A working paper published on 9 October 2026 proposes using specialised chips to help verify a global pause on training frontier AI models. The plan would last at least 10 years. It is a proposal, not an agreement by governments.

The paper, by an interdisciplinary group of 26 researchers, is available at hardwired-pause.ai. UC Berkeley News also reported on its assessment that such a pause could be feasible under certain conditions.

Keep approved AI running, halt new training

The proposal, called a “hardwired pause,” would prohibit training new frontier models while allowing approved existing models to remain in use. It would also permit chips designed for inference, which means running a trained model to produce answers.

  • Proposed pause: at least 10 years.
  • Approved models could continue operating.
  • New chips would be designed to run models, not practically train a new frontier model.

The authors argue that these specialised chips should not be able to train a new frontier model even if stolen or seized. But the plan would also need safeguards for training-capable chips already in circulation.

Chip controls would need international cooperation

The researchers propose monitoring key points in the chip supply chain and tracking existing hardware to detect attempts to evade the pause. That would make cooperation among countries that produce or own chips essential, including the United States and China.

Chip design alone would not enforce the plan. The paper assumes governments would consider a pause if they could trust one another to comply or quickly detect violations. It also identifies risks that remain unresolved, including covert evasion and gains in training capacity as chips become more efficient.

A feasibility study, not a prediction

The authors describe their work as a preliminary assessment of how a pause might be implemented if world leaders were willing to consider one. They do not argue that a pause is desirable or predict that governments will agree to it.

The proposal’s practical test is whether governments and private chip and AI companies can agree on approved models, existing hardware, inspections and enforcement, then sustain those rules for a decade or more.

For users and businesses, the distinction is significant: the proposal aims to keep some existing AI services available while restricting the hardware used to train more powerful models. Whether that balance is workable depends on political agreement and effective oversight, not chip design alone.