Epoch AI, a research group that maintains a public database of notable machine learning models and their training requirements, reports that the compute used to train frontier language models has grown at roughly 5 times per year since 2020, equivalent to doubling about every 5.2 months. Over a longer window, Epoch finds that the compute used to train notable AI models has grown about 4.5 times per year since 2010.
These growth rates are far faster than the historical pace of hardware improvement. Moore’s Law, the long-running trend of transistor density roughly doubling every two years, would translate to compute doubling on the order of every couple of years, not every several months. The gap means that the explosion in AI training compute has come not just from chips getting better but from companies spending dramatically more, deploying ever-larger clusters of accelerators, and engineering the software to use them efficiently.
For a business reader, this figure puts the AI capital-spending boom in perspective: a roughly 5-times annual increase compounds extraordinarily quickly, so the resources behind a leading model can be an order of magnitude larger than those behind a model just two years older. It also explains why AI datacenters, power supply, and chip availability have become strategic concerns rather than mere engineering details.