Methods · Model

Frontier training compute growth (doubling time)

Doubling time of frontier models' training compute, fitted log-linearly against publication date from 2020 on Epoch's frontier-model table. Derived metric training_compute_doubling_days.

Reason through this

The reading

Fitted on Epoch's estimates, most of which are inferred for closed models; the newest models are missing until Epoch estimates them.

Every readingeach dot links to its record

days

a readingnormal rangefast range

estimate

1 reading drawn from the derived metric training_compute_doubling_days; 67 observations in the evidence.

Derived (training_compute_doubling_days) from: Epoch AI notable and frontier models. Epoch AI, Notable AI Models (CC BY 4.0).

The numbers

as of	reading	value	record
2026-09-03	-11.4 39 0.898	170 days	estimate

Why it matters

The invention stage's own speed limit. The brief's reading is 4-5x a year; a slowdown here is the first thing the data-and-compute constraint would show.

What cuts against it

Compute estimates for closed models carry wide uncertainty; the frontier set is Epoch's selection; training compute is an input, not a capability.

How it is measured

derived metric training_compute_doubling_days (formula in the semantic layer)
source Epoch AI notable and frontier models · default tier 6 · CC BY 4.0
proxy benchmark, model_release · cadence annual · valve invention to product

Normal = doubling no faster than every 300 days (about 2.3x a year, the pre-2010 pace of compute growth); fast = every 182 days or less (4x a year or more, Epoch's 2010-2025 trend). Between is emerging.

Applied to metric:training_compute_doubling_days.

Derived rows (1)

as of	dims	value	inputs
2026-09-03	-11.4 39 0.898	170 days	sources1 2 3+64 more

Latest readings

2026-09-03 · 170 daysas of 2026-09-03(67 obs)

Why this status, and its history

Log-linear fit on Epoch's 38 frontier models published since 2020: training compute doubles every 170 days (95% CI 152-193), inside the fast band (182 days or less, 4x a year or more) and matching Epoch's own 4-5x reading. Single tier-6 source; the estimates are inferred for most closed models, so the status is capped at emerging. Latest frontier model with an estimate: July 2025. Initial seed.

2026-09-10unmeasured to emergingconf — → 60 · the evaluator (automatic)
Log-linear fit on Epoch's 38 frontier models published since 2020: training compute doubles every 170 days (95% CI 152-193), inside the fast band (182 days or less, 4x a year or more) and matching Epoch's own 4-5x reading. Single tier-6 source; the estimates are inferred for most closed models, so the status is capped at emerging. Latest frontier model with an estimate: July 2025. Initial seed.
sources1 2 3 4 5 6+60 more

Why this confidence

Rests on 67 observations from Epoch AI notable and frontier models; best evidence tier 6, grade B; current.

Confidence is independent of status: 90–95 multiple strong independent sources; 70–89 good evidence, some ambiguity; 50–69 mixed or hard to operationalise; 0–49 limited or vague evidence.

Predictions that lean on this

Happening
Gwern argued that once a scalable architecture exists, simply training ever bigger neural networks will keep yielding more sophisticated general abilities without fundamentally new designs. · Gwern
Slower than said
Ray Kurzweil forecast that before 2030 the combined thinking capacity of non-biological intelligence would exceed that of all human brains put together. · Ray Kurzweil
Too early to tell
Ajeya Cotra's biological anchors report modelled a median in the early 2050s for transformative AI, which she rounded to about 2050, with most aggressive and most conservative plausible medians of 2040 and 2080. · Ajeya Cotra
Too early to tell
Tom Davidson's takeoff model put a median of 2043 on AI able to readily automate all cognitive work, with about one chance in ten by 2030 and four in five by 2070. · Tom Davidson
Too early to tell
Ege Erdil and Tamay Besiroglu doubt that AI improving its own software alone will set off an intelligence explosion, holding that compute and experiments bound progress. · Ege Erdil and Tamay Besiroglu

More predictions that lean on this

Too early to tell
Leopold Aschenbrenner projected a trillion-dollar training cluster by 2030, drawing about a hundred gigawatts, more than a fifth of current US electricity output. · Leopold Aschenbrenner
Too early to tell
Yafah Edelman and Anson Ho of Epoch AI argue that longer data-centre build times push the first trillion-dollar AI compute cluster from about 2030 back to about 2035. · Yafah Edelman and Anson Ho

Reading

emergingA model's judgement, not a reading: probably faster than normal. The fitted reading on the site has frontier training compute multiplying a little faster than the fast threshold, so it was read as fast, in line with the long-run trend since 2010. It is close to the edge, though, and the newest closed models are missing or inferred with wide error. The strongest alternative is that growth has eased as single training runs hit power and chip limits.

Latest

170 daysas of 2026-09-03(67 obs)

What the rule reads

170 days(153 days–191 days)as of 2026-09-03(67 obs)

The rule

Normal ≥ 300 days; fast ≤ 182 days. Between them it reads emerging.

Confidence 60 of 95: mixed or hard to operationalise. grade B

Leading: training compute sets the capability of models released months later and deployed years later.

Updated 2026-09-11