A 26-researcher academic team, including two UC Berkeley professors, has published a roughly 200-page report arguing that it is technically feasible for the world's governments to enforce a mutually verified pause on frontier AI training lasting at least 10 years. The plan would halt new high-capability model training, inventory existing training-capable chips, and shift chip production toward devices that can run approved AI systems but cannot meaningfully train new ones.
The report, released Oct. 9 by the Working Group on AI Pause Feasibility, includes UC Berkeley professors Barry Eichengreen and Stuart Russell alongside faculty from Princeton, Stanford, Harvard and Oxford, according to UC Berkeley's news office. As the Daily Californian first reported, the project was organized by Will Fithian, a UC Berkeley associate professor of statistics, who said some people support the idea of a pause but remain unsure whether it's actually achievable.
How a Hardwired Pause Would Work
The proposal calls for prohibiting the training of new frontier AI models during the pause period, while creating a whitelist of existing models deemed safe enough to keep running, per the working paper hosted at hardwired-pause.ai. The plan would also ban production of specialized chips capable of both training new models and serving existing ones, according to the Daily Californian's reporting on the proposal. In their place, manufacturers would ramp up so-called inference-only chips, which the report describes as capable of fast, energy-efficient AI inference on approved models but not practical for training new frontier systems, per UC Berkeley's news office.
Fithian told the Daily Californian that existing training-capable chips already in circulation would be used for inference under monitoring, with some traded into internationally governed data centers the report calls scientific preserves. The report's own framing, per the same Berkeley account, is that such a plan would permit continued use of currently available AI products while preventing new models from overwhelming states — with most or all of the pre-pause stock of training chips phased out over time.
Why Chip Supply Chains Matter
The researchers lean heavily on the idea that the AI chip supply chain is highly concentrated with several chokepoints, which UC Berkeley's report summary says gives governments outsized leverage to enforce production rules. That concentration is already straining the industry: in 2026, AI chip production became a binding constraint on the pace of the broader AI compute buildout, according to a CNAS report, which also notes that building new manufacturing capacity takes years. CNAS points out that the rental price of NVIDIA's H100 chip, first released in 2023, is actually higher today than it was several years ago, even as computing costs have historically fallen. Meanwhile, Microsoft, Alphabet, Amazon, Meta and Oracle plan to spend almost $700 billion on capital expenditures in 2026 alone, the majority going toward AI infrastructure, per the same CNAS analysis.
Part of that bottleneck traces back to a single company: ASML is the sole supplier of extreme ultraviolet lithography machines needed to produce chips below 7 nanometers, according to a supply-chain map from MacroPolo, which notes each machine costs more than $150 million and takes 18 months to build. The Daily Californian's reporting also notes that the United States and China can each produce graphics cards — Nvidia makes Blackwell chips and Huawei makes Ascend chips — while Japan, the Netherlands, Germany and South Korea all produce components vital to the AI supply chain.
Consumer Chips Seen as a Non-Issue
One wrinkle the working group addressed head-on: millions of consumer-grade graphics cards, like Nvidia's 3090, are already in homes and offices worldwide. Per the Daily Californian, Fithian said those cards are less powerful than the chips the United States or China use to train frontier models and are far more diffusely held, meaning the paper's analysis suggests consumer graphics cards wouldn't pose a problem under the plan.
Newer hardware designed specifically for inference illustrates the direction the proposal envisions. Startup Taalas produces inference chips specialized to run one model per chip, with that model hardwired directly into the chip's construction rather than programmed like a typical graphics card, according to the Daily Californian. The outlet reports that a Taalas inference chip can serve an AI model 48 times faster than an Nvidia B200 graphics card — a preview of the kind of inference-focused hardware the pause proposal wants to see scale up as training-capable chips are phased out.
A Steep Logistical Climb
Independent estimates suggest the hardware transition, while large, might not take as long as it sounds. At the shipping rate researchers cited, replacing the global stock of training hardware would take about 16 months, while the broader transition could be completed within five years, according to TokenPost, which also reported that the plan would require countries to inventory training chips already in circulation. That urgency is rooted in how fast AI compute has been scaling: computing budgets for frontier AI training had increased by roughly an order of magnitude every 18 months over the preceding five years, researchers reported, according to AI Front Page.
Fithian compared the current moment to February 2020, when early concerns about a new respiratory virus prompted calls to shut down some international travel, according to the Daily Californian's report on his remarks. The comparison underscores the working group's central caveat: the proposed policy requires adoption by many countries to work at all, and the published paper itself does not assess whether governments are actually likely to implement it.
Momentum Beyond the Lab
The idea of deliberately slowing AI development isn't new. More than 1,000 technologists and researchers signed a 2023 open letter calling on AI labs to pause training of systems more powerful than GPT-4 for at least six months, according to the hardwired-pause.ai paper — though Lawfare notes that no lab actually paused in response. More recently, Anthropic said in June 2026 that it would be good for the world to have the option to slow or temporarily pause frontier AI development, and more than 1,000 employees of frontier AI companies signed a July 2026 letter asking the U.S. government to support international tools for deliberately pacing AI development, per the hardwired-pause.ai paper. Anthropic has also said publicly it would slow or pause its own development if other developers near the frontier did the same in a verifiable way, according to Lawfare's analysis.
The proposal has also drawn political voices into the conversation. Former national security advisor Susan Rice wrote in the New York Times that the U.S. should negotiate with China toward a freeze in place of advanced AI, per the hardwired-pause.ai paper, and the Daily Californian reports that Rice and Senator Bernie Sanders have each voiced support for pausing frontier AI development. Still, the dossier underlying the proposal makes clear that legal analysts see coordinated private-sector pauses as carrying substantial antitrust risk — a reminder that even a technically feasible plan faces a long road through politics and law before any chip factory actually slows down.
