The unit economics of AI are breaking in the wrong direction, and the trillion-dollar infrastructure buildout might be running on borrowed math.

The Summary

The Signal

The AI boom has a pricing problem. According to Steve Hou, head of research at Silicon Data, the cost per token generated is moving in the opposite direction of token prices. Compute isn't getting cheaper fast enough to offset the commodity pricing pressure on AI outputs. That's the kind of margin squeeze that makes venture capitalists reach for the Tums.

This matters because the entire valuation structure of companies like OpenAI and Anthropic depends on a world where AI services scale profitably. If every marginal token you sell costs more to produce than the last one, you're not building a software company. You're building an airline.

"Token prices are falling while the cost of compute needed to produce them is rising."

The infrastructure buildout tells the story. Hyperscalers are spending hundreds of billions on GPU clusters, power contracts, and cooling systems. The bet is that demand will absorb all that capacity at prices that cover the build. But what if the demand curve and the cost curve are racing toward each other? What if the gap closes before the promised AI productivity revolution materializes?

The question extends beyond today's models to the agent economy everyone's building toward. If millions of AI agents are running on behalf of users, and each agent is burning tokens at scale, who pays? At what price? And can that price support the infrastructure underneath?

Here's the uncomfortable math:

  • Compute costs are sticky, governed by chip supply chains and energy costs
  • Token prices face competition, open models, and customer resistance to premium pricing
  • Infrastructure investments are already committed, based on optimistic demand forecasts

The gap between those three realities is where companies die. Or pivot. Or discover they're not software businesses after all, but compute arbitrage plays with a very narrow spread.

The Implication

Watch for two signals in the next twelve months. First, pricing strategies from the major AI labs. If OpenAI or Anthropic start bundling, tiering, or creating enterprise lock-in mechanisms that don't depend on per-token pricing, they're feeling the squeeze. Second, watch the agent platform companies. The ones building Web4 tooling need a stable cost structure to promise developers. If the economics don't work at scale, the agent economy doesn't get built.

For builders, this is a moment to get serious about efficiency. The free lunch on inference costs is ending. If your agent or application assumes cheap, abundant compute forever, stress test that assumption. The boom survives, but not everyone riding it does.

Sources

Bloomberg Tech