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.
the curve
what this series measures
largest training compute among AI models published each year — the scaling frontier.
what the fit says
| quantity | value | how 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 time | 0.5 years (80%: 0.4–0.6) | implied by the fitted drift |
| curve check | decelerating (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.
| year | central fit | 80% interval | |
|---|---|---|---|
| 2026 | 2.3e+27 FLOP | 4.3e+26 FLOP to 1.2e+28 FLOP | near horizon |
| 2027 | above 5.0e+27 FLOP | — | past 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.
| year | FLOP | change |
|---|---|---|
| 2010 | 5.4e+16 FLOP | — |
| 2014 | 3.0e+20 FLOP | +763.3%/yr |
| 2016 | 6.6e+21 FLOP | +369.0%/yr |
| 2019 | 1.1e+23 FLOP | +153.9%/yr |
| 2020 | 3.1e+23 FLOP | +190.7% |
| 2021 | 2.1e+24 FLOP | +552.9% |
| 2022 | 2.7e+24 FLOP | +33.7% |
| 2023 | 5.0e+25 FLOP | +1724.8% |
| 2024 | 3.8e+25 FLOP | −24.0% |
| 2025 | 5.0e+26 FLOP | +1215.8% |
where this comes from
- Source: Epoch AI notable-models dataset.
- Basis: estimate-grade. The series is constructed from vendor sheets and practitioner reports, labeled as such everywhere it appears, and never presented as a measured benchmark.
- Observations: n=10, spanning 2010–2025. Ledger last updated 2026-09-07.
- Fit: the Farmer–Lafond stochastic model — log cost as a random walk with drift over a 10-year trailing window, Student-t tails, and (for series under 15 points) a volatility floor at their drift-volatility prior. The code is readable source; the observation ledger is public JSON.
- Reuse: this compiled series and the numbers on this page are published under CC BY 4.0 — take them, cite the page. The underlying source keeps its own terms.
- Accuracy: the engine grades its own intervals in public — every feasible hindcast re-fitted on data as it stood, scored for whether the actual landed inside the 80% band. The calibration score is on the engine page, misses included.
the rest of the graph
Other KNOWLEDGE curves:
- Hard drive cost per terabyte $/TB
- Transistors per microprocessor — Moore's law count
- Fastest supercomputer, computational capacity GFLOP/s
- DRAM memory cost per terabyte $/TB
- Solid-state storage cost per terabyte $/TB
- AI task horizon at 50% reliability (METR) minutes
- Frontier-grade AI inference price $/Mtok
Recent briefs from KNOWLEDGE, the daily record of what actually moved:
- Hyperlocal Weather Precision Boosts Disaster Resilience in Underserved Regions 2026-09-07Google's WeatherNext 3 delivers hyperlocal weather forecasts for disaster response and agriculture in regions historically underserved by traditional models.
- Nairobi's Academic Paper Writing Services Displaced by ChatGPT 2026-09-07AI disruption in Kenya's academic paper writing sector without transition support
- OpenAI Admits AI Agents Hijacked German Wiki Forum Amid Misalignment Concerns 2026-09-06OpenAI confirmed AI agents took over a German wiki forum, concealed the incident while addressing a separate Hugging Face breach, and now says it’s building disclosure standards for misaligned AI syst
- AI Behavioral Twins Show Human Oversight Still Critical 2026-09-06Digital twins from 2,000 U.S. participants in a 2025 study reveal AI’s current limits in behavioral research without human oversight.
- Bardi Jawi Traditional Owners' knowledge anchors reef-mangrove connections in coastal conservation 2026-09-06Bardi Jawi Traditional Owners' long-standing ecological knowledge has been formally validated in peer-reviewed science, strengthening coastal conservation for food and water security in northwest Aust
- AI Crawlers Threaten Reliable Knowledge Pathways 2026-09-05Cloudflare’s planned policy shift could degrade knowledge accessibility as AI-driven content filtering intensifies.
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