Revenue is on metric. Another important metric is gross margin.
OpenAI's gross margin is estimated at 33%[0]. They also have to pay Microsoft 20% of revenue.
So, for each $1 of revenue:
Revenue $1.00
COGS (inference) (0.67)
Microsoft (0.20)
Gross margin $0.13
That $0.13 must cover "everything else": R&D, payroll, etc., and ideally leave some profit on the table.
The problem for OpenAI (and other pure AI companies) is that inference is not like software that sells at marginal cost (build once, sell everywhere), but each token costs money. Inference gets cheaper, but newer models require more computing power and consume more tokens. So the gross margin does not improve over time.
Break-even in the future won't come from just growing API usage and subscriber base.
$20b is December 2025 revenue multiplied by 12. Actual 2025 annual booked revenue is more in the order of $13b[0].
Also, those $122b raised are not cash-in-hand.
For example, NVIDIA commits to $30b, but OpenAI must build 5gw inference+training capacity using NVIDIA's Ruby Vera systems. NVIDIA's Huang has said[1] that in 1gw of compute, around $35b are NVIDIA hardware. So that $30b investment from NVIDIA goes back to NVIDIA as $175b in revenue. (Besides, NVIDIA gets non-voting shares in OpenAI.)
Deals like these are one of the few ways OpenAI can sustain growing in capacity, but it comes at a huge premium. That's one of the reasons they need the IPO, to get access to cheaper money.
Thanks to Twitter recent storms, many people are migrating to Mastodon. As the Mastodon user numbers go up, many servers will have to face the question of how to pay the bills for hosting and other costs.
Revenue $1.00
COGS (inference) (0.67)
Microsoft (0.20)
Gross margin $0.13
That $0.13 must cover "everything else": R&D, payroll, etc., and ideally leave some profit on the table.
The problem for OpenAI (and other pure AI companies) is that inference is not like software that sells at marginal cost (build once, sell everywhere), but each token costs money. Inference gets cheaper, but newer models require more computing power and consume more tokens. So the gross margin does not improve over time.
Break-even in the future won't come from just growing API usage and subscriber base.
[0] https://sacra.com/c/openai/