Your GPUs are American even when you aren’t: export control for AI clusters in Europe

There’s a comfortable assumption behind a lot of cross-border AI infrastructure: we’re not a US company, the cluster isn’t on US soil, so US export rules are someone else’s problem. It’s an assumption worth examining carefully, because when it’s wrong it tends to surface as a seized shipment or a penalty rather than a polite letter. US-origin GPUs can carry continuing US compliance obligations, and where your company is registered doesn’t by itself make those obligations go away.

The rule people miss

US export controls can continue to matter after the original export. The law firm Greenberg Traurig makes the strong version of the point: the controls can apply to GPU hardware in perpetuity, and even a non-US data-centre operator can face liability under the Export Administration Regulations — for example when it hosts a client’s controlled GPUs, or when restricted end users or end uses come into the picture.

The precise scope matters. Certain US export-control obligations can continue to apply after the original export — particularly when controlled GPUs are subsequently reexported, transferred, remotely accessed, serviced, or operated by restricted end users or for restricted end uses. It isn’t that every American chip drags its owner under permanent direct regulation in all respects; EAR applies to specific items, technology, destinations, end users and end uses, and much depends on the classification (ECCN), the route, and the licence or exception in play. But the idea that obligations simply end when the pallets clear customs is the dangerous one. As one compliance analysis of the 2026 rules put it, a data centre that legally received H100 servers in 2024 can still carry obligations in 2026 about who accesses that compute, for what, and under what logging and audit trail. Miss that, and the exposure isn’t a fine on paper — it’s a cluster that can be forced offline mid-contract while the gap is closed, with every GPU-hour of that downtime lost.

Why this deserves attention from an Asian owner in Europe

Picture the common case: an Asian company buys Blackwell-class hardware and places it in a European data centre, reasonably assuming Europe is neutral ground between Washington and Beijing. The relevant jurisdictional and licensing analysis depends on factors including the item’s origin and classification, the destination, the end user, the end use, and what happens to the hardware next — not on the location of the rack or the passport of the owner alone.

Two specifics make this concrete. First, reexport and transfer. A later reexport or in-country transfer can create a new licensing question even when the original export was lawful — the applicable authorization depends on the item, the destination, the end user and the relevant EAR provisions. “I already own it” doesn’t settle that question; moving the cluster onward later is its own event with its own rules — and a transfer made without the right authorization can strand a multi-million-euro asset in place, unusable until it’s resolved. Second, remote access, which is where the regulatory focus is visibly heading. A proposed Remote Access Security Act would extend attention from physical possession to certain forms of remote access to controlled items — its text reaches foreign persons accessing a US-jurisdiction item over a network, including cloud. It passed the House in January 2026 but, as of August 2026, remains pending in the Senate. A separate Chip Security Act (H.R. 3447), ordered reported by the House Foreign Affairs Committee in March 2026, would add security-mechanism requirements for advanced AI chips. Neither is law yet — but together they show the direction, and the direction is toward following the access, not just the box. For an operator that means a compliance gap can surface not at the border but years in, over who logged in — long after the hardware is paid for and earning.

What it costs to get wrong

The scale of current US export-control enforcement is not theoretical. In February 2026, BIS announced a $252.5 million settlement with Applied Materials over unauthorized exports of semiconductor manufacturing equipment to China — its second-largest standalone penalty. That case was not about GPUs, but it shows the size of the enforcement environment around advanced semiconductor technology. The more common cost is quieter: hardware held at the border. A documented deployment scenario had a large GPU order stuck in customs for four months because the required registration wasn’t in place before shipment — and while it sat, demurrage ran daily, SLA penalties accrued, and millions in compute revenue never happened. The hardware was legal. The paperwork wasn’t ready. That gap was the whole loss.

What this means when you commission a cluster

Compliance isn’t a lawyer you call after the hardware lands — it’s something mapped before anything ships, and it leaves a trail through the whole build. The export classification of the GPUs and the licence or exception they move under has to be established. Certification should be collected from the exporters and intermediaries confirming their own compliance — and a party who won’t provide it treated as the red flag it is. And once the cluster runs, the access controls, logging and audit framework have to be maintained to show, on an ongoing basis, who is using the compute and for what. This is continuous work, not a one-time form.

The uncomfortable part for a newcomer is that this sits on top of the physical build, not beside it. Rack and validate the cluster but skip the export trail, and you have an asset that runs beautifully and can still be frozen by a compliance gap you didn’t know you had.

Where GTVA fits

This is exactly the seam we’re built for: European infrastructure, with people who understand both the hardware and the rules that come attached to it. We commission clusters in the EU, under European jurisdiction, with the technical controls, documentation and operational processes needed to support GDPR and NIS2 compliance built into the infrastructure rather than bolted on afterwards — and with the export-control trail treated as part of commissioning, not an afterthought. Classification and licence documentation checked before hardware moves, certification collected from the parties in the chain, access and audit framework stood up so the cluster is defensible on day one.

For an operator bringing American hardware into Europe, the value isn’t only that we can build the cluster. It’s that we understand the jurisdiction it’s landing in — and the obligations it can carry with it.