The “dollar per token” business

One of the interesting business models in AI right now is what I like to call the “dollar per token” business. It’s a bit like how cloud companies charge you per I/O request, except here, you’re paying for every token, a chunk of text your AI sends or receives.

The more tokens you use, the happier the seller.

This simple and intuitive model has a twist: if the same company that sells you tokens also builds the model, it suddenly has a quiet incentive to make the model talk more. Longer replies mean more tokens. More tokens mean more dollars.

And that sets up a little tension between users (who want quick, to-the-point answers) and model providers (who wouldn’t mind a bit of extra verbosity).

But where there’s tension, there’s opportunity.

We might soon see a new kind of AI business emerge — one that charges per task, not per token.

Imagine an AI help-desk service that bills you only when it actually solves your tech issue, not for every word it types along the way. That’s like how Walmart or Costco makes money by selling goods. Their incentive aligns with their customers, i.e., to keep the prices down. A task-based AI would think the same way:

“How do I give you the answer in fewer tokens?”

If that happens, the economics of AI could shift from charging for how much the model talks to charging for what the model gets done.

Doesn’t that feel like a healthier direction for everyone?