Researchers propose replacing AI chips to enforce a decade-long global training pause

A team of 26 researchers released a study on October 9, 2026, proposing a global pause of at least 10 years on training the most advanced AI models. Its enforcement tool is hardware: stop producing training-capable chips and replace existing stock with chips that run models but cannot practically train new frontier systems.

The Working Group on AI Pause Feasibility’s 200-page report presents a framework, not an adopted international agreement. According to Berkeley News, the team includes researchers from UC Berkeley, Princeton, Stanford, Harvard and Oxford, combining technical expertise with economics, international relations and experience in diplomacy and government.

Keep models running, remove the means to advance them

The researchers’ starting point is that frontier training—the work of developing ever more powerful AI models—requires chips that governments can regulate. They argue that the highly concentrated chip supply chain offers chokepoints for enforcement, while bypassing that supply chain would be extraordinarily difficult.

Under the proposal, production of AI training chips would stop. Most or all of the stock made before the pause would then be phased out and replaced with specialized inference-only chips. Inference means using a trained model to produce answers or other outputs, rather than training a new model.

The report says these replacement chips would run approved models quickly and efficiently, preserving continued use of existing AI products. Constraints built into the hardware would make them impractical for training new frontier models, even if the chips were stolen or seized.

Replacement would take time. During that transition, the researchers propose additional verification measures to prevent remaining training-capable chips from being used for frontier training. The hardware restrictions and monitoring are intended to work together, rather than treating a production ban as sufficient on its own.

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

Wesley Holliday, UC Berkeley philosophy professor and study co-organizer

Verification must cover secret training and open defiance

The report’s verification framework addresses a strategic obstacle: countries or companies may reject restraint if they fear rivals will cheat and gain a commercial or geopolitical advantage. The authors distinguish two ways a state could break the proposed agreement, each requiring safeguards:

• Covert evasion: a state secretly trains new frontier models while appearing to comply with verification procedures.
• Overt breakout: a state openly abandons the pause and tries to build a much more powerful model before other states can respond.

The authors acknowledge that success depends on world leaders cooperating and accepting difficult decisions about existing chip stock. Those decisions could include retiring chips or transferring them to other jurisdictions or internationally governed data centers, including scientific preserves.

A pause would also change chip economics

Study co-organizers Will Fithian and Holliday say technologies needed for inference-only chips are already being developed for economic reasons. But rapid turnover in AI models currently limits their appeal. They argue that prohibiting new frontier training would strengthen the economics of specialized chips and encourage further innovation in their design.

The report also outlines ways to make participation politically attractive: preserve AI-powered scientific and medical research, retain strategic insurance against states that defect, and pass cheaper inference costs on to consumers. These are proposed features of the arrangement, not benefits demonstrated by an operating global pause.

The researchers anticipate that the replacement chips’ lower resource demands would reduce environmental impacts and ease friction between U.S. data center operators and local communities. Their feasibility argument ultimately remains conditional: leaders must value longer-term outcomes, and monitoring must make consequences for breaking the agreement credible.