Control Selects for Copies

A regulator can block a foreign AI service by denying a license, restricting cloud access, or stopping a hardware shipment. The same regulator may be unable to remove a model whose weights have already been downloaded, mirrored, renamed, and installed on private machines. Distribution controls therefore do more than reduce access: they alter which form of access survives.

A centralized service remains attached to an institution. It needs data centers, payment networks, contracts, support staff, software updates, and permission to operate. Each dependency creates a point where authorities can demand an audit, impose conditions, or terminate access. Accountability makes the service safer to negotiate with and easier to stop.

Released model weights detach capability from the organization that created it. The original laboratory may still face penalties, but later copies can acquire different hosts, interfaces, modifications, and jurisdictions. Some will be public; others will not. The provider with a legal department is easier to govern than the copy with a checksum.

Three kinds of control are often blurred together. Controls on creation restrict advanced chips, training clusters, capital, energy, or specialized labor; these can reduce how much frontier capability is produced. Controls on inference restrict the computing infrastructure needed to run a model; these can limit scale, speed, and reliability. Controls on distribution restrict who may receive a service or artifact. This last layer is effective against official deployments but weak against files already released.

A copied model is not a copied data center. Large-scale inference still consumes hardware and electricity, and smaller operators may accept lower speed, reduced capacity, or degraded performance. Yet they do not need to repeat the original training process. The immense fixed cost of creating a model can coexist with a low marginal cost of reproducing its weights.

The result can look successful in administrative records: licenses denied, vendors excluded, contracts canceled. But the measurable decline may occur mainly among providers that preserve identities, logs, update channels, and emergency controls. Local or modified copies can remain available without those attachments.

This pattern extends beyond AI whenever a regulated capability can be separated from its accountable supplier and copied cheaply. Control still constrains quantity, quality, or scale, but it also changes the population of what remains. When capability is available either as a governed service or as a portable file, distribution controls select against the form with an owner, an audit trail, and an off switch.