They Traded People for Tokens. The Returns Never Showed Up.

Jensen Huang has a napkin math test for whether an engineer earns their seat. At the close of Nvidia’s GTC 2026, he floated a number: a half million dollar engineer should be burning through at least a quarter million dollars in AI tokens a year, and Nvidia itself is steering toward a two billion dollar annual token bill for its engineering force. It is a memorable line from the man who sells the compute. It is also the cleanest description you will find of a trade that thousands of companies have quietly made without saying it out loud. Money that used to pay people is now paying for tokens.

The question almost nobody asked while the trade was happening is the only one that matters now. Is it working?

The trade everyone made quietly

Here is where the record turns awkward. Gartner surveyed 350 executives at companies pulling more than a billion dollars in revenue, every one of them running AI agents or automation. Roughly eight in ten had already cut headcount. And the correlation between those cuts and any improvement in returns? There was none. The reductions freed up budget. They did not manufacture results. The organizations that actually moved their return on investment were the ones that used AI to make their people faster and sharper, not the ones that used it to show their people the door.

Read that twice, because it inverts the whole pitch. The promise behind the trade was substitution. Swap a salary for compute, let the machine absorb the work, and pocket the difference. The data coming back from the companies that moved first says the difference mostly never showed up.

The Substitution Trap

Call it the Substitution Trap: the belief that if AI can produce the output, you no longer need the human who used to produce it. It sounds airtight on a spreadsheet. It falls apart in the building.

Uber handed AI coding tools to five thousand engineers in December and torched its entire 2026 AI budget by April. Four months. Leadership admitted that even with a huge share of committed code now machine generated, the line connecting all that output to anything a customer actually feels was still missing. Engineers got capped at fifteen hundred dollars a month. Klarna is the cautionary tale everyone points to, having walked back a headcount cut of seven hundred roles after the customer experience and the brand took the hit. Across the industry, the informal culture of “tokenmaxxing,” where burning the most tokens marked you as the most innovative, cooled fast once the invoices landed.

The pattern is the same every time. Output went up. Outcomes did not follow. And that gap, the distance between something being produced and something being worth producing, is exactly where the trade breaks.

Why the tokens never became returns

Here is the missing variable, and it is not a better model. It is a fluent human standing between the model and the result.

An AI system generates a draft, a chunk of code, a customer reply, a forecast. That output is raw material, not a finished outcome. Somebody still has to know whether it is right, aim it at the actual problem, catch the confident mistake, and carry it the last mile to something a customer or a student or a patient benefits from. When companies removed the people and kept the tokens, they did not just cut cost. They cut the judgment layer that turns generated output into real return. The machine kept producing. Nobody fluent enough was left to make the production matter.

Think about what a token actually buys. It buys a probability, a very good guess at the next word or the next line of code. What it does not buy is accountability, context about your specific customer, or the instinct to notice that a plausible answer is quietly wrong. Those live in a person. Strip the person out and you are left with a firehose of plausible material and no one holding the hose. The bill still arrives every month. The value does not, because value was never a property of the output alone. It lived in the loop between a capable human and the tool, and that loop is the first thing a headcount cut severs.

This is why the winners in the Gartner data amplified their people instead of replacing them. It tracks with what we saw when Samsung handed AI to every employee and discovered access alone changes nothing, and with the two track job market that keeps splitting workers by fluency rather than by title. The tool is now everywhere. The fluency to make it pay is not, and that scarcity is the entire story.

What this actually means for you

You probably do not sign off on your company’s token budget. That is fine, because the lesson is not really about budgets. It is about which side of the trade you want to be standing on.

The Substitution Trap creates a strange kind of opening. Every organization that cut too deep now has generated output piling up with too few fluent people to direct it. The person who can take an AI system, aim it well, verify what it produces, and finish the job becomes the one who makes the whole expensive bet finally work. That person is not defined by a job title or a degree. They are defined by fluency, the practiced skill of working with these tools rather than around them or under them. That is the entire premise of staying in the driver’s seat instead of getting automated out of it.

Building that fluency is not a weekend download or a certificate you frame. It is a habit. That is what SeedStacking is built for: small, repeatable reps that stack real capability over time, one Seed at a time. You do not need to control the trade. You need to become the person it cannot work without. The fastest way to start is with other people doing the same thing, which is exactly what the free community is for.