|
OpenAI burned $3.7 billion in the first quarter of 2026, on $5.7 billion of revenue, according to documents the company shared with investors and reported by The Information. Revenue and burn both roughly tripled from a year earlier.
The number that traveled was the $3.7 billion. It is large, and “burned” is an alarming word. It is also the wrong number to watch.
A company tripling its revenue and spending hard to do it is the most ordinary thing in technology. What matters is what the spending buys. Most of OpenAI’s buys models that do not exist yet.
The biggest line is training: about $32 billion this year, $65 billion next, something like $440 billion by 2030. That is enormous, and it is mostly a choice. Training is the cost of building the next model, not running the last one.
And OpenAI can pay for it. It ended the quarter with more than $73 billion in cash, up from $40 billion in December. By that measure the $3.7 billion is not burning. It is spending, which is the generous word and the accurate one.
The part that is not a choice is inference, the cost of running the models people already pay for. That cost reportedly quadrupled last year and will reach about $14 billion this year. It has dragged OpenAI’s adjusted gross margin down to about 33 percent, against an internal target of 46.
Gross margin is what is left after you pay to deliver the product, before research or salaries or anything else. For software it is usually high, and usually climbs as a company grows, because the millionth customer costs less to serve than the first.
OpenAI’s is low, and falling. Every new user makes the core business a little less profitable, not more.
So there are two burns, and they are not alike. Training is a bet OpenAI can stop making; ship nothing new and the bill drops. Inference is the cost of success, and it rises when the product works. You cannot grow out of that, because growth is what makes it worse.
It is also why there is now a market for cheaper, good-enough models: serving the best model at scale costs enough that not everyone wants to pay for the best model.
This matters now because OpenAI is about to let the public price it. It has filed confidentially for an IPO that could come as soon as September, at up to $1 trillion, the first clean public test of what a standalone frontier lab is worth. It has also told investors it will not be cash-flow positive until about 2030.
Markets know how to fund years of losses when unit economics improve with scale. They have less practice with the opposite, which is what the gross margin is quietly describing.
None of these numbers are official. They come from investor documents reported by The Information; Reuters could not verify them, and OpenAI has not confirmed them.
But if they are roughly right, the headline pointed at the wrong number. OpenAI is not in trouble for spending $3.7 billion in a quarter. It can afford that, and most of it buys the future. The question is the gross margin, the 46 that became 33, and whether the most important company in AI can sell its product for more than it costs to deliver.
|
