MANY MINDED · THE COST CURVES · KNOWLEDGE

How much compute does it take to train a frontier AI model?

Frontier AI training compute, FLOP · FLOP · measured 2010–2025 · fitted live from 10 observations by the Many Minded cost-curve engine.

The last measured value is 5.0e+26 FLOP (2025, Epoch AI notable-models dataset). Over the fitted window the series rises 361.8% a year (80% interval +204.0% to +601.5%) — a doubling every 0.5 years (80%: 0.4–0.6). Carried forward on that fit, 2026: 2.3e+27 FLOP (80%: 4.3e+26 FLOP–1.2e+28 FLOP).

Estimate-grade series. This one is constructed, not read off an authoritative sheet — largest training compute among AI models published each year — the scaling frontier. Treat the level as an estimate and the slope as the claim.

5.0e+26 FLOPlatest measured (2025)+361.8%fitted, per year0.5yper doubling10observations, 2010–2025estbasis

the curve

frontier AI training compute — measured history and fitted projectionfrontier AI training compute: measured 2010–2025 in FLOP, last value 5.0e+26 FLOP, fitted at 361.8% a year, projected to 2033 inside an 80% interval. Logarithmic vertical axis.FLOP1.00e+5T1.00e+7T1.00e+9T1.00e+11T1.00e+13T1.00e+15T1.00e+17T201020132016201920222025202820315.0e+26 FLOP (2025)
Vertical axis: FLOP, log scale. Solid: measured observations. Dashed: the fitted central projection. Shaded: the 80% interval, which widens with horizon because shocks accumulate.

what this series measures

largest training compute among AI models published each year — the scaling frontier.

what the fit says

quantityvaluehow it is computed
fitted rate+361.8%/yr (80%: +204.0% to +601.5%)Farmer–Lafond drift over the trailing window · n=10 points, 15 years
doubling time0.5 years (80%: 0.4–0.6)implied by the fitted drift
curve checkdecelerating (375% → 295%/yr)first half of the record versus the second, judged against the direction that helps this metric (rising)

Not shown for this curve, because the machinery returns nothing: Wright's law (no cumulative-deployment series for this technology); a regime break (needs ≥12 observations).

the projection, with its interval

Log cost as a random walk with drift: the forecast variance grows with horizon (τ + τ²/m), which is why these bands widen instead of staying parallel. The middle column is the least useful number on this page; the interval is the claim.

yearcentral fit80% interval
20262.3e+27 FLOP4.3e+26 FLOP to 1.2e+28 FLOPnear horizon
2027above 5.0e+27 FLOPpast the point where a number would be theater

The table stops at 5.0e+27 FLOP — a decade above the highest value ever observed here. Past that, quoting a number would be theater rather than forecast.

questions this page answers

How much compute does it take to train a frontier AI model?

5.0e+26 FLOP as of 2025, the latest measured value in the series (Epoch AI notable-models dataset). The fitted trend has it rising 361.8% a year, with an 80% interval of +204.0% to +601.5%. This is an estimate-grade series: the level is constructed, so read the slope as the claim.

How fast is frontier AI training compute rising?

+361.8% a year over the fitted window, an 80% interval of +204.0% to +601.5% — a doubling every 0.5 years (80%: 0.4 to 0.6 years). Fitted from 10 observations spanning 15 years.

What will frontier AI training compute be in 2026?

The central fit says 2.3e+27 FLOP, inside an 80% interval of 4.3e+26 FLOP to 1.2e+28 FLOP. The interval is the forecast; the middle number is only its midpoint. Bands widen with horizon because shocks accumulate — a constant-width band would be overconfident.

Is the rise in frontier AI training compute accelerating or slowing?

Splitting the record in half, the fitted rate went from 375% to 295% a year — decelerating. Regime breaks, not window choice, are what dominate this method's errors — which is why every projection here carries an interval.

Where does this data come from?

Epoch AI notable-models dataset. 10 observations spanning 2010–2025. The series is estimate-grade — constructed from vendor sheets and practitioner reports, and labeled as such wherever it appears. Both the observation ledger and the fitting code are public, and the engine publishes its own calibration score and its misses.

the raw numbers

Every observation behind the fit, unrounded by us and unsmoothed. This table is here on purpose: graphs make people underestimate exponential change, and the raw series beside the curve is the one correction shown to work.

yearFLOPchange
20105.4e+16 FLOP
20143.0e+20 FLOP+763.3%/yr
20166.6e+21 FLOP+369.0%/yr
20191.1e+23 FLOP+153.9%/yr
20203.1e+23 FLOP+190.7%
20212.1e+24 FLOP+552.9%
20222.7e+24 FLOP+33.7%
20235.0e+25 FLOP+1724.8%
20243.8e+25 FLOP−24.0%
20255.0e+26 FLOP+1215.8%

where this comes from

Other KNOWLEDGE curves:

Recent briefs from KNOWLEDGE, the daily record of what actually moved:

every KNOWLEDGE brief →

Every tracked curve: utility solar, installed · battery pack price · battery cell price · solar module price · human genome sequencing · onshore wind electricity · solar electricity (LCOE) · space launch, to LEO · wheat yield · offshore wind electricity · modular housing, turnkey · AI video, finished minute · humanoid robot, capable unit · cheapest humanoid, list price · farm labor share · healthcare spend per person

← the whole engine: build clock, cascade, calibration · the archive · how we know