The amount of computing power behind artificial intelligence is growing at a pace that is hard to picture. According to a New York Times report highlighted by Techmeme, a deluge of AI computing power is set to come online in the coming years — and the research group Epoch AI expects the number of AI chips in use to double roughly every nine months.
That doubling rate is the number to sit with. A quantity that doubles every nine months grows more than fourfold in a year and a half, and more than tenfold in about three years. Applied to the specialized chips that train and run AI systems, it describes an infrastructure buildout compounding on itself rather than expanding in a straight line.
The Times frames this against how quickly AI capabilities have advanced. As the piece notes, the milestones keep getting grander: in 2023, an AI system passed the bar exam — a benchmark that, not long before, would have sounded like a distant hypothetical.
The link between those two threads is the point. More chips in use means more capacity to train larger systems and to serve them to more people at once. Compute has been one of the main inputs driving what AI models can do, so projections about chip supply are, indirectly, projections about capability.
It matters because the pace of AI's advance is no longer an abstract forecast about algorithms — it is being poured into physical hardware on a schedule, and that hardware, once built, will shape what the technology can do for years.