AI Policy

The US Now Polices Rented Nvidia Chips Too

The US Commerce Department is reviewing whether Chinese AI firms bypass chip export limits by renting Nvidia hardware abroad. What EU and UK cloud GPU buyers should check now.

AI PolicyBy Servola Tech Desk2026-08-084 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

The US Now Polices Rented Nvidia Chips Too

Key takeaways

The US Commerce Department's Bureau of Industry and Security has opened a formal review into whether Chinese AI firms are reaching Nvidia's export-restricted Blackwell-generation chips by renting server capacity in a third country such as Thailand rather than importing hardware directly, and any EU or UK company buying discounted cloud GPU capacity through a reseller now has a reason to ask exactly where that hardware physically sits.
The trigger was Moonshot AI's Kimi K3, a 2.8 trillion parameter model that matched frontier performance from US labs within days of release; White House OSTP director Michael Kratsios publicly accused the firm on 22 July of acquiring Nvidia GB300 servers and running workloads through Thailand-based infrastructure.
Existing export rules were written around a physical shipment: who imports a chip, and where it lands. They were never written around a rented hour of compute on a chip that never crosses a border, which is exactly the gap this review is trying to close.
Treasury Secretary Scott Bessent said sanctions and Entity List placement 'will be on the table' for firms found to have used offshore rentals to reach controlled hardware, and separate legislation already moving through Congress aims to extend export-control language to remote and cloud access. Nothing here is finalized law yet; this is a review and a bill in progress, not an enacted rule.

The investigation Washington just widened

The Bureau of Industry and Security, the Commerce Department division that enforces US chip export rules, has opened a formal review into how Chinese AI companies are obtaining access to Nvidia's most advanced processors without buying them outright. The trigger was Moonshot AI's release of Kimi K3, a 2.8 trillion parameter model whose performance drew immediate comparisons to frontier systems from Anthropic and OpenAI. Days later, on 22 July, White House Office of Science and Technology Policy director Michael Kratsios said publicly that Moonshot had acquired Nvidia GB300 servers, part of the export-restricted Blackwell generation, and had run training workloads through infrastructure based in Thailand rather than importing the chips into China.

Reporting on the review, drawing on people familiar with the matter, describes the Commerce Department as now formally investigating whether Chinese firms reached export-controlled American AI chips this way as a matter of course, not an isolated case. Treasury Secretary Scott Bessent has said sanctions and placement on the Entity List, the US blacklist that cuts a company off from American suppliers, 'will be on the table' for firms confirmed to have used the practice at scale.

Renting was the workaround, not the ban

Chip export controls were built around a physical transaction: a chip ships from a US or allied factory, crosses a border, and lands in a buyer's data center. Every rule Washington has written for five years has tightened that chokepoint, restricting which chips can be sold to which buyers and requiring export licenses for the most capable hardware. What those rules did not clearly reach is a Chinese firm renting compute by the hour from a data center in Thailand, Malaysia, or another third country that never itself imported the chip in violation of anything. The hardware stays put; only the workload crosses a border, invisibly, over a network connection.

That gap is exactly what a bill already moving through Congress, the Remote Access Security Act, is designed to close, by extending export-control obligations to remote and cloud access rather than physical shipment alone. The bill and the BIS review point at the same problem from two directions, legislative and enforcement, which is itself a signal that Washington now treats offshore rental as the live edge of the export-control fight, not a footnote to it.

Why a European compute buyer should care

Most owners will read this as a US-China story with no European angle. It has one, and it is not about which AI model you use. For five years, an EU or UK company doing AI vendor due diligence has asked one question about hardware: who did you buy the chip from. That question is becoming the wrong one. The question a compliance-minded buyer needs now is where the physical hardware behind a rented GPU contract actually sits, and who else is renting time on the same cluster, because a reseller offering cheap GPU-hours through a chain of subsidiaries can be sitting on exactly the kind of offshore, multi-tenant capacity this review is built to catch.

The exposure is not that a European firm gets accused of anything. It is that a shared cluster named in a future BIS enforcement action, or a reseller placed on the Entity List, can strand a European customer's workloads with no warning and no easy fallback, and can turn a routine vendor relationship into a compliance incident a board has to explain. That is a new line item for AI supply-chain risk, sitting next to the model-provenance and data-residency questions most compliance teams already track.

The provenance check to add before your next GPU contract

Add one question to every cloud GPU procurement from now on: ask the vendor, in writing, which jurisdiction physically hosts the hardware behind your contract, and whether the cluster is shared with tenants outside your own compliance perimeter. A reseller who cannot answer that cleanly is telling you something. Document the answer the same way you already document model provenance and data-residency choices under the EU AI Act's vendor-risk expectations, so a change in a supplier's status shows up as a decision you made on purpose, not a surprise an auditor finds first.

None of this is settled law yet. The BIS review is a review, and the Remote Access Security Act is a bill in progress, not an enacted rule, so nothing forces a vendor switch today. What it does force is the habit of asking the location question before signing the next contract, while it still costs nothing to ask and everything to have skipped.

Frequently asked questions

Has the US already banned renting Nvidia chips through offshore data centers?

No. As of this review, offshore rental access sits in a legal gray zone that existing export rules were not written to cover. The Bureau of Industry and Security is formally reviewing the practice, and the Remote Access Security Act would extend export-control obligations to remote and cloud access, but neither the review nor the bill has produced an enacted rule yet.

Does this affect European companies renting cloud GPUs from mainstream providers?

Not directly, and most EU and UK buyers using major cloud providers have nothing to change today. The exposure sits with resellers and brokers offering discounted GPU capacity through opaque, multi-tenant, or offshore arrangements, which is exactly where a compliance-minded buyer should now ask where the hardware physically sits.

What actually triggered the Commerce Department's review?

Moonshot AI's release of Kimi K3, a 2.8 trillion parameter model that matched frontier performance from US labs within days. White House OSTP director Michael Kratsios publicly accused the firm on 22 July of acquiring Nvidia GB300 servers and running training workloads through Thailand-based infrastructure rather than importing the chips into China.

For five years the compliance question was who you bought the chip from. This review is the first sign that the real question is becoming where the compute you rent physically sits, and Washington has not finished writing the rule that will decide it. Ask the location question now, while it is a line in a contract and not a line in an audit finding.

Export controlsNvidiaBureau of Industry and SecurityChina AI chipsOffshore computeCloud GPU complianceAI Policy

25 AI Firms Ask Washington to Back Open Models

Twenty-five US technology firms signed a letter asking Washington to avoid premature restrictions on open-weight AI. Here is what it means for owners in Europe who want off vendor lock-in.

3 min read

Infrastructure

SpaceX Bets 10 Gigawatts on Nvidia's Rubin Chips

SpaceX says it will run about 10 gigawatts of AI compute by 2027, exclusively on Nvidia's Rubin chips - and the customer list is not who the headlines say it is.

4 min read

Decisions

The Codename Worked. The Trading Controls Did Not

An indictment unsealed on 24 July says two Volkswagen engineers traded on a joint venture known internally as Project Climb. The control that failed is a common one.

5 min read

← All articles