OpenAI's Revenue Story Has a $20 Billion Problem

The number everyone quoted was $70 billion. OpenAI told its investors closer to $50 billion. That gap isn’t a rounding error. It shows how far the story ran ahead of the business.

Read the attribution carefully, because the FT does. According to the Financial Times, OpenAI recently told investors its annualized revenue was “approaching $50bn” in September. The $70 billion was the figure that had been widely reported, and the FT’s headline calls the new number less than “previously signalled.” Who did the signalling is the question people will fight over. It doesn’t change the thesis. Whether the $70 billion came from a pitch, a leak, or a game of telephone, it was the number in circulation. Twenty billion dollars, gone. Not from the business. From the story told about the business.

Here’s the part that holds either way. Nobody had a reason to puncture $70 billion. Bulls didn’t. Secondary buyers didn’t. Founders raising on “we’re the next OpenAI” comps didn’t. A number that flatters everyone survives a long time, no matter where it started.

This matters more than the headline. If you priced OpenAI paper off the $70 billion everyone was quoting, swap in $50 billion and the revenue multiple on that same price jumps 40%. That’s just arithmetic. Then you need a very specific set of assumptions to make the price work: margins expanding, compute costs falling faster than they have, enterprise lock-in deepening before Microsoft or Anthropic or Google eats the margin. At $70 billion, those assumptions get a lot more comfortable. The $20 billion gap is the difference between “expensive but defensible” and “everything has to go right.”

Everyone says OpenAI is winning the AI race. In one dimension, the opposite is closer to true: OpenAI is winning the narrative race, which is a different race with different stakes. The market has blurred ChatGPT engagement with enterprise revenue durability, and growth rate with margin trajectory. None of that is a lie. All of it is optimism. And when the real number shows up, the gap between the framing and the fact is exactly what turns a growth stock into a busted story.

To be clear: $50 billion in annualized revenue is extraordinary. The number isn’t the problem. The problem is that the valuation needs the story to be bigger than the number. Same price, more than a quarter less revenue underneath it, for a company with massive compute costs, a Microsoft dependency that gives Redmond the best seat in the house, and a competitive field that keeps getting more expensive to stay ahead of. That’s a lot more faith to ask of the same price.

Look at how Anthropic competes. This week it launched Claude Haiku 5.5 at exactly GPT-6 Luna’s price: $0.10 per million input tokens, $0.50 per million output. The parity only holds up to 100,000 tokens. Past that, Haiku 5.5’s price jumps 5x to $0.50/$2.50, while Luna’s step-up comes later, at 272,000 tokens, and is milder at $0.20/$0.75. For long-context work, Luna is the cheaper model.

Even that isn’t the whole story. Haiku 5.5 also has a new tokenizer, and Simon Willison found the same long prompt uses about 1.25x as many tokens on it as on Haiku 4.5, which he calls a hidden price increase. The price sheet doesn’t show it. Same lesson, smaller scale: the headline number and the real number are different things, and you only find the gap if you go looking. If the most checkable numbers in AI still need a footnote, the uncheckable ones deserve a discount.

The model market is compressing. Haiku versus Luna looks like the cloud price wars: relentless margin pressure, fast commoditization, and winners decided by distribution and switching costs, not raw capability. OpenAI’s moats are ChatGPT’s consumer mindshare and its enterprise sales motion. Both are real. Neither is infinite.

The stakes are structural. Anyone holding OpenAI paper valued against the $70 billion figure now has to re-anchor to $50 billion. That doesn’t just change a spreadsheet. It changes the story every board deck, every secondary trade, and every future fundraise has to tell. Once people start marking to the real number, the reflexivity runs in reverse.

You can raise at a great story, or you can raise at a great business. The trick is knowing which one you’ve got before the market figures it out first.