Stop Me If You've Heard This One Before
Last week, audited financial statements from OpenAI leaked to the press. The numbers are staggering: $13.07 billion in revenue for 2025, against $34 billion in costs and expenses, producing a net loss of $20.92 billion. That's nearly eight times the losses the company posted the year before. Eight. Times.
But here's the part that should make you pause, and I mean really pause — not the kind of pause where you nod sagely and then open another incognito tab to play with the latest model. OpenAI is simultaneously the most valuable private AI company on earth and one of the most extraordinary money incinerators in technology history. It has raised $122 billion in a single March 2026 round, reportedly has $25 billion in cash on hand, and still managed to burn through more operating cash than most countries' GDP.
Something doesn't add up. Or rather, it adds up perfectly if you're willing to ask the uncomfortable questions.
The Revenue Story Is Impressive. The Cost Story Is Terrifying.
Let's give credit where it's due — $13 billion in revenue is not nothing. OpenAI went from effectively zero revenue in 2022 to thirteen billion in three years. That is, by any normal business standard, a miracle. ChatGPT is one of the fastest-growing software products ever built, and the API business is powering an entire ecosystem of startups that would collapse overnight if OpenAI turned off the tap.
But here is the number nobody wants to stare at directly: $34 billion in costs. That's not a company scaling efficiently. That's a company that has built its entire business model on the assumption that the cost curve bends faster than the revenue curve. And the audited statements suggest that assumption is not holding up.
The breakdown is instructive:
- Compute costs dominate absolutely everything. AI training and inference at OpenAI's scale requires clusters of GPUs that cost more to run than entire airlines. Microsoft's partnership absorbs some of this, but the raw compute bill is still measured in the tens of billions.
- Headcount has exploded without proportional output growth. OpenAI crossed 4,000 employees in 2025. Each of those employees costs the company well over half a million dollars when you factor in compensation, equity, and the San Francisco real estate that houses them.
- The API pricing war is self-inflicted. Every time a competitor drops their per-token price, OpenAI feels compelled to match or undercut. The result is a market where everyone is racing to the bottom, and nobody has a clear path to sustainable margins on inference alone.
The Uncomfortable Question Nobody Wants to Ask
Here's where the skepticism kicks in, because I think the industry has been collectively avoiding a fairly simple question: What if OpenAI never becomes profitable?
Not "what if it takes longer than expected." What if the business itself — selling access to a model that requires billions of dollars in compute to train and maintain — simply cannot generate the margins that justify its $300+ billion valuation? The leaked financials are not a lagging indicator of the past. They are a leading indicator of a structural problem.
The defense, and it's the defense you'll hear from every OpenAI bull, is that AGI will unlock entirely new markets and revenue streams that don't exist yet. True believers point to agents, to enterprise automation, to the idea that once AI can do the work of a knowledge worker, the total addressable market is measured in the trillions.
Maybe they're right. But here's what history teaches us: companies that lose $20 billion a year while telling investors "just wait until we get to the really good stuff" tend to hit a wall before the really good stuff arrives. We've seen this movie before. It was called the dot-com boom. It was called WeWork. It was called every hype cycle that convinced itself it had transcended the laws of economic gravity.
What the Leaked Financials Actually Tell Us
If you strip away the AGI mysticism and the founder narratives, the leaked OpenAI financials tell a much simpler story than the one you'll read in the tech press:
- Compute is not getting cheap fast enough. Despite the hype around inference cost reductions, training costs continue to scale super-linearly with model capability. Every order-of-magnitude improvement in model quality requires an order-of-magnitude increase in compute spend.
- The moat is thinning. OpenAI's early-mover advantage is being eroded by open-source models, cheaper competitors (DeepSeek, Qwen, Z.AI's GLM series), and hyperscaler custom silicon. The pricing power that let OpenAI charge premium rates is gone.
- The cash runway, while enormous, is finite. $25 billion sounds like a lot until you're losing $21 billion a year. The math does not require a PhD to work out.
The Verdict: Brilliant Technology, Unproven Business
Let me be clear about what I'm not saying. I am not saying OpenAI is going to fail. I am not saying the technology is overhyped. GPT-5 is genuinely impressive. The engineering team at OpenAI is probably the most concentrated pool of AI talent on the planet. The product is real, the usage is real, and the impact is real.
What I am saying is that $20.92 billion in annual losses is not a footnote. It is not a rounding error. It is not "investment in the future" in the way that Amazon's early losses were investment in the future — because Amazon was building a capital-efficient logistics machine, not renting someone else's GPUs to run a model whose marginal cost per query barely declines with scale.
OpenAI's financials are a mirror held up to the entire AI industry. The question is whether anyone is willing to look.
Comments