A report by 26 scholars from Berkeley and other institutions says a global pause on training more powerful AI models could be technically achievable for at least 10 years if leaders in the US and China both wanted it. The proposal focuses on hardware rather than shutting down AI products already in use: stop making training chips and shift production toward chips designed only to run existing models.
The roughly 200-page report was released on Oct. 9 and was initiated by UC Berkeley statistics associate professor Will Fithian and philosophy professor Wesley Holliday. Under the framework, services already online, including ChatGPT and Claude, could keep operating if they were placed on a whitelist. What gets blocked is the training of stronger new frontier systems.
A hardware-locked pause
The authors describe the approach as a pause “written into hardware.” It has five parts:
- ban the training of new frontier models, meaning the strongest models available at the time;
- create a whitelist so already deployed models judged safe enough can continue serving users;
- stop producing, or sharply reduce output of, AI chips used for training;
- switch production to inference-only chips;
- conduct a global inventory of training chips that existed before the pause and place them under oversight.
The report says the pause should last at least 10 years and could be extended if governments agreed.
Why inference chips could become economically viable
According to the report, buying inference chips is not attractive under current conditions because new models change so quickly that hardware can become outdated fast. If training new models were banned, that economics would change and inference chips would become more compelling.
For legacy training chips, the authors propose using them for inference during a transition period, with verification measures aimed at preventing them from being redirected back into training. Over time, those chips would be phased out through trade-in programs and government buybacks. The report calls this installed stock “kindling,” arguing that if it is left in place, it could fuel a renewed race later on.
How long the transition might take
The report estimates that by the end of 2026, global training-grade compute will amount to about 47 million H100-equivalent chips, using NVIDIA’s H100 as the reference unit. Monthly new shipments are put at about 2.9 million. Replacing the full stock would take around 16 months, and the entire transition could “possibly” be completed within five years.
Could countries train in secret?
The authors argue the real issue is scale, not abstract technical possibility. Training a small model quietly on a single machine would be nearly impossible to detect. Training a system that surpasses the current frontier would require hundreds of thousands to more than a million chips running continuously for months, which would be much harder to conceal.
The report splits cheating into two categories. One is covert cheating: for a country to secretly train a strategically meaningful model within five years, it would need about 6.4 million chips. The report says that is roughly equal to OpenAI’s expected entire fleet by year-end, while China as a whole is estimated to have only 3.1 million chips. The other is open defection: if a country tore up the agreement and tried to produce a decisive model within two years, it would need about 180 million chips, nearly four times current global stock.
The report highlights two weak links. First, if the US retains a large number of training chips for transition-period inference, it preserves the capacity for open defection. Second, undeclared “dark compute” in China could be the biggest opening for covert cheating.
The supply chain as an enforcement lever
The report says the AI hardware supply chain offers another layer of control. Fithian writes: “Circumventing existing supply chains is extremely difficult, and international prohibitions can exploit this feature of the hardware required for AI training.” The report points to ASML in the Netherlands as the sole supplier of EUV lithography systems needed for the most advanced chips, TSMC as the owner of the most advanced fabs and packaging capacity, and NVIDIA chips as accounting for more than 60% of global AI compute.
Politics remains the real barrier
The authors make one condition explicit: the report asks whether a pause could be enforced if US and Chinese leaders wanted one. It does not ask whether they actually do.
Current geopolitics points the other way. The report says US-China competition in AI is intensifying, not slowing. It cites Donald Trump as saying, “Whoever wins AI wins.” On Sept. 21, Finnish President Alexander Stubb and Norwegian Prime Minister Jonas Gahr Støre launched an appeal to control frontier AI models. About 20-plus national leaders responded, but neither the US nor China joined.
There have also been calls from industry to slow down. Anthropic CEO Dario Amodei has urged people to “slow down on the frontier,” and Sam Altman and Elon Musk have also signaled agreement. But Amodei said plainly that slowing down is not the same as stopping training. The report adds that Beijing may view any pause agreement as an extension of US chip export controls.
Who bears the cost
The report argues the overall economic shock would be limited because most companies do not rely on the most advanced models. Even if the market value of the “Magnificent Seven” fell back to early-2025 levels, wiping out about $6.5 trillion, the hit to US consumption would be about $195 billion, or around 0.6% of 2026 GDP. The Federal Reserve, the report says, would still have room to cushion the blow.
NVIDIA would face the hardest adjustment. The report says the company is heavily committed to training-grade chips and would need to pivot toward inference-only products, a transition it describes as “neither easy nor cheap.” The authors also acknowledge other costs, including the risk of entrenching incumbent leaders and the growing staleness of model knowledge cutoffs.