AI chips gained about 49% more compute per dollar each year

Epoch AI has put a number on something the industry usually describes with hand-waving: how fast AI compute is getting cheaper. Looking at the chips actually purchased each quarter from early 2023 through the end of 2025, they find that performance per dollar grew about 49% per year, which works out to a doubling roughly every 1.7 years. The 90% confidence interval runs from 36% to 66%, so the trend is solid even if the exact figure is not.

The growth did not arrive smoothly. Through mid-2024 the gain was almost flat, around 6% a year, and then it jumped sharply as Blackwell-generation Nvidia chips took over the bulk of spending. Epoch built the estimate by dividing total computing throughput by total spending each quarter, in inflation-adjusted 2025 dollars, using a metric called Total Processing Performance that multiplies peak operations per second by the numerical precision width. The sample covers 24 chip models from Nvidia, AMD, Google, Amazon, Huawei and others. One comparison captures the longer arc: Nvidia's GB300 costs about 5.5 times what a 2016 P100 did, but delivers roughly 200 times the performance, so the cost-effectiveness is about 37 times better.

Epoch is careful that these are theoretical specs. Real workloads depend on software and utilization, and the paper numbers overstate what you get in practice. Still, the direction and the pace are what matter for anyone planning multi-year compute budgets.

Why it matters

If you are budgeting for training or inference over the next few years, this sets a defensible expected rate: assume the hardware you buy roughly halves in cost per unit of compute every 1.7 years, and treat plans that count on much faster or much slower drops with suspicion.