Alex Karp — CEO of Palantir, the data-analytics company best known for its defense and intelligence work — went on CNBC and tore into the AI industry. The labs, he said, sell you "tokens that create no value."¹
Yes, but that is not the real problem. It's that per-token pricing is a category error — and a telling one.
Price per unit is how you sell a commodity. Raw material in, add a margin, out comes a price per finished unit. It's the right model for toilet paper, for aluminium cans, for anything whose worth is in the stuff itself. Bulk goods, priced by the pound.
Intelligence is not bulk goods.
Take my own industry. No one prices a drug at raw-material cost plus markup — the idea is absurd. A pharmaceutical's price comes from the R&D poured into it, from how well you understood the patient's need, how precisely you fit the product to its markets, and whether that value is recognised and rewarded. Cost-plus never enters the room. The worth was never in the compound. It was in everything built around it.
AI is the same. The value was never in the output tokens. It's in how well you turn input into outcome: the system of prompts, context, agents, and judgment you wrap around a model. Two people push identical tokens through the same model; one files a regulatory dossier, the other produces slop. The tokens cost the same but the value is not even close. A token has no value on its own.
So counting tokens to reach a price measures the one thing that doesn't matter.
The pricing model of the labs gives them away. Per-token billing is an honest confession that they sell a commodity — raw compute, metered by the pound. The value was always downstream, with the person doing the wrapping. Karp's fix is to move that meter inside his own castle and call it sovereignty.² Same category error, longer lock-in.
I've written that when tools converge, people make the difference. Pricing proves it from the other direction. When the model itself is billed like a commodity, the only thing left worth paying for is what a human builds on top of it.
Stop counting tokens. Start measuring outcomes.