Inference-time compute makes cost-per-outcome a choice — and that's the application layer's counterattack on the labs

Gupta isolates what’s genuinely new about inference-time compute: “no prior software had a dial where spending ten times more compute bought a better answer. It makes cost per outcome a choice. A ten-second query and a ten-minute query on the same model are different products at different prices, so margin depends on the system’s judgment, not just model pricing.” This is “the application-layer counterattack against the labs”: as frontier models absorb product logic, every application has to prove “it can allocate the customer’s tokens better than the provider can.” The incentives diverge cleanly — “The lab is incentivized to expand usage; the application is incentivized to spend only where the outcome is worth it.”

This makes token allocation a moat, which is why Routing across the whole model market — and absorbing every migration — is a defense the labs can't copy and AI is the computer — orchestration across 19 models is the product, not any single model — spending the right amount of the right model per sub-task is the judgment being sold. It is the application-layer expression of The system of work is the moat, not the model — the model is fungible underneath: the labs own the model, but the system that decides how much intelligence each step deserves is the defensible surface. Operationally it extends Production agents route routine cases through decision trees, reserving humans for complexity — routing isn’t just latency/cost hygiene, it’s the core economic decision. The length counterpart to this depth dial is Task horizon breaks seat-based pricing — usage scales with workflow depth × length, not headcount.