The Pulse
Researchers Map a Decade-Long Pause for Frontier AI Training
A working group’s report examines how controls on training chips could support a long pause in frontier AI training. Its feasibility assessment depends on governments cooperating on the proposed governance and verification measures.
AI.info Team · Written with AI ·
Governments considering a long halt to frontier AI training face a strategic bind: any country that pauses may fear rivals will keep building more powerful models. A new report from a 26-member international working group argues that a decade-long pause could be verified through controls on the chips used to train those models. Its conclusion is conditional, however: the researchers do not say governments are likely to agree, or that a pause is the best policy. They examine whether an agreement could work if leaders wanted one.
Training chips are the proposed pressure point
The report, released October 9, proposes halting production of chips capable of frontier-scale training, then gradually replacing existing stock with specialized inference-only chips. Inference is the use of an already-trained model; the proposed replacement hardware would run approved models quickly and efficiently but would not be practical for training new frontier systems. The authors say the chip supply chain has concentrated production chokepoints that governments could regulate.
Existing training-capable chips would not all disappear at once. During the transition, the report proposes monitoring those chips and preventing their use for frontier training, while new inference-only hardware takes over commercial services. The authors also discuss transferring or retiring much of the old stock, including possible placement in internationally governed research facilities or custody outside a state’s borders.
“Our assessment is that an implementation similar to that described in § 3 would likely be technically feasible for the immediately foreseeable future, if the governments of the U.S., China, and the supplier and chip owner states listed in § 3.1 implemented the governance and verification measures discussed in § 3.3.”
The Working Group on AI Pause Feasibility, the report’s authors
Monitoring must catch cheating—and deter a breakout
The plan addresses two distinct ways an agreement could fail. In covert evasion, a state secretly trains new frontier models while appearing to comply; the researchers argue that training at frontier scale would require hundreds of thousands or millions of chips and months or years of use, making it harder to conceal than smaller-scale training. They say verification need not catch every violation to make the pause workable, but would need to support credible consequences.
Open defiance poses a separate problem. A state could abandon the agreement and try to assemble a large enough chip fleet to race ahead before others respond; the report says monitoring alone cannot prevent that. Limiting the accumulation of training capacity and moving much of the declared pre-pause stock out of national control are among the measures it considers for reducing that risk.
The decade depends on chip counts and training efficiency
The authors identify four conditions for a pause lasting at least a decade: states must cooperate on chip manufacturing rules, share information to locate existing chips, produce inference-only hardware in time to replace training chips, and keep gains in training efficiency within the report’s assumed range. A global census of pre-pause chips would help identify undeclared capacity, while commercial access to transitional inference services could give owners an incentive to declare their hardware.
Training efficiency is a key uncertainty because algorithmic and data improvements can make a fixed quantity of computing power more productive. The report says the pause could likely endure for at least a decade if efficiency gains for training models near the paused frontier do not exceed an order of magnitude every five years. That estimate depends on the researchers’ assessment of historical gains and on the possibility that those gains could slow when frontier-scale experiments stop.
A feasibility study, not a proposed treaty
The working group explicitly does not assess whether a pause is desirable or how likely leaders are to implement one. Its analysis assumes a hypothetical in which governments would prefer to pause if they could trust others to do the same or detect defections quickly enough. That distinction matters: the report outlines a possible verification system, not an agreement adopted by any government.
The authors also say a pause would not resolve every problem created by existing AI systems. Misuse and malfunctions would still require governance, and limits on large-scale training could make it harder to fix model security flaws through retraining. The report flags another risk: depending on how rules are designed, restrictions could entrench the market positions of leading AI companies. Its central political test remains whether states can agree to disclose, monitor and ultimately retire the training hardware on which their own AI programs depend.