OpenAI loses $1.69 for every dollar of revenue. Anthropic pays over $1 billion per month for compute.

These are not small companies making small mistakes. These are the leaders in AI, and the math does not work.

The Expectation

We expected AI companies to behave like software companies.

Software has beautiful economics:

  • Build once, sell infinitely
  • Marginal cost approaches zero
  • Gross margins climb to 75-85%
  • Scale is your friend

This is the SaaS playbook we all internalized.

The Reality

AI companies are infrastructure businesses wearing software company clothing.

Every token generated costs real compute. Marginal cost stays stubbornly positive.

And token prices collapsed 75% in one year.

So you sell more, but you earn less per unit, and your compute bill grows with usage.

Volume is not the cure. Under these dynamics, volume is part of the problem.

The Numbers

AI-native companies sit at 45% gross margins. Mature SaaS sits at 75-85%.

That 30-point gap is structural, not transitional.

OpenAI’s leaked projections show losses widening through 2028. Not shrinking. Widening. $74 billion in losses on $100 billion in revenue.

The Trap

These companies do not own their infrastructure. They rent it from the same companies that fund them.

Microsoft funds OpenAI. OpenAI pays Microsoft for Azure compute.

Amazon funds Anthropic. Anthropic pays Amazon for AWS compute.

And both companies compete with models from their infrastructure providers.

You do not typically hand your gross margin to your competitor. Here it is the default arrangement.

The Bets

Two different theories on how this gets fixed:

Scale - Build the largest platform, capture all demand, and assume pricing power emerges at sufficient scale.

Wedge - Find work valuable enough that buyers tolerate real prices. Coding tools, agents, automation measured against labor cost rather than token cost.

Claude Code at $2.5 billion annualized proves the wedge exists. Whether it is large enough to matter is the open question.

What Has To Happen

Three things must move:

Gross margins must climb from 45% to 60%+ and stay there while volume grows.

Model efficiency gains must outrun price deflation instead of getting passed straight to buyers.

The product mix must shift toward work that resists deflation.

None of this is proven. None of this is guaranteed.

The Open Question

The question is not whether AI companies are growing. They obviously are.

The question is whether revenue growth and margin improvement are the same trend or opposing ones.

The SaaS era taught us that scale fixes economics. The leaked numbers describe a business where scale makes the loss bigger.

The math does not work. Yet.