Inference is profitable; training is what costs

The claim that AI inference is inherently unprofitable, propped up only by VC subsidies, does not survive basic arithmetic. A dense 70B model on four A100s produces roughly 2 million tokens per hour at about $1 per million output tokens when you amortize GPU hardware over a five-year lifespan and add industrial power costs. Frontier providers charge $4.50 to $27 per million tokens, implying the claimed 70 to 80 percent gross margins are plausible if not conservative. DeepSeek proved this from the supply side: they claimed 80 percent margins while charging less than half of OpenAI or Anthropic, and the open market price for equivalent inference settled around 87 cents per million tokens.

The high margins exist because inference revenue must subsidize training costs for AI labs. OpenAI and Anthropic are not just selling predictions; they are funding the next model. A pure inference provider with no training program can be profitable at a fraction of frontier pricing. This means the agent ecosystem does not depend on the survival of any single lab: whoever acquires the model weights can continue serving inference profitably. For builders, the practical takeaway is that token costs (see: token economics dominate the cost of agent projects) will fall as more providers compete on inference margins, and open-model inference (via cheaper providers (see: model choice trades quality for cost and speed)) is already viable for most agent workloads today.