Hardwired pause: 26 researchers say a 10-year ban on frontier AI training is feasible

A group of 26 researchers says a worldwide pause on training frontier AI models for at least ten years is technically feasible, and describes how to hardwire it into chips. The Working Group on AI Pause Feasibility published its report, "The Feasibility of a Hardwired Pause of Frontier AI Training", on October 9, 2026. It is explicit about what it does not do: it conditions on world leaders who want a pause and does not assess whether a pause is desirable or likely.

The proposal in plain terms

The mechanism is chips, not code. The report proposes that states agree to stop training new frontier models, keep a whitelist of existing models deemed safe to deploy, halt or sharply reduce production of training-capable chips, and run a global census of the chips that already exist. Chips made during the pause would be "inference-only": able to run approved models quickly, but impractical to use for training even if stolen or seized.

The paper defines a training-capable accelerator by a throughput threshold set in a treaty. It also introduces a residual training factor: the ratio of an inference-only fleet's peak training throughput to its licensed inference rating. Its example value of 0.001 would cut a fleet's training capacity by 99.9 percent.

The numbers behind the feasibility claim

The report's executive summary rests on a few quantities. Frontier training budgets have grown by a factor of 10 every 18 months, and the pause holds if training efficiency gains stay at or below 10 times every five years, a rate the authors compare to Moore's Law. They say it would likely withstand a further 100 times in efficiency gains if the first 10 times is handled in time.

On verification, the summary says detecting covert frontier training is far easier when it takes hundreds of thousands or millions of chips, and that "reasonably effective verification methods would likely be more than sufficient in the short term". On breakout, an open attempt to resume would need a chip fleet larger than today's total world stock, so it is not an immediate threat.

The four conditions

The report says the pause is feasible only if four conditions hold: states cooperate to govern AI chip manufacture; states share information to locate pre-pause chips; inference-only chips are developed and built in time to phase out training-capable ones; and post-pause efficiency gains stay at or below the pre-pause rate. Each is a political or industrial bet, and the report does not claim they are likely.

UC Berkeley News, which covered the release, adds that inference-only chip technologies are already in development for economic reasons, and a training ban would raise their appeal. The report also argues a pause could be made politically acceptable through AI-powered science and medicine, insurance against falling behind, and cheaper inference for consumers. The main focus is the US and China, since both host leading developers.

What the authors admit

The paper lists its limits. A pause would not fully stop non-whitelisted models or small-scale fine-tuning of sub-frontier models. Residual training capacity in inference-only chips and untracked "dark compute" must be bounded, not eliminated. It cites weak institutional precedent: the Biological Weapons Convention's under-resourced support unit and the paralysis of the WTO Appellate Body. Taiwan's contested status is flagged as critical to chip manufacture. And the game theory may be a Stag Hunt rather than a Prisoner's Dilemma, meaning cooperation depends on timely, publicly observed monitoring and credible consequences for defection.

The authors are also clear on credit. The site notes that each listed co-author contributed ideas or text and that co-authorship does not imply agreement with every claim in the paper. The 26 members come from Berkeley, Princeton, Stanford, Harvard, Oxford, Tufts, MIT, Georgia Tech and other institutions, and include Stuart Russell, Barry Eichengreen, Rose Gottemoeller and David Krueger.

Where it sits in the argument

This is a technical answer to a political question the US Congress is already debating. We have covered the Sanders and Casar bill to ban artificial superintelligence and the enforcement problem in chip smuggling cases. A hardwired pause only works if the chips are under control, and those cases show how hard it is to keep even today's export controls watertight. The report's authors respond that supply chains are concentrated, with chokepoints such as a lithography machine weighing 180 tons, and that bypassing them is extraordinarily difficult.

What the report cannot answer is whether labs and governments want to be in the position it describes. Nothing in the documents we opened shows a government or lab adopting it, and none claims they will.

What is already in motion

The report notes that inference-only chip technologies are being developed for economic reasons, which is why the authors treat the hardware half of the proposal as an acceleration of existing trends rather than a new industry. It describes four phases, from pre-agreement through a phase where most inference runs on inference-only chips, and it defines its unit of account with care: an H100e is 989 TFLOP/s of dense 16-bit throughput, or roughly 3.1 times 10 to the 22nd FLOPs per year.

Those definitions matter because a treaty has to count something. The paper's answer is hardware capacity, not content: verification relies on quantitative accounting of chips, not on surveillance of prompts. That is a deliberate choice to make a pause checkable without reading anyone's data.

Our take

The strength of the report is that it moves the pause debate from "whether" to "how", with numbers a chip policy analyst can argue with. The weakness is the first word of every condition: states must cooperate, share, build and hold. We would watch for one signal. If a government negotiating position or a lab's policy cites the residual training factor or the census idea, the report has moved from paper to policy. Until then it is the best blueprint for an agreement nobody has agreed to want.