Training power draw growth (doubling time)

Doubling time of notable models' estimated training power draw, fitted log-linearly from 2020 on Epoch's notable-model table. Derived metric training_power_doubling_days.

Why it matters. Power is the physical constraint that compute growth eventually hits; the brief's reading is 2.2x a year.

Proxy types

model_release

Unit

days

Cadence

monthly

Valve

invention to product

• derived metric training_power_doubling_days (formula in the semantic layer)

Estimated from hardware counts and thermal design power; a proxy for the buildout's physical footprint, not a metered figure.

213 observations. Hollow points are disputed (see counterevidence). Every point links to its observation.

Log-linear fit on 168 notable models' estimated training power draw since 2020: doubling every 506 days (CI 332-1067), inside the normal band (yearly or slower), but the fit explains almost none of the variance (r2 0.08) because the notable set mixes small and frontier models. Single tier-6 source and a weak fit: emerging. Initial seed.

• 2026-08-07 · 506 daysas of 2026-08-07(213 obs)

1. 2026-09-10○unmeasured →to ◐emergingconf — → 50 · evaluate

   Log-linear fit on 168 notable models' estimated training power draw since 2020: doubling every 506 days (CI 332-1067), inside the normal band (yearly or slower), but the fit explains almost none of the variance (r2 0.08) because the notable set mixes small and frontier models. Single tier-6 source and a weak fit: emerging. Initial seed.

   obs:0064d6a3obs:0345be86obs:075cc617obs:085ec98aobs:0b75879dobs:0d3630dc+207 more

50 / 95 — mixed or hard to operationalise

Confidence is independent of status: 90–95 multiple strong independent sources; 70–89 good evidence, some ambiguity; 50–69 mixed or hard to operationalise; below 50 limited or vague.

• Frontier training compute growth (doubling time) ◐emerging
• Circular financing scale ▲concentrating
• Crosswalk: Methods ⇄ Compute & physical (input to)