Tokenmaxxing
We're measuring the wrong thing
Help us grow our audience. If you like this piece, please share it with one friend. Interested in joining our network? Click below.
I blame Jensen.
An offhand comment on an outlandish podcast set a new standard for AI productivity. On the All-In Podcast, Jensen Huang quantified the amount of burn he expects to see his senior engineers produce annually, otherwise they’re in trouble.
“If that $500,000 engineer did not consume at least $250,000 worth of tokens, I’m going to be deeply alarmed.” The hosts laughed while the audience took notes.
I also blame the media.
Because of its thirst for attention, this stupid concept now has a dumb name that makes you feel wrong when you type it and even worse when you say it out loud.
Tokenmaxxing.
It is the concept that the only true form of AI productivity is to burn the living shit out of every token in your digital vicinity. Are you writing a deck? Why use Figma/Keynote/Slides to create it when you can vibe it up from scratch? Are you scraping data? Why put it in a doc when you can vibe-code it into an interactive page your team will ignore?
Tokenmaxxing - even if you don’t say it - is now a mainstream behavior affecting individuals as well as enterprises like Nvidia. How many memes have you seen in the last two weeks that depict the moment you hit your token limits?
Companies are adopting it in their operational manual. Meta even had an actual leaderboard. And then they realized — duh — that engineers are gaming the system and inflating tokens for clout.
What are we talking about here?
PROVOCATION OF THE WEEK
Tokenmaxxing is right (for now).
It’s a dumb name. It can be gamed. But it’s the closest measurement to change we have at the moment. And that says more about us than it does about the metric.
An NBER study from February surveyed 6,000 executives. 90% of firms reported zero productivity gains from AI. So the bar for tokenmaxxing being useful is embarrassingly low. At least it means someone is trying.
We talked about this on the pod this week. Dan brought up the gold rush. Most miners didn’t get rich. Levi Strauss did, he sold denim to the people digging. “The outcome wasn’t gold.” Tokenmaxxing is digging. And right now, digging is all we know how to measure.
So, you know… grab a shovel.
WHAT’S NEXT?
If you have thoughts, feedback, or a perspective worth sharing, reach out: chmiel@ondiscourse.com. You might see your reaction in next week’s edition.
ON_Member Events
ON_Discourse Group Chat for AI Leaders with Nicholas Thorne, Co-Founder, Audos
Thursday, April 16, 11am EST
OpenAI just closed $122 billion. Vibe coding is a $4.7 billion market. More apps are being built than ever — and more founders are building them completely alone.
Audos, an AI platform that helps people launch businesses from scratch using autonomous agents. Since the Domain AI Summit last year, his product thesis has shifted, his tech stack has been both enabled and threatened by the same model providers, and he’s landed on a funding model borrowed from the music industry that rejects VC logic entirely.
This conversation covers what’s changed in AI since we last gathered, why the “co-pilot” framing broke, what happens when you bet your company on a model provider that can pull the rug at any time, and the question nobody in AI wants to talk about: the human cost of building with machines instead of people.
ON_Discourse Speakeasy: Monthly Cocktails for AI Leaders
Wednesday, April 22, 5pm-8pm EST, NYC
Join us for cocktails and excellent conversation with ON_Discourse members, including leaders, founders, builders, and operators.
ON_Podcast
Dan, Toby, and Chmiel break down the Silicon Valley flex nobody's asking enough questions about: burning tokens as a workplace KPI. Is maxing your token usage a signal of genuine AI-first thinking — or just a new way to dress up the same old productivity theater?