US Commerce Department closes cloud loophole that let Chinese firms rent AI chips

57 minutes ago 24

A new interim final rule targets remote access to advanced chips abroad after the Nvidia-Megaspeed controversy

For years, US export controls worked like a bouncer who checked IDs at the front door and ignored the side entrance. Chinese companies couldn’t buy top-tier AI chips outright, but they could rent processing power from data centers sitting outside China.

On Monday, the US Commerce Department moved to shut that side door. The rule arrived largely in response to the Nvidia-Megaspeed saga, which put the gap in an uncomfortable spotlight.

What the new rule actually does

The Commerce Department’s Bureau of Industry and Security (BIS) issued the measure as an interim final rule on November 16, 2026. It is expected to be published soon in the Federal Register.

The core target is remote rental of advanced semiconductors. Until now, the controls focused on where physical chips were shipped, not on who was logging in to use them.

Under the new framework, data centers must obtain a license if they provide more than 12 million aggregate total processing performance to entities based in China, Macau, or certain other restricted jurisdictions. That threshold works out to approximately 760 Nvidia H100 GPUs.

Providers already serving these customers get a 90-day window to come into compliance.

The rule also leans heavily on due diligence. Data centers will need know-your-customer protocols, plus monthly reporting on licensed transactions.

Geography is part of the package too. Five countries—Indonesia, Malaysia, Singapore, Thailand, and Brazil—will now require licenses for exports of advanced AI chips because they are seen as carrying elevated risks of diversion.

The stated goal is to bring remote-access controls in line with the existing bans on physical exports.

How Megaspeed forced the issue

The trigger was a controversy involving Megaspeed, a Singapore-based cloud provider. The company was under investigation for alleged diversion or circumvention practices involving billions of dollars in Nvidia hardware.

That episode exposed how porous the system had become. Chips could legally land in a third country, sit in a data center there, and still end up powering work for Chinese customers through a network connection.

Southeast Asia’s appearance on the license list is notable. Singapore, Malaysia, Thailand, and Indonesia have become popular destinations for data center buildouts, and the rule now treats that growth as a potential leak rather than purely a commercial win.

Brazil’s inclusion signals that BIS is not only thinking about China’s immediate neighbors.

What this means for AI labs, chipmakers, and cloud providers

The most direct effect lands on data center operators in the five named countries. They now face licensing hurdles, customer vetting, and recurring reports, all of which add cost and slow down sales to certain clients.

For operators whose business models depended on Chinese demand, the math may change quickly. A 90-day runway is not much time to replace a major customer base.

Chipmakers are exposed as well. Companies with significant international trade exposure, Nvidia among them, may see market volatility as these controls take hold.

The 12 million total processing performance line is a clear target for those seeking to structure activity around the threshold, including through smaller distributed rentals or new intermediaries in countries not yet on the list.

Several things are worth watching in the weeks ahead. The first is the Federal Register publication, which will lock in the final text and any changes from the interim version. The second is how data center operators in Singapore, Malaysia, and the other named countries respond during the 90-day compliance window. The third is enforcement, given that know-your-customer rules and monthly reports only work if someone reads them and acts on what they find—and the Megaspeed case showed how much can slip past before anyone notices.

Disclosure: This article was edited by John Chen. For more information on how we create and review content, see our Editorial Policy.

Read Entire Article