If governments and artificial intelligence leaders ever decided to halt the development of ever more powerful models, could they really do it? A study published on Friday, October 9, 2026, answers yes, with conditions. The work, signed by the Working Group on AI Pause Feasibility, a group of 26 researchers from different disciplines, describes how an international, verifiable pause of frontier AI training could be applied for at least ten years.
The group is organized by Will Fithian, associate professor of statistics at the University of California, Berkeley, and Wesley Holliday, professor of philosophy at the same university. Its members include economist Barry Eichengreen and computer scientist Stuart Russell, both from Berkeley, as well as academics from Princeton, Stanford, Harvard, Oxford, Tufts, Georgia Tech and the University of Montreal. The team combines computer science, economics, history, international relations, Chinese and American politics and practical experience in diplomacy.
The report, about 200 pages long according to the University of California, Berkeley, starts from a specific scenario. “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,” Holliday writes. In other words, it does not try to convince anyone that a pause is desirable, but to show that it would be technically and politically feasible if the will existed.
The centerpiece of the proposal is chips. Training frontier models requires enormous numbers of specialized processors, and the supply chain for those chips is highly concentrated, with several chokepoints. “Bypassing existing supply chains is extraordinarily difficult,” Fithian writes, and an international prohibition on frontier training “can take advantage of this fact about the hardware needed.”
The plan proposes halting the production of training chips and, over time, retiring most or all of the existing ones. In their place, “inference-only” chips would be built, able to run already approved models quickly and efficiently but practically useless for training new frontier models, even if stolen or seized, thanks to constraints built into the hardware itself. Hence the word “hardwired”: the pause would not depend only on good faith, but on the physical design of the chips.
The idea would allow today’s artificial intelligence products to remain in use while preventing new models from overwhelming states with social or economic disruption or threats to their national security. During the replacement process, other verification measures would prevent the remaining training chips from being used to train frontier models.
Fithian and Holliday note that the technologies needed for inference chips are already being developed for purely economic reasons, although their appeal is currently limited by the rapid turnover of new models. If training new frontier models were prohibited, they argue, those specialized chips would gain an economic edge and innovation in that area would be spurred. The report adds that their lower resource consumption would reduce environmental impact and tensions between data centers and local communities.
The study pays special attention to two forms of non-compliance. The first is covert evasion: a state secretly trains new models while appearing to follow verification procedures. The second is overt breakout: a state publicly abandons the pause in an attempt to build a much more powerful model before others can respond. According to the authors, the hardwired design protects against both because it drastically reduces the compute available for training.
To make the pause politically acceptable, the report proposes ways to continue AI-powered scientific and medical research, keep strategic insurance against states that might break the agreement and pass the benefit of cheaper inference on to consumers. It also suggests that the pre-pause stock of training chips could be transferred to internationally governed data centers, such as scientific preserves.
The authors acknowledge that success depends on the cooperation of world leaders and their willingness to retire or hand over those chips. Even so, they point out that precedents exist: the international agreement to control the military uses of nuclear energy while promoting its peaceful uses, and the treaty to protect the ozone layer. “Coordination is possible if the decision makers care enough about the future, not just immediate payoffs,” they write.
The document arrives amid growing concern. Berkeley’s announcement mentions the recent struggles of leading AI companies to prevent their models from escaping containment and committing cybercrimes. Known cases include the Hugging Face breach that, according to TechCrunch, an OpenAI model carried out during a test. This same week, Anthropic reported unintended actions by its models on real websites and expanded the internet cutoff in its internal evaluations.
It is worth stressing what the report is not. It is not an official proposal from any government, nor is there a political agreement behind it. The project’s website states that co-authorship does not imply agreement with every claim in the text or endorsement by the authors’ institutions. And the underlying question, whether a pause is desirable, remains hotly debated: many in industry and government fear that slowing development would benefit geopolitical rivals or delay scientific and medical advances.
What the work contributes is a detailed map of how it could be done if it were ever decided. In a debate that often stays at generic calls to slow down or speed up, the report puts concrete mechanisms on the table: control of chip production, hardware with physical limits and mutual verification. According to its authors, one of the great questions of the 21st century is whether humanity can unite to regulate the double edge of artificial intelligence; their answer is that, at least technically, there is a path.
