Researchers propose a hardware-enforced pause of frontier AI training

A proposal aimed at the hardware supply chain

A group of 26 researchers has outlined a path to an internationally verified pause of frontier AI training lasting at least ten years, Berkeley News reported on October 9. The university’s account describes a roughly 200-page report that would halt production of AI-training chips and gradually retire existing stock. Specialized inference-only chips would instead support approved existing models while being impractical for training new frontier systems.

The proposal therefore distinguishes using existing AI products from training more powerful models. During the transition, the researchers propose verification measures for remaining training-capable chips. Their argument relies in part on concentrated chip supply chains and manufacturing chokepoints that governments could regulate.

Feasibility is the authors’ argument, not an enacted agreement

Berkeley’s report says success depends on cooperation among world leaders and willingness to retire or transfer existing hardware. This is not an announced treaty, a current ban or proof that evasion can always be prevented. The group’s own English-language website also states that co-authorship does not imply agreement with every claim, and that institutional affiliations do not constitute institutional endorsement. The new development is a detailed governance proposal, not confirmation that governments have accepted it.