David George frames the internal economics shift bluntly: “Token spend per engineer is investment, not cost.” The practical numbers: $1,000/month token spend per engineer is table stakes, top engineers see order-of-magnitude productivity gains and manage 20-30 agents simultaneously, and the organizational unit shrinks from 10-person committees to 4-person strike teams (design + product + engineering collapsed). The hiring model inverts — find ~5 people who will deliver 100x value regardless of seniority, then scale through compute, not headcount.
This is the internal R&D economics that makes Sell the work, not the tool — the services budget dwarfs the software budget, and every model improvement helps whoever sells the outcome possible: if your own engineers are 10x more productive via token spend, you can build the autopilots that capture the 6x-larger services TAM. The “20-30 agents simultaneously” is exactly what An orchestrator agent that manages other agents solves the parallel coordination problem without human bottleneck enables operationally — a single engineer managing an agent swarm, not manually coordinating human teams. The counterintuitive implication connects to AI automation amplifies demand for expert human judgment rather than replacing it: fewer engineers total, but each one doing far higher-value work at dramatically higher effective output. And the margin structure explains why Platform economics beat labor arbitrage — margins fund flywheels that body shops cannot holds — companies that invest in compute infrastructure create flywheels, while those that hire bodies cannot keep pace. Boris Cherny provides the clearest evidence: Uncorrelated context windows are a form of test time compute — fresh perspectives multiply capability — at Anthropic, he starts 80% of sessions in plan mode across multiple terminal tabs and desktop app windows simultaneously, each with an independent context window. The plugins feature was “entirely built by a swarm over a weekend” with no human intervention.
Tom Blomfield compresses the same prescription to four words for YC founders: “burn tokens, not headcount.” His evidence is the shape of the current batch rather than a spend figure — founders reaching demo day with a million dollars of revenue, and some reaching Series A with ten million, at a fraction of the headcount those numbers used to require. He is careful about the multiplier, declining to name one: “you can argue about whether it’s a thousand X or whatever, but it’s multiple times more.” The claim underneath is that a polyglot who really knows how to use these tools is worth many times what the same seat was worth before, so the headcount of the whole company comes down — and specifically that the layer which disappears is middle management, per When the system routes information, IC and DRI are the only roles left. Compare Deliberately overspending on inference buys lead time, not efficiency, which buys the same tokens for a different reason: not to replace a hire but to occupy the capability curve early.