AI chip race calculator

Set how fast demand for AI computing grows, how fast ASML makes EUV machines and how much more each chip does a year, and see how many chips AI would want against how many the machines could make, year by year to 2035.

AI’s growth against the machines

Set how fast each one grows a year. Both sides start at 1 in 2025. It’s a simple illustration of compounding, not a forecast.

Growth each year
3.3× a year

Epoch AI: the computing power of AI chips sold has grown about 3.3 times a year since 2022.

1.3× a year

ASML expects about 65 in 2026 and is reported to be aiming for more than 110 by 2028: about 1.3 times a year.

1.35× a year

Epoch AI: the performance of AI chips doubled every 2.3 years from 2010 to 2023, about 1.35 times a year.

The gap in 2035

550 times

AI would want about 550 times as many chips as the machines could make

Chips AI would want in 20357,600 times 2025

Chips the machines could make in 203514 times 2025

AI’s demand for computing grows 3.3 times a year. Each chip does 1.35 times more each year, so the number of chips AI would want grows 2.44 times a year. The machines, and the chips they can make, grow 1.3 times a year. So the gap grows 1.88 times a year, passing 2 times in 2027 and 10 times in 2029, and reaches about 550 times by 2035.

Year by year, against 1 in 2025
YearChips AI would wantChips the machines could makeGap
202511None
20262.41.31.9×
20276.01.73.5×
2028152.26.6×
2029362.913×
2030873.724×
20312104.844×
20325206.383×
20331,3008.2160×
20343,10011290×
20357,60014550×

This isn’t a forecast. A big gap can’t actually open, because AI can’t use chips that were never made: when demand outruns supply, prices rise, buyers wait and AI’s growth slows to what can be made, or more machines and factories arrive. Both sides start level in 2025, though memory was already running short by the end of that year. The machines are ASML’s standard EUV machines; how many chips they make also depends on the machines already working, upgrades, other tools and factory space, which the sum leaves out. The 2028 aim is reported, not confirmed by ASML.

What goes in

Three growth rates, each a multiple a year. AI’s demand for computing starts at 3.3 times, Epoch AI’s estimate of how fast the computing power of AI chips sold has grown since 2022. EUV machines made starts at 1.3 times, from ASML’s expected 65 standard EUV machines in 2026 to a reported aim of more than 110 by 2028. Computing per chip starts at 1.35 times, from Epoch AI’s finding that the performance of AI chips doubled every 2.3 years from 2010 to 2023.

What the result means

Both sides start at 1 in 2025. The chips AI would want grow by the demand rate divided by the gain per chip, and the chips the machines could make grow by the machine rate. The gap is one divided by the other.

At the starting figures, the number of chips AI would want grows about 2.44 times a year and the machines grow 1.3 times, so the gap grows about 1.88 times a year. It passes 2 times in 2027 and 10 times in 2029, and reaches about 550 times by 2035. With no gain per chip it reaches about 11,000 times. The gap only stays shut when the chips AI wants grow no faster than the machines.

It’s a simple illustration of compounding, not a forecast. A gap that size can’t actually open: when demand outruns supply, prices rise, buyers wait and AI’s growth slows to what can be made, or more machines and factories arrive. Both sides also start level in 2025, though memory was already running short by the end of that year.

The figures behind it

Epoch AI measures AI computing as quarterly AI chip sales from Nvidia, Google, AMD, Amazon, Huawei and others, counted in Nvidia H100 equivalents: 3.3 times a year since 2022, with a likely range of 2.7 to 4.1 times.

ASML recognised 44 standard EUV machines in 2025, up from 42 in 2024, and expects about 65 in 2026. It plans at least 80 in 2027 and about 30% more capacity in 2028. Reuters, citing JPMorgan analysts, reported in September 2026 that ASML is exploring more than 110 a year by 2028; ASML hasn’t announced that, and the report doesn’t say whether it counts only standard machines. Growing from 65 to 110 in two years is 1.3 times a year. ASML’s own plan, about 104 in 2028, gives 1.26.

Epoch AI’s chip figure is the peak computing performance of 47 AI chips at one level of precision (FP32), which doubled every 2.3 years: 1.35 times a year. Moving to smaller number formats has given chips bigger one-off jumps.

The sum treats the chips the machines can make as growing in step with new machines. It leaves out the machines already working, upgrades, other tools and factory space. Figures were checked on 23 October 2026.

Sources

AI chip shortage AI compute growth Chip demand calculator ASML EUV Epoch AI Memory shortage Exponential growth

More tools to try