Simon Willison argues that Anthropic and OpenAI have found product-market fit, but at one precise point. The enterprise value is real and concentrated: coding agents burn far more tokens than consumer products, and they work where there are well-paid professionals whose companies accept API prices without flinching. Translated for a board: when management brings a PoC, the right question isn't "does it work". It's "at what token price does it work". Many PoCs shine because they run on launch rates, subsidized to grab market share. Change the price list, or scale up volume in production, and the economics break. A director who doesn't ask for cost per task reasoning at full prices, not today's promotional ones, is evaluating a demo, not a deal. There's one number to get: cost per use case at steady state, with real margins.
Why this matters for anyone building enterprise AI: a PoC that only holds up at subsidized prices isn't a product, it's a subsidy on a clock.