The Models Got Cheaper. The Bills Didn't.
Inference costs fell off a cliff. So why is your AI line item bigger than ever?
Every quarter the labs announce cheaper tokens. Every quarter my client's AI bill goes up. Both statements are true, and that is the interesting part.
Falling unit price does what falling unit prices always do: it expands use. When a call costs a tenth of what it did, you stop rationing calls. You wrap everything in a model.
Then the second-order costs show up. Vector stores. Monitoring. The engineer whose full-time job is "make the prompts stop hallucinating." None of that is in the token price.
The companies winning on cost are the ones who set a budget and treat the model like a utility, not a toy. The ones bleeding are the ones who let every team spin up their own.
Cheap intelligence is still intelligence, and intelligence still needs someone to aim it. The aiming is where the money goes.