Skip to content
Slow Variables
Ask

Model

Frontier lab gross margin (estimated)

OpenAI's full-year revenue minus its inference compute spend, over revenue, from Epoch's estimates (derived metric lab_gross_margin_estimate, grade C). Anthropic's figure is on the metric's saved analysis; OpenAI is the headline because it has more than one year.

Reason through this

The reading

An estimate, not a filed margin: revenue minus inference compute only, so it overstates a true gross margin. Read the direction, not the level.

Direction over the last 2 readingseach dot links to its record
openai · 51.4% · 2024-12-31 · estimateopenai · 38.5% · 2025-12-31 · estimate
in the rule's windowmoves within ±5.0 pts read as stable, and a move counts only if more than half the steps go its way; this window moved −12.9 pts and 1 of 1 steps fell; higher reads concentrating
estimate

2 readings drawn from the derived metric lab_gross_margin_estimate; 4 observations in the evidence.

Derived (lab_gross_margin_estimate) from: Epoch AI data hub — AI companies. Epoch AI, 'Data on AI Companies' (CC BY 4.0), epoch.ai/data.

The numbers
as ofreadingvaluerecord
2025-12-31openai38.5%estimate
2024-12-31openai51.4%estimate

Why it matters

Whether the model layer keeps a software margin or a cloud margin decides whether it is a chokepoint or a commodity.

What cuts against it

Every input is one estimator's (Epoch, tier 5), so the status cannot rise above emerging; revenue minus inference compute leaves out serving staff, data and hosting; reported margins from journalists differ and sit behind paywalls.

How it is measured
  • derived metric lab_gross_margin_estimate (formula in the semantic layer)
  • source Epoch AI data hub — AI companies · default tier 5 · CC BY 4.0
  • proxy price · cadence annual

Year on year; a move under five points is noise. A rising margin means the lab layer keeps more of what it sells.

Applied to metric:lab_gross_margin_estimate.

Derived rows (2)
as ofdimsvalueinputs
2025-12-31openai38.5%sources12
2024-12-31openai51.4%sources12
Latest readings
  • 2025-12-31 · 38.5%as of 2025-12-31(2 obs)
  • 2024-12-31 · 51.4%as of 2024-12-31(2 obs)
Why this status, and its history

First scoring. OpenAI's revenue minus inference compute, over revenue, fell from 51% in 2024 to 38% in 2025 on Epoch's estimates, a move past the five-point dead band that would read dispersing; every input is one estimator's (Epoch, tier 5), so the single-source rule holds the status at emerging until a second, independent source for the same measure exists, not until more years accrue.

  1. 2026-09-27unmeasured to emergingconf — → 30 · Claude (model)

    First scoring. OpenAI's revenue minus inference compute, over revenue, fell from 51% in 2024 to 38% in 2025 on Epoch's estimates, a move past the five-point dead band that would read dispersing; every input is one estimator's (Epoch, tier 5), so the single-source rule holds the status at emerging until a second, independent source for the same measure exists, not until more years accrue.

    Weighed against it: The estimate leaves out serving staff, data and hosting, so it overstates a true gross margin; a falling ratio can also mean a lab is pricing to win share while inference volumes grow, not that it is losing pricing power.

    sources12
Why this confidence

Rests on 4 observations from Epoch AI data hub — AI companies; best evidence tier 5, grade C; 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