Reselling tokens isn't a business
OpenAI is reportedly burning $1.69 for every dollar earned, and Anthropic is already raising prices. If your AI business is mostly API markup, the price floor is coming for you.
There’s a thoughtful piece doing the rounds at the moment about the AI bubble, and the punchline isn’t the one you usually hear. The bubble isn’t AI itself, which is going nowhere. The bubble is the price of AI, propped up by investor subsidies that the frontier labs are running out of patience for. OpenAI is reportedly burning around $1.69 of cash for every dollar it earns. Anthropic has already begun raising prices and tightening free-tier limits. Whatever the precise figure, the labs cannot indefinitely subsidise inference for tens of millions of customers, and the unit economics need to find a floor.
What I keep coming back to is who that hurts. It isn’t Anthropic and OpenAI directly, both of whom have ample paths through, and it isn’t the businesses for whom an LLM is one part of a larger product, where the model is doing one job alongside many others. The squeeze lands hardest on the middle: startups whose entire offering is a thin shell wrapped around a frontier API, charging customers slightly more than the API costs to call.
That model has only ever worked because the input price has been artificially low. If frontier API prices double, which seems modest given the burn rates being quoted, a token-markup business has to either double its own prices or absorb the difference. Customers, who have their own thoughts about what an AI feature is worth, are unlikely to absorb a doubling cheerfully. So the margin compresses, and there’s nowhere obvious for it to come back from, because the only thing the business was ever really selling was access.
The businesses that look comfortable in this scenario are the ones whose model isn’t really the product. They’re selling a workflow, a body of data, a set of integrations, a piece of judgement that happens to use an LLM as one component among many. They can swap providers, downsize models, or move to open weights when the maths demands it, because the value isn’t in the call to OpenAI, it’s in everything wrapped around it.
So if you’re a founder whose pitch deck describes a business model that is, when you strip away the marketing, mostly markup on API tokens, the next twelve months are going to ask hard questions of that pitch. Reselling tokens isn’t a business, it’s an arbitrage, and arbitrages don’t survive price shocks, which is worth thinking about now rather than later.