Why AI Companies Burn Billions — and Why They're Still in Loss

The game they are playing

OpenAI's leaked 2025 financials show $13.07 billion of revenue against $34 billion of costs. Anthropic's IPO filing shows $4.6 billion of 2025 revenue and an $8 billion operating loss — and a pledge to spend $518 billion on infrastructure over ten years. Around them, the hyperscalers moved from roughly $410 billion of capital spending in 2025 toward $730 billion planned for 2026.

Nobody is spending this way by accident. The money burns for three deliberate reasons:

The race never pauses. A frontier model is only worth anything until someone trains a better one. So each lab must fund the next training run before the last one pays off — OpenAI's research line alone hit $19.18 billion in 2025, more than all its revenue, including $10.59 billion paid to Microsoft for compute. Stopping the race to become profitable means becoming obsolete. This is the core logic: *lose money on purpose, fast, to own the technology everyone will need.*

Serving users costs a fortune. Every chat message and API call runs on GPUs that burn electricity and were bought at the top of a shortage. OpenAI's cost of revenue — mostly inference — grew from $2.65 billion to $7.5 billion in one year, scaling directly with success. Video generation is worse: reporting put OpenAI's Sora at roughly $1 million a day of inference against $2.1 million in lifetime revenue. Every free chatbot is subsidized, which makes "free" the most expensive marketing in corporate history.

Talent is priced like a lottery ticket. Top researchers command nine-figure packages, and stock-heavy compensation is how labs compete with each other and with their own investors.

Add the sales push — OpenAI spent $5.73 billion on sales and marketing in 2025, up fivefold in a year — and the burn is structural, not waste. It is the bill for trying to win a market before the market exists.

So why the loss, if revenue is exploding?

Here is where the numbers get strange. OpenAI tripled revenue in a year. Anthropic grew twelvefold. And yet the losses got bigger, not smaller. Three reasons:

Revenue is growing slower than spending. OpenAI's operating loss went from $8.78 billion in 2024 to $20.92 billion in 2025. It did improve as a share of revenue — from 237% to 160% — meaning each dollar earned loses less than before. But the finish line only exists if costs someday grow slower than revenue, and every new model generation resets the cost base. The company tells investors it hopes to be profitable by 2030; The Information reported a projected $115 billion of burn through 2029.

Net losses are not cash. Anthropic's headline number — a net loss of $80.7 billion in the first half of 2026 — included a $72 billion accounting charge on convertible bonds. Those bonds gain value *because the company is doing better*: investors expect the shares to be worth more, and accounting rules book that expected gain as a loss. The actual cash spent on infrastructure in those six months was about $9 billion, against $16.8 billion of revenue in the latest quarter and a claimed adjusted profit. The scarier-looking number is partly the smell of success.

The losses are financed on purpose. Labs raise tens of billions from Microsoft, Nvidia, and sovereign-scale funds precisely to cover these losses — the way Amazon ran unprofitable for its first decade while buying the market. The loss is not an accident happening to these companies; it is the strategy investors are underwriting. What would actually end the party is not the loss itself but the financing drying up.

Is this the dot-com bubble again?

The rhymes are real: infrastructure built ahead of demand (today's data centers, like the 1990s' dark fiber), circular financing (Nvidia invests in OpenAI, whose data centers buy Nvidia chips — a deal web above $750 billion by 2026, echoing telecom equipment makers financing their own customers in 1999), and valuations priced on transformation rather than earnings.

Two differences matter. The buildout is funded by the most cash-rich companies in history from actual profits — the IMF's chief economist expects a possible bust, but not a systemic banking crisis like 2008. And real revenue exists: $20 billion annualized at OpenAI, $16 billion a quarter at Anthropic, versus dot-coms whose revenue was a rounding error.

The honest reading: a bubble in *investment and expectations* sitting on a genuine technological shift — exactly what 2000 was. The Nasdaq needed 15 years to recover, most dot-coms died, and the survivors changed the world.

Real use versus hype

MIT's NANDA initiative found roughly 95% of enterprise generative AI pilots produced no measurable profit impact despite $30–40 billion of spending. That is the hype. The 5% that works is boring and verifiable:

• Writing code — field experiments, including Microsoft's own research, show real productivity gains, and coding assistants were the first AI products with organic enterprise demand.
• Customer support — companies like ClassPass quietly route most routine inquiries to AI. Measurable, unglamorous, real.
• Search, drafting, summarization — a good-enough first draft that saves an hour.

The gap between that reality and the "transform everything" pitch is exactly where the unpriced losses live.

The verdict

AI companies burn money because the game rewards spending faster than the market rewards earning — pausing the race means losing it. They are in loss because costs reset upward with every model generation while revenue, though exploding, started from nearly nothing. And the headline loss numbers overstate the danger: Anthropic's $80 billion was mostly accounting, OpenAI's true burn is the operating line.

The bet is that revenue keeps doubling until it catches costs. Anthropic's numbers say that bet is not crazy. OpenAI's cost lines say it is not safe either. The dot-com comparison, correctly told, ends with a question, not a prophecy: when this settles, who is Amazon — and who is Pets.com?