Tokens vs. outcomes

Show me the incentive and i'll show you the outcome

Downstream of compute, there are basically two types of businesses emerging in AI:

1. Those that sell tokens or usage (you make more money the more people use you/workloads run)

2. Those that sell outcomes (you make more money the more correct you are)

We can think of these colloquially as the frontier labs versus the application layer companies, but there are application layer companies that sell tokens, and it appears that the model labs are at least exploring selling outcomes. These approaches have fundamentally opposed incentives.

Selling tokens

The frontier labs, neoclouds, inference clouds and others are in the business of selling tokens. A lot of the products they ship treat token efficiency as a byproduct vs as a constraint, which makes sense because their teams have access to way more compute than you, way better models, and minimal constraints on token spend/usage. In certain cases they will try to make things more token efficient, but at they’re always selling you tokens. Unlimited token budget basically does not exist if you dont work for these companies. Outcomes are your problem.

This is an exceptional business. If you sell intelligence by the token, you don’t need to be sensitive to whether or not the actual task was completed with any specific level of precision. You just need to ensure that the user feels like it was, so they hit the slot machine again next time. As long as this continues, you’re fine.[1]

Selling outcomes

Selling outcomes is super different. Depending on the industry, you take on varying levels of liability for the outcome being incorrect. This might be implicit liability in terms of you not getting paid unless the task is actually completed, or actual statutory liability (particularly true if you’re fully vertically integrated or building AI enabled services). In most of these cases, you have some kind of compute or token budget constraint, and as a result, your incentives diverge - you you price based on outcomes and efficiency = margin. To put it in context, Anthropic’s current $100b fundraise is probably greater than the entire amount raised by application layer companies YTD on earth. Even if you wanted to, you literally cant compete.

The economy is made out of tasks. Do you sell completed tasks or tokens?

If you sell tokens, your incentive is to get people to use as many tokens as possible because that’s how you get paid. If you sell outcomes, your incentive is to be correct as cheaply as possible because that’s how you get paid. As a result for companies that sell tokens, shifting to sell outcomes will be hard structurally (because its not their DNA to have precision here) and culturally (because employees likely havent ever had to be sensitive to token budget in a specific domain). ALSO, selling tokens is an amazing business. Why would you want to compromise it?

One way to think about it is who is responsible when the model hallucinates

If you sell tokens, the customer is responsible.

If you sell outcomes, you are.

AGI

The one caveat to this is if you are able to get to AGI or ASI, because theoretically, especially if AGI means that you have perfect decision-making in every scenario given the same context versus anyone else (or at worst, decision quality that is sufficiently better vs. the rest of the world that your feedback loop means you pull away/hit escape velocity faster than they can catch you). What that means is you now know you can always be correct, and token budget just becomes an optimization you can choose. The frontier labs are dramatically more likely to achieve AGI or ASI, if it’s achievable, than the application layer companies are, and they can also just choose not to make it available, cementing an escalating advantage vs. everyone else.

Thanks to Dimitri Dadiomov and Daniele Perito for reading this in draft.

[1] To be clear, very often the outcome is or can be achieved, but its on the user to get it there, and the token provider isn’t on the hook if it’s not.