Export controls on advanced semiconductors rest on a specific theory: frontier capability requires frontier compute, so restricting the compute restricts the capability. The theory is sound. Its weak point is the assumption that capability must be created rather than acquired.
The logic of the regime
Training a frontier model requires very large quantities of specialised accelerators. Those come from a small number of suppliers with traceable supply chains, which makes them an unusually good control point: physical, countable and licensable.
Restrict the chips and you raise the cost and time required to train a frontier model. That part demonstrably works.
Sovereign AI
The controls produced a predictable response. Many states concluded that depending on foreign models for critical infrastructure was itself a strategic risk, and began funding national AI programmes.
Sovereign AI usually means some combination of: a model trained or controlled domestically, hosted on domestic infrastructure, operating under domestic law, and not subject to a foreign provider's decisions about access. These are reasonable objectives independent of any dispute.
The difficulty is that most states pursuing them cannot access the compute to train a frontier model, or cannot justify the expense. Which leaves acquisition.
The hole
Distillation transfers capability without repeating its creation. A student needs a small fraction of the teacher's compute, and it needs no controlled item at all. Nothing is shipped. There is no export, no licence to deny, and no border crossing to interdict.
A national programme unable to obtain enough accelerators to train a frontier model can still obtain frontier-level behaviour, using ordinary hardware and an API account. The controls constrain who can build a frontier model. They do not constrain who can end up with one.
This is the structural argument that made distillation a policy question rather than a contractual one, and it is the strongest version of the frontier labs' concern. We follow the consequences in The cascade problem.
Why the enforcement instruments fit badly
Every proposed response targets a party and requires establishing what they did.
- Entity List designation requires an evidentiary record, and the technical evidence does not exist
- Sanctions require attributable conduct, and querying a public API is not obviously that
- Terms of service enforcement is a contract remedy scaled for damages, not statecraft
- Compute thresholds measure how a model was made, and a distilled model was made cheaply. See What Is a Frontier Model?
Where this stands
Through 2026 the gap was identified repeatedly, sanctions were repeatedly raised, and no enforcement action was taken. That is not indecision so much as a reflection of the mismatch: the available instruments were designed for physical transfers, and this is not one.
A regime that addressed it would have to regulate behaviour rather than hardware, which is a considerably harder thing to write and a considerably harder thing to verify.
Common questions
What are AI export controls?
Restrictions on the sale and transfer of advanced semiconductors and related equipment, intended to slow the development of frontier AI capability in targeted countries by limiting access to the compute required to train it.
Why does distillation undermine export controls?
Because the controls restrict the compute needed to create capability, while distillation transfers capability that already exists. A student model needs a fraction of the teacher's compute, so a lab constrained on hardware can still obtain frontier-level behaviour without ever training a frontier model.