Answers are grounded in this paper and cite their sources.
Computes how supply severance impacts AI capacity decay in nations, suggesting significant implications.
Every argument for restricting the export of advanced AI chips rests on a claim about time: controls buy time. Nobody has ever said how much. This paper computes it, and finds that the usual explanation of why controls work is the wrong one. The mistake: treating chips like gold National AI capacity is measured constantly. At least eleven public indices rank countries by how much computing power, talent and infrastructure they hold. All of them take a photograph. None asks what happens to the picture afterwards. That matters, because installed computing power is not a reserve sitting in a vault. It is a fleet of machinery that breaks. Using published failure data from a real training run, roughly 9% of accelerators fail in a year from causes that cannot be repaired in the field. They also age against a target that keeps moving: the scale of a state-of-the-art training run roughly doubles every ten months. A gold reserve does not corrode. A chip fleet does, twice over. The engineering literature has measured the corrosion precisely, but always assuming replacement chips can be bought. That assumption is exactly what a supply cut-off removes. Removing it is what this work does.
No takes yet. Share an insight, caveat, or question.
Kevin Escoda (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: