LLenny's Podcast
← All frameworks
Finance

Token Budgets as a Trust-Proportional Resource

Kill the token-spend leaderboard; cap AI spend like any resource, sized to trust in someone's ROI judgment

Difficulty
Easy
Time to result
~months to results
Steps
3
Confidence
84%

A framework for governing engineering/AI token spend. Reject vanity leaderboards that reward incineration; instead treat tokens like GPUs, headcount, or opex — a finite resource you deploy by ROI. As burn rates approach an engineer's salary, introduce caps sized proportionally to how much you trust that person to spend ROI-positively. Use when setting AI-tool budgets for teams.

Origin

Mosseri flatly calls Meta's famous token-spend leaderboard 'a terrible idea,' explains they reined in costs by shutting down 'token incinerators,' and lays out how he'd cap spend as burn rates rise toward salary levels.

Core principles

  • 01No leaderboards for token spend — they reward building token incinerators that create little value
  • 02Treat AI spend like any deployable resource (GPUs, opex, payroll), allocated by ROI
  • 03When burn rate approaches salary, caps become healthy — sized proportionally to trust in the person's ROI judgment

How to run it

  1. 1

    Kill vanity spend metrics

    Remove token-spend leaderboards. They incentivize burning tokens for status rather than value.

    Pro tip It's not hard to build a token incinerator; look at dollars-in versus value-out and the bad ideas become obvious.

  2. 2

    Treat tokens as a deployable resource

    Allocate AI spend the way you allocate GPUs, CPUs, storage, labeling opex, and headcount — a finite resource deployed across teams by expected return.

  3. 3

    Cap proportional to trust as burn rises

    As a strong engineer's token burn rate approaches their cost of employment, introduce caps sized to the company's trust in that person's ability to spend ROI-positively.

    Pro tip Costs may rise first (more tokens used) then fall as frontier models enter a pricing war — expect a roller coaster.

    Watch out Impose caps too early, before burn is material, and you throttle useful experimentation.

In the wild

Reining in costs by killing silly projects

Instagram got costs under control simply by shutting down the 'silly things' — token incinerators that, on a dollars-in/value-out basis, were obviously bad ideas — rather than by imposing hard per-engineer limits.

Costs came down without formal token limits for engineers.

Common mistakes

Gamifying spend with a leaderboard

Ranking people by token spend rewards incineration and signals that more spend is better, which is a terrible incentive.

Is it for you?

Best for

Engineering and product leaders setting AI-tool budgets as per-person burn approaches salary

Not ideal for

Tiny teams where AI spend is still trivially small

From the transcript

it's a terrible idea. No leaderboards for token spend.

21:00

the the burn rate of a strong engineer might be the same as their salary or their cost of employment. And if in that world…

22:30

the cap should probably be com like a proportional to your sort of you know the company's sort of trust in

22:30

From the episode

Adam Mosseri: AI is a tailwind for authenticity