OpenAI's Next Raise Assumes a Number Nobody Will Say Out Loud

There is a figure in OpenAI's fundraising model that everyone quotes and no one underwrites out loud.⁶ The next round, per people who have seen the deck, is priced on a revenue curve that only clears if two things happen at once¹: inference keeps getting cheaper, and usage never slows down.

Two assumptions, one company

Each half is plausible alone. Together they are the whole company. Cheaper inference is what lets usage grow; growing usage is what justifies the spend that makes inference cheaper.³ The model assumes the flywheel never skips.

The trouble is the denominator

The trouble is the denominator. Compute cost per token has fallen fast⁴, but the frontier model that drives the headline demand is the most expensive one to serve, and it is the one enterprises actually deploy.

The ad business is supposed to fill the gap

The ad business, barely six months old, is supposed to fill the gap. This week OpenAI narrowed which partners can run ads on its surfaces², a move that reads as protecting margin more than expanding it.

Rivals are circling the same enterprise dollar with a reliability pitch, and state attorneys general are circling the ad model.⁵ The raise has to close before either circle tightens.