Deemed exports and the compute pipeline nobody talks about

4 min read · 714 words
Share:
Michael Darius Eastwood
Michael Darius Eastwood · Independent AI alignment researcher
Published
Michael Darius Eastwood · Policy · 3 July 2026
Michael Darius Eastwood, independent researcher, London: originator of the embedded-correction alignment thesis (manuscript 8 December 2024, SHA-256 anchored: f0d1f38f).

The deemed-export rule is a piece of US export-control administration that treats the release of controlled technology to a foreign national inside the United States as an export to that foreign national's country. It was designed for classified physical or design information handed to non-US persons in visits, labs, or classrooms. It maps awkwardly onto compute.

The rule in one line. Access to controlled technology by a foreign national is treated as an export. Where the technology is knowledge, access is what triggers the control. Where the technology is compute, the question of what "access" means is not settled by the classical text, and 2026 policy is filling the gap in real time.

Where the classical rule works

A physically instantiated piece of technology, drawings or specifications for a controlled item, or the details of a manufacturing process: the deemed-export framework maps cleanly. The person in the room gets access; the release is the release. Universities, labs and companies file the same license paperwork the shipment abroad would need.

Where it maps awkwardly onto compute

Cloud-based frontier training is a different object. The technology being controlled is not the drawing of a chip; it is the ability to schedule tens of megawatts of AI compute for extended runs, on architectures the export regime treats as controlled. A foreign national with a cloud account can, in principle, submit workloads to that compute without ever occupying US soil. The deemed-export framework was not written for that pattern.

What 2026 policy is doing

The US export-control apparatus has spent 2025 and 2026 pushing controls at the compute layer directly, through model export orders (the Fable 5 case), through end-use and end-user restrictions, and through the semiconductor supply chain (Chips Act 2.0 in Europe; SAFE Chips in the US). The deemed-export doctrine is being extended in practice through workflow and licensing conditions attached to who can rent controlled compute rather than through statutory revision. That is administrative flexibility, and it is also administrative fragility.

The obvious failure modes

The first is uneven enforcement. A cloud provider without visibility into ultimate end use cannot police access in the way a physical facility can. The second is jurisdictional arbitrage: workloads submitted through third countries with weaker controls. The third is the definitional problem the classical rule leaves open: what counts as "access" to a training run that is a black box the customer never sees inside?

Why this matters to the programme

The measurement instrument the programme proposes, beta over k on frontier models, has no export-control dimension in principle. In practice, however, running the sweep requires access to compute the export regime is increasingly asking questions about. A public measurement programme that cannot cross national boundaries becomes a governance instrument only within one jurisdiction, which is not what governance instruments are supposed to be.

What could be written into policy without a statutory revision

Two clauses. A public-interest carve-out for research measurements published under preregistration and open replication. And a mutual-recognition clause with trusted partners on such measurements, so that a beta over k number produced in one jurisdiction is admissible in another. The Macron trusted-partners framework at G7 Évian is one of the few current policy vehicles that can carry both.

What the reader keeps

The deemed-export rule is being stretched by cloud compute. 2026 is patching the stretch administratively rather than statutorily. Measurement programmes that want to be global governance instruments need explicit carve-outs, and those carve-outs are cheap for policy makers to write and expensive for them to omit.

From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.

Buy on Amazon UK Amazon US

Stay informed

New posts on AI alignment, convergence evidence, and the ARC/Eden research programme.

Get updates →