A TOKENPOST analysis says U.S.-China chip controls could enforce an AI pause lasting at least 10 years.
A proposed global pause on training more powerful artificial intelligence models could be enforced through advanced-chip controls, if U.S. and Chinese leaders chose to halt or sharply reduce training-chip production, according to an analysis. The pause would last at least 10 years. Existing AI services could continue if they were placed on a whitelist. New frontier-model training would be prohibited, and chipmakers would shift toward hardware designed only to run existing models.
The roughly 200-page analysis, released Oct. 9, was written by 26 scholars led by Berkeley professors Will Fithian and Wesley Holliday. It examined whether governments could stop frontier-model training while keeping approved services operational.
Under the proposed system, countries would inventory the training chips already in circulation. Existing hardware could initially be used for inference, the process of running an existing model, under verification controls. Governments could later phase out those chips through buyback programs and exchanges for newer inference-specific hardware.
The analysis estimated that global training capacity would reach the equivalent of 47 million Nvidia H100 chips by the end of 2026, with about 2.9 million additional chips shipping each month. At that pace, replacing the global stock of training hardware would take about 16 months, and the broader transition could be completed within five years.
Small-scale secret training would be difficult to detect. However, developing a model beyond the current frontier would require hundreds of thousands to millions of chips operating for months. A country secretly seeking a strategically significant model within five years could need about 6.4 million chips. A country openly abandoning the arrangement and pursuing a decisive model within two years could need about 180 million chips, nearly four times the estimated global stock.
The analysis identified two vulnerabilities. The United States could retain enough training hardware used for transitional inference to restart training. Undeclared computing capacity in China could also support covert development.
Supply-chain controls would be central to the arrangement. ASML holds a monopoly on manufacturing the extreme ultraviolet lithography machines needed for the most advanced chips. Taiwan Semiconductor Manufacturing Co. operates leading-edge foundries and packaging facilities, and Nvidia chips account for more than 60% of global AI computing capacity.
The political conditions for such an arrangement are not yet in place. Leaders from more than 20 countries responded to a Sept. 21 appeal launched by Finnish President Stubb and Norwegian Prime Minister Støre to control frontier AI models, but the United States and China did not participate.
The broader economic impact could be limited because most businesses do not use frontier models. Even a $6.5 trillion decline in the market value of the seven largest U.S. technology companies, back to early-2025 levels, would reduce U.S. consumer spending by about $195 billion, or 0.6% of 2026 gross domestic product.
Nvidia would face a particularly difficult adjustment because of its heavy investment in training-class chips. Shifting toward inference-specific hardware could be difficult and expensive. The pause could also strengthen dominant companies and leave AI models with increasingly outdated knowledge.