US export controls on AI chips cannot stop capable models from being built abroad, and DeepSeek is the clearest proof yet, according to a new analysis from the Brookings Institution.

The Chinese AI lab released models that compete with leading American systems while operating under the very restrictions Washington designed to slow Chinese progress. Brookings argues that this outcome exposes a structural weakness in the controls, not a failure of enforcement.

What did DeepSeek actually demonstrate?

The controls rest on a simple premise. If Chinese firms cannot buy the most powerful accelerators, they cannot train frontier models. DeepSeek showed that software engineering can substitute for raw compute. The lab squeezed more capability from restricted hardware through efficiency work rather than by buying more chips.

This matters because the rules measure restrictions in hardware terms. They set thresholds on chip performance and on how many units a buyer may acquire. They do not account for what a skilled team can extract from the silicon it already has.

The takeaway is plain: compute caps slow development, but they do not cap it.

Why do the controls still matter?

Brookings does not argue the rules are useless. Restricted access to top-tier parts raises costs, lengthens training timelines and forces workarounds. That burden is real, and it shapes investment decisions across the Chinese AI sector.

But the analysis draws a sharp line between slowing a competitor and stopping one. Efficiency gains compound. Each generation of software improvements narrows the gap that the hardware rules create. A policy built only on chip ceilings therefore loses force over time.

What does the analysis suggest instead?

The Brookings piece points toward controls that cover more of the AI supply chain. Chips are one layer in a stack that also includes model weights, algorithms, data and expertise. A regime that targets only the hardware layer leaves the rest of the stack open.

This reframing has commercial consequences for US semiconductor firms. Nvidia and its peers sell restricted parts into a market that Washington keeps redefining. Each revision of the rules reshapes their addressable revenue. If policy shifts toward software and model-level controls, the compliance burden would spread beyond chipmakers to cloud providers and AI labs.

The open question is whether future rule revisions will follow this broader framing, or whether hardware thresholds will remain the only lever Washington pulls.