Infrastructure · Trend

Collapsing inference cost

Price per token for a given capability keeps falling by an order of magnitude a year.

Trend strength 7/10

Momentum +3/q

Confidence high

Status rising

Forecast horizon

Cheap intelligence diffuses into every product; volume, not access, becomes the differentiator.

Connections

Connections · 2

How this node ties into the rest of the map, and the evidence behind each link.

Cheap inference makes long agentic runs affordable.

Smaller models slash cost per capability.

Signal sources

Dated facts from primary sources in this direction.

Training compute for frontier AI models grew ~4–5× per year from 2010 to 2024 — a trend Epoch projects to continue toward 2030.

Global data-centre electricity use is projected to roughly double from ~485 TWh (2025) to ~950 TWh by 2030 — about 3% of world electricity.

The price to query a GPT-3.5-level model fell from $20.00 to $0.07 per million tokens between Nov 2022 and Oct 2024 — a 280× drop in ~18 months.

DeepSeek disclosed its open-weight R1 reasoning model's final training run cost just $294,000; R1 was released under open weights on 20 Jan 2025.