A global 10-year pause on AI training is feasible, say 26 Berkeley-led researchers — if we hardwire it into the chips

A 26-researcher team led by UC Berkeley scholars has released a 200-page blueprint for an international, mutually verified pause on frontier AI training of at least 10 years — enforced not by promises, but by chips that physically can't train new models.

A 26-researcher team led by UC Berkeley scholars has released a 200-page blueprint for an international, mutually verified pause on frontier AI training of at least 10 years — enforced not by promises, but by chips that physically can't train new models.

Twenty-six researchers from UC Berkeley, Princeton, Stanford, Harvard and Oxford say a global pause on frontier AI training is feasible — not someday, but mapped out today in a 200-page report from the Working Group on AI Pause Feasibility, released Friday. Their mechanism: a 10-year, mutually verified halt enforced in hardware, by replacing training chips with chips that physically can't train.

The study is the first academic attempt to answer the full question: if the world's governments and AI leaders ever agree to pause frontier training, how would it actually work? The team — organized by UC Berkeley's Wesley Holliday and Will Fithian, and including economists Barry Eichengreen, computer scientist Stuart Russell, and faculty from Princeton, Stanford, Harvard and Oxford — blends computer science, economics, history, international relations and diplomacy with practical government experience.

The plan: retire the training chips

Everything hinges on hardware. Frontier training requires AI chips, and states can govern chips — the supply chain is highly concentrated, with chokepoints that give governments real leverage. "Bypassing existing supply chains is extraordinarily difficult," writes co-organizer Will Fithian, UC Berkeley associate professor of statistics. "An international prohibition on frontier AI training can take advantage of this fact about the hardware needed for AI training."

The proposal: halt production of AI training chips, phase out the pre-pause stock over time, and replace them with specialized inference-only chips — chips that serve fast, energy-efficient inference on approved models but are not practical for training new frontier models, even if stolen or seized, thanks to constraints hardwired in. During the transition, verification measures would keep residual training-capable chips out of frontier training runs. Existing AI products would keep running; new, ever-more-powerful models simply couldn't be built.

Verification, not trust

The report's core insight is political: a pause could be in everyone's interest and still fail, because no one wants to be the one who paused while a rival raced ahead. The framework therefore centers on mutual verification — measures that let world leaders assure nervous rivals that others are upholding the deal, and detect defections quickly.

"We consider a scenario in which world leaders want to pause if they can be confident that others are pausing or that defections would be detected quickly enough," writes co-organizer Wesley Holliday, a UC Berkeley philosophy professor, in the study.

The report names the two failure modes explicitly: covert evasion, where a state secretly trains new frontier models while ostensibly complying, and overt breakout, where a state openly abrogates the pause to try building a much more powerful model before anyone can respond. The hardwired approach guards against both — breakout attempts can't conjure training-capable hardware from a retired supply chain.

What makes a pause politically palatable

The researchers don't hand-wave the politics. The report sketches ways to keep AI-powered scientific and medical research going during a pause, retain strategic insurance against being left behind by a defecting state, and pass the economic benefits of cheaper inference to consumers. Inference-only chips already exist as an economic proposition, Fithian and Holliday note — their appeal has just been limited by the rapid turnover of new models. Prohibit new frontier training and suddenly those specialized chips become the most attractive hardware around.

There are side benefits too: the new chips' reduced resource demands would lower the environmental footprint and ease frictions between U.S. data center operators and local communities. And the report leans on precedent — an international agreement controlling nuclear energy's military uses while promoting peaceful ones, and the pact that protected Earth's ozone layer.

The real question

The researchers agree the whole thing stands or falls on leaders' willingness to cooperate — to retire or transfer pre-pause chip stock to other jurisdictions or internationally governed data centers, like scientific preserves. "Coordination is possible if the decision makers care enough about the future," the authors write, "not just immediate payoffs, and if monitoring is adequate to support credible consequences for defections."

The report frames it as one of the century's defining questions: whether humanity can unite to regulate AI's "dual-use dynamics" — its capacity to benefit or to harm. Whether that coordination is possible is a question for leaders, not engineers. But for the first time, someone has written down how the engineering could actually work. (Berkeley News) (The full report, hardwired-pause.ai)