Epoch AI

Type: Organization (independent research group that tracks and forecasts AI scaling)

Epoch AI publishes the most-cited estimate of when training will use up the public stock of human-written text. On this rung it matters for two reasons: the estimate itself (the ~300T-token effective stock fully used between 2026 and 2032, central 2028), and the fact that Epoch moved it — from "before 2026" in 2022 to 2028 in 2024 — and said why in public (Epoch AI).

Its August 2024 report Can AI scaling continue through 2030? weighs four limits on training — power, chip manufacturing, data and latency — and concludes that ~2e29 FLOP training runs are likely feasible by 2030, with power the limit most likely to bind first, then chips; data comes after both.

Why the source is graded as it is

Epoch's figures are forecasts built from explicit, published assumptions (stock size, growth, quality filtering, reuse). That makes them checkable, and also revisable — as the 2022-to-2024 move shows. The 2030 report and the data paper share authors, so they are one group's view, not two independent ones. Graded moderate.

Mentioned In

• The data stock and when it runs out

Related Entities

• Reddit — one of the sellers who gain if free public text stops being enough