AI Frontier
Everywhere else in FutureGrid you see AI's impact on jobs. Here you see the raw training-compute records behind it — an unrelenting exponential drawn from Epoch AI's tracked catalog.
Compute doubling time
~5.7 months
Modern era 2010–present · r²=0.65
The small trendline is decorative; the labeled figure above is the reported value.Compute-known records
528
Models with reported or estimated training compute
Top country by recent tracked releases
United States
189 recent tracked · 664 total dated
Largest reported training run
500 YFLOP
Grok 4 · xAI · 2025
The small trendline is decorative; the labeled figure above is the reported value.Training Compute Over Time
Each point is a model from Epoch AI's compute-known subset — records with reported or estimated training FLOPs. Frontier models (Epoch's historical top 10 by training compute at release) are highlighted. The trend line shows the fitted modern-era exponential.
Compute doubles every ~5.7 months (modern era, r²=0.65, n=466)
Where Tracked Models Are Developed
A share-of-records view of Epoch AI's tracked models by attributed country. Each tile's area is that country's share of all tracked records for the selected metric — showing how concentrated tracked development is across the catalog, not a ranking of national AI capability, output, or impact.
Tracked Model Activity by Organization and Country
Rankings reflect Epoch AI's tracked catalog — not general AI capability, product adoption, or economic impact. The default sort is recent tracked releases (past 3 years). Compute and frontier metrics apply only to the compute-known subset and are biased toward organizations that disclose training compute. Org entities are preserved as recorded in the source; no editorial consolidation is applied.
Training Cost & Power Draw
Frontier training is becoming exponentially more expensive and energy-intensive. Median and peak costs shown in 2023 USD; power in Watts.
Model Landscape
Why This Drives Workforce Disruption
A doubling of AI training compute every ~5.7 months means that capabilities which seemed out of reach just two years ago are standard today. This pace — sustained since 2010 and still accelerating — has coincided with the displacement signals FutureGrid tracks across labor markets, sectors, and occupations. It is not driven by one company or one country; it is a systemic, global scaling of machine intelligence.
Capability expands faster than adaptation
When compute doubles faster than organizations can retrain workers or redesign workflows, the gap between what AI can do and what jobs currently require has continued to widen.
Compute disclosure reflects concentration
The compute-frontier records in Epoch AI's catalog are heavily concentrated among organizations that report training compute — primarily large, well-capitalized labs. This reflects a disclosure pattern as much as a capability pattern; many active AI developers worldwide do not publish compute figures, so compute and frontier metrics systematically undercount non-disclosing participants.
Cost and energy amplify the stakes
Reported or estimated peak training runs in the source have reached $387.8M and 109.9 MW. Only well-capitalized entities can push the compute frontier — further concentrating the advantage among those who disclose it.
Data Attribution
Source caveats
- Training-compute figures are Epoch AI estimates with varying confidence; not all models report compute/cost/power.
- Coverage is not exhaustive; Epoch curates 'notable' models.
- Compute estimates carry uncertainty (see Confidence field).
- Future-dated entries reflect the source snapshot and are shown as-is.
- Descriptive trends, not predictions.
- Multi-country collaboration models are co-attributed to each participating country in the country leaderboard.
- Epoch AI's 'Frontier model' flag reflects top-10 training compute at release time, not capability or impact.
- Org/country model counts use the full dated catalog. Compute-derived metrics (frontierCount, maxComputeFlop) use only the compute-known subset.
- openWeightsCount is a proxy for tracked open-release activity only.