FINDING
The Compute Disclosure Blackout: Why the Trend Line Says 6×1027 FLOP But No 2026 Model Shows It
Fitting a standard log-linear regression to every record-setting ("frontier") AI model's publicly disclosed training compute since 2018 produces a clean trend: compute has doubled roughly every 5.2 months (R² = 0.98 — a strong fit). Extrapolated forward, that trend predicts frontier compute should reach roughly 6.2e+27 FLOP by the end of 2026.
But the actual data tells a stranger story: the highest publicly disclosed training-compute figure for any model remains Grok 4 (xAI, 2025-07-09) at 5.0e+26 FLOP — over a year old. Every model released since, including this month's frontier systems, either discloses a lower figure or none at all. That is not evidence that scaling stalled — independent benchmark scores (see our AI Model Scoreboard) show clear capability gains through 2026. It is evidence that frontier labs have simply stopped publishing exact compute figures, most likely for competitive reasons. The forecast number above should be read as "what the pre-2025 disclosure trend implies," not as a claim about what any specific 2026 model actually used.
*Extrapolation from the pre-2025 disclosure trend — see finding above. Not a claim about actual 2026 model compute.
METHODOLOGY
How This Analysis Was Built
Data source: Epoch AI's public "Notable AI Models" dataset (epoch.ai), downloaded 2026-07-29. 8,330 total model entries; 470 had both a publication date and a disclosed training-compute figure usable for this analysis.
Method: for each model with disclosed compute, we identify the "frontier" subset — every model that set a new all-time compute record at its release date (16 such models since 2018, the start of the modern deep-learning scaling era). We fit an ordinary least-squares regression of log₁₀(compute) against time on this frontier subset. The slope gives a doubling time; extrapolating the fitted line gives the forecast figures above.
Why this method, and its limits: record-setting models (not the average of all models) are the standard way to measure the frontier, matching Epoch AI's own published methodology. The R² of 0.98 indicates the pre-2025 trend was genuinely log-linear, not cherry-picked. The method's core limitation is exactly what the finding above describes: it can only fit on disclosed figures, and disclosure has become sparser for the newest frontier systems — so recent-year forecasts should be read as trend extrapolation, not measurement.
What we did not do: we did not estimate undisclosed compute figures for any 2026 model, and we did not adjust the regression to force-fit a particular narrative. The chart shows real data points only.