In an Oct. 9 working paper, the Working Group on AI Pause Feasibility proposes how countries could pause training the most capable AI models for a decade or more while continuing to run models on an approved list. Participating states or interim treaty authorities would decide which models make the list. The plan would eventually replace general-purpose accelerators, which can train models as well as run them, with specialized chips designed mainly to serve approved models.
Training uses computing power to change a model’s internal weights; inference is the stage when a finished model answers requests. Under the proposal, participating governments would stop training new frontier models, halt or sharply reduce production of training-capable chips, and account for the chips already in circulation.
Those existing chips could keep serving approved models during a monitored transition. Once specialized inference chips were ready for mass production, the proposal would phase out much of the older training-capable fleet. That would preserve access to selected AI services while reducing the stock of hardware available for large new training runs.
The report says inference chips may retain some training capacity. It proposes certifying an upper bound on how much training each chip could still perform, then capping how much of that capacity operators and countries could license. Hardware would also have to remain impractical for high-throughput training even if a state seized it.
The group’s preliminary analysis says a pause lasting a decade or more could be feasible if governments coordinate chip-manufacturing rules and share information on existing chips, replacement inference hardware arrives on time, and future training does not become more efficient faster than it did before the pause. The report does not assess whether world leaders are likely to seek a pause or whether one is desirable. It says this approach would complement other work on risks from current models. A state that openly breaks the pact would still pose an enforcement risk.
Sources
- Working Group on AI Pause Feasibility, “The Feasibility of a Hardwired Pause of Frontier AI Training” (working paper, v1.0, Oct. 9, 2026), especially §§1.1–1.2, 3.1–3.3, 4.1–4.4 and 7.6.
- Report executive summary.
- Illustrative photo of training-capable GPUs, not the proposed inference-only chips: Fukumoto, Wikimedia Commons, licensed under CC BY-SA 4.0.
