Deliberately overspending on inference buys lead time, not efficiency

Tan argues the frontier labs behind the consumer products are “pretty constrained on… amount of compute they want to give you,” so anyone who wants to run at full strength has to reach for a harness they control and “tune it all the way up” — loading “a million tokens or 800,000 tokens” into any given request. He puts the cost at “50 or 100,000 dollars a year to like use the agents at full strength,” calls it “a crazy amount,” and says that for a CEO or a founder “it actually makes a lot of sense” because “you get to live in 2028 today.” The operative phrase is permission: “you have to give yourself permission to token max in that way.”

This is the mirror image of Inference-time compute makes cost-per-outcome a choice — and that's the application layer's counterattack on the labs, and the two are worth holding in tension rather than reconciling. That insight is about product margin — spend tokens only where the outcome justifies them, because the lab wants to expand usage while the application wants discipline. Tan’s move is deliberately indisciplined at the unit level: you are not buying a better answer to this task, you are buying an earlier position on the capability curve, and the return shows up as lead time rather than as margin. The distinction that decides which applies is who bears the cost — a per-customer inference path demands the discipline, a founder’s own leverage does not.

The implication for The price of intelligence is the new organizing axis — labs, applications, and countries are all fighting to set it is that the same falling price curve supports two opposite strategies simultaneously: ride it down to widen margins, or spend ahead of it to compress your own timeline. It also sharpens AI compresses the distance between idea and execution but not between good and bad judgment — if compute is no longer the binding constraint for someone willing to pay, then the constraint really is what you choose to point it at.