Burn Tokens, Not Headcount

In a recent YC batch talk, general partner Tom Blomfield made a claim that should reframe how every founder and enterprise leader thinks about AI spend: YC companies are now reaching demo day with about 5x more revenue per employee than they did 18 months ago. The constraint isn't headcount any more. It's token consumption.

"Burn tokens, not headcount."

The Roman-legion company is built around humans as the conduit for information flowing up and down. Nested hierarchies, named spans of control, middle managers translating intent into action. Blomfield's argument is that the legion was an engineering choice forced by the limits of human communication, and that AI removes the constraint that made it the right shape. Once a company's domain knowledge is legible to AI, the hierarchy stops being load-bearing.

The mental model he wants founders to drop is the copilot. A copilot is AI bolted onto an existing workflow to make a human 20 or 30% more productive. That's the old shape with a bigger engine. The new shape is recursive self-improving loops: a sensor layer pulling in support tickets and product telemetry, a policy layer deciding what's safe to act on, a tool layer of deterministic APIs, a quality gate, and a learning mechanism that closes the loop. Run those overnight and the company improves while you sleep.

YC's own example is the user manual. Five-to-ten-year-old advice, manually maintained, treated as gospel by founders who've barely opened it. Over a single weekend they regenerated it from 2,000 hours of recorded office hours into a 150-page document that updates itself every month. The artefact stopped being a document and became a living interface to the combined judgement of 16 partners. Same principle for product analytics. Same principle for customer support. Each function becomes a loop that compounds rather than a process that depreciates.

The implication for enterprise leaders is awkward. Token usage as a directional signal of who's experimenting is gameable the moment it becomes a performance metric, but Blomfield is right that it's the cleanest proxy we have right now for who in the organisation is figuring out what the new intelligence can do. The leaders worth spending time with are the ones token-maxing. The functions worth restructuring are the ones still wired as a Roman legion.

The harder claim, and the one worth sitting with, is on middle management. Blomfield says it's done. The coordination problem that used to need humans translating intent down and information back up is now an AI problem. What's left at the human edge are the high-stakes moments where models can't reach: the conference, the sales conversation, the founder breaking up with their co-founder. Everything in the middle is software that should be ephemeral, generated against a preserved corpus of business context, and regenerated when the next model rolls.