AI labs are running the burn-now-dominate-later playbook. The cost curve may not cooperate.

Last week I posted about the historic revenue growth at OpenAI and Anthropic. Numbers that have no precedent in the history of enterprise software. That part of the story is real and remarkable.

The WSJ just published confidential financials from both companies ahead of their IPOs. The revenue numbers are extraordinary. So are the losses.

But the most revealing detail is not the numbers. It is how both companies present them.

Both companies present investors with two versions of their own profitability. One that includes model training costs. And one that excludes them.

Think about that for a moment.

Excluding training costs from an AI company’s financials is like an oil company excluding drilling costs. Or a chip manufacturer excluding fab costs. Or an airline pretending airplanes are optional. The framing quietly says: “Our core business is profitable if you ignore the thing required to remain competitive.”

Investors appear to be underwriting these businesses using a familiar Silicon Valley playbook: burn aggressively now, dominate later. But unlike Amazon, the underlying cost curve here may not improve with scale. Frontier model training costs are escalating with every generation. The WSJ article says it plainly: “Each jump in intelligence is harder to come by, and costs more than the last one.” Inference may commoditize over time. But the training arms race shows no signs of slowing.

There is a subtler risk for operators building on top of these platforms. Current AI pricing may reflect growth-stage economics rather than steady-state economics. Cheap tokens. Effectively unlimited experimentation. Low-cost API calls. Developers and enterprises are building business models around economics that may not be permanent.

If pricing power eventually reasserts itself, the AI-native products built on today’s token prices may not survive tomorrow’s.

And through all of this, the hyperscalers may be the only participants guaranteed to win regardless of which model lab ultimately dominates.