You probably saw the headline this week: Anthropic is telling investors it will be profitable for the second straight quarter. The AI darling is finally making real money, right? The revolution is here, and it’s paying for itself.
Not so fast.
If you read past the press release, you’ll find a tiny, magical phrase: “adjusted operating income.” This is a sort of Enron-style accounting trick that conveniently excludes some of the most massive, unavoidable expenses a business can have. Namely? Revenue sharing, the cost of model training, and hardware deployments.
Profitability without the cost of building your core product isn’t a milestone; it’s a magic trick.
Anthropic’s gross margins are reportedly above 80%—before accounting for revenue shared with distribution partners like Amazon, and before paying to train the very models that generate the revenue. This is the equivalent of a bakery claiming it’s profitable if you ignore the cost of flour, ovens, and the lease on the building.
You’ve probably noticed this playbook before. We saw it with the gig economy companies right before their IPOs. Uber and Lyft spent years trying to convince us they were “profitable” by stripping out driver payouts, depreciation, and massive R&D costs. They masked deeply unprofitable core operations with accounting tricks until they were big enough to monopolize or pivot to ad revenue.
The idea of removing model training from your costs is wild. The company is famous for training models. That is the entire moat. The core competitive value of an AI company requires continuous, massive capital expenditure on training and hardware. If you exclude the billions spent on compute, of course you’re profitable. You’re just ignoring reality.
When you have to invent a new metric to make your business look viable, your business isn’t viable.
The AI sector is adopting the exact same financial engineering theater we’ve seen fail before. It’s frustrating to watch investors and observers allocate capital and trust based on these false metrics. We are being gaslit by corporate accounting masquerading as genuine business milestones.
Next time an AI company claims it’s crossed the profitability threshold, ask them one question: GAAP or non-GAAP? Because “profitable without COGS” doesn’t actually mean anything at all. Stop believing the adjusted numbers. If the core operations don’t make money when you include the cost of building the product, the emperor has no clothes.
FAQ
Q: Doesn't adjusted operating income give a clearer picture of operational efficiency?
A: No. Excluding the cost of training models and deploying hardware in an AI company is like a restaurant excluding the cost of food. It doesn't show efficiency; it hides the actual cost of doing business.
Q: Should investors completely ignore these profitability claims?
A: Yes. Investors must demand GAAP metrics. If a company can only claim profitability by inventing a metric that excludes its primary expenses, the core operations are deeply unprofitable.
Q: Is this just standard Silicon Valley accounting?
A: It’s standard Silicon Valley theater, but the scale of the deception is historic. We're watching the gig economy IPO accounting tricks on steroids, applied to an industry burning billions on compute.