direct or indirect Chinese ownership.
That design matters for Southeast Asia, where much of the rented AI capacity serving Chinese demand is physically located. By adding destination licences for Singapore, Malaysia and Indonesia, alongside Thailand and Brazil, Washington is treating the location of the server as a risk in its own right, not only the nationality of the end user. Companies headquartered in the United States are exempt from the destination licence if exporters meet reporting requirements on each deal, which leaves American cloud builders a route to keep expanding in the region while other buyers face case by case review.
Timing next to Nvidia's results
The rule lands one day before Nvidia reports fiscal third quarter results on Tuesday, its first earnings release since Jensen Huang resigned as chief executive last week and Colette Kress stepped in on an interim basis. Analysts expect earnings of about $2.47 a share on revenue of about $109 billion, but the policy change speaks to a different question than the quarter itself: how much of the demand sitting outside China is genuinely foreign, and how much has been Chinese demand wearing a foreign address. If a meaningful share of Southeast Asian data centre orders falls into the second category, the licence process becomes a direct variable in Nvidia's forward orders, not a background compliance cost.
The 90 day window for existing operators is the first test of how hard the rule will bite in practice. A licensing regime that approves most applications quickly would mainly add paperwork and customer checks. One that moves slowly, or refuses licences for centres with large Chinese client bases, would strand capacity that has already been built and paid for, and push Chinese labs toward domestic chips sooner than their performance would otherwise require.