Measuring AI Adoption: Why Token Counts Went the Way of Lines of Code
In May, Nvidia’s CEO Jensen Huang said this out loud: “if your $500K engineer isn’t burning $250K in tokens, something is wrong.”
Around the same time, OpenAI’s Sam Altman said he was excited to see what tokenmaxxing startups would do. Meta ran an internal competition rewarding staff for token usage. For a few months in 2026, the amount of AI you consumed was treated as evidence of how good you were at your job.
By August, the Associated Press was reporting the backlash. Mozilla’s chief technology officer Raffi Krikorian put it at its bluntest: “tokenmaxxing is a dumb thing.”
That’s a fast round trip for a management idea, and it’s worth understanding why, because the specific metric is going away, and the mistake underneath it absolutely is not.
What tokenmaxxing was
A The unit AI companies bill by, and roughly three-quarters of a word. Every question you send and every answer you get back is counted in tokens, so a long document costs more to process than a short one. is the thing you’re charged for. “Tokenmaxxing” meant maximizing how many of them you burned, on the theory that heavy usage signals someone extracting real value from the tools.
The stereotype, as the AP described it, was an employee up late orchestrating an army of AI agents working through the night on their behalf.
You can see why it appealed. Adoption is notoriously hard to measure, and here was a number sitting right there in the billing dashboard: precise, current, and comparable between people. Very tempting.
It’s lines of code, all over again
Krikorian’s analogy is the useful part, and it’s why this post isn’t really about tokens.
Software companies once measured programmer productivity by counting lines of code written. It seemed reasonable: more code, more work done. It fell apart because the metric was trivially gameable and pointed the wrong way: the best engineers often delete code, and the clearest solution is usually the shortest one. Rewarding volume rewarded bloat.
“I think tokenmaxxing is moving through the exact same pattern,” Krikorian told the AP. “I think this is going to be an interesting blip that we’re all going to look back to laugh at in a year.”
The pattern is always the same. Something real and hard to measure (value, skill, adoption) gets replaced by something easy to measure that correlates with it. Then people optimize the A stand-in measurement. Something you can count that’s related to the thing you actually care about but isn’t the same as it. Proxies are useful right up until someone is rewarded for moving them, at which point they stop tracking the thing they stood in for. instead of the thing, and the correlation dies.
Nicholas Arcolano of Jellyfish makes the point precisely: “Extreme token use often isn’t a sign of good engineering.” It can just as easily indicate poorly specified tasks, repeated rework, or someone fighting the tool. High usage and high competence look identical on that dashboard.
Then the bills arrived
The collapse wasn’t philosophical. It was financial.
Bain’s Jue Wang told the AP that at the large companies her firm advises, “the token cost for them has been doubling, almost every other month.” Her worked example is worth sitting with: $200 per developer per month, across 20,000 developers. “a number that is not a line item that any general manager has planned for.”
Vincent Gusdorf, who runs AI analytics at Moody’s Ratings and wrote a report recommending more discipline, described the turn plainly: “As bills started to pile in, people realized that those new tools are quite expensive and you need to use them wisely.” And, more pointedly: “It’s very easy to create something you don’t need with AI.”
Palantir’s Alex Karp was less diplomatic, telling CNBC that American businesses were privately “livid” about paying for tokens that created no value.
The cost that isn’t on the invoice
There’s a second bill, and it’s the one most adoption metrics can’t see at all.
Microsoft’s Satya Nadella has argued that companies are paying twice: once in tokens, and again by feeding their proprietary data to the model provider. Worth noting he was promoting Microsoft’s own approach while raising doubts about competitors’ data assurances, but the underlying point stands regardless of who makes it, and it’s the same argument as shadow AI from the other direction.
A usage metric rewards putting more through the tool. If nobody has written down what may go into it, “more” and “more exposure” are the same instruction.
What to measure instead
Here’s the awkward part: the reason a bad metric spreads is that measuring adoption properly is genuinely harder, and vaguer, than reading a number off a dashboard.
Three that survive contact with reality, in rough order of usefulness:
- What’s still in real use at 30 days. Not what was built during the enthusiasm: what’s still running a month later, when the novelty is gone. This is the number that survives scrutiny, and it’s the one the manager’s AI toolkit ends on.
- Time to first useful output for a new starter. How long before someone who joined last week produces something with AI that you’d actually use? It measures whether the knowledge is in the team or in two people’s heads.
- How many workflows have a named owner. Anything in real use that nobody owns is a future outage. Counting owners is counting durability.
None of these can be gamed by being busier, which is the whole point.
The cheaper levers, briefly
If cost is the actual problem, most of the saving isn’t in typing less. It’s in not using an expensive model for cheap work. Wang again: “Not everything needs a Claude Opus 4.6. And yet you see so many companies, so many users, default to using Opus for everything, including generating emails.”
That’s driven a shift toward Automatically sending simple requests to a cheaper, smaller model and only the hard ones to the expensive model. The user usually doesn’t notice; the bill does. , which is where most of the real money is.
Two smaller ones that anyone can do today. Output usually costs several times more than input, so asking for a shorter answer saves more than writing a shorter question. And a long chat thread re-sends the entire conversation with every turn, so starting a fresh chat for a fresh task costs less than a sixty-message thread, and it’s why you hit usage limits faster than you expect.
Don’t overshoot in the other direction
A spending crackdown is its own failure mode, and it’s the more expensive one.
Hassan El Mghari of Together AI told the AP the better approach is “to kind of just empower employees on how to use this stuff and let them use AI when and however much they need to.” His company sells access to cheaper models, so read it with that in mind, but the logic is sound. An hour of someone’s time costs far more than the tokens they’d spend in it, and AI already gives back a real chunk of its time savings as rework.
Rationing a tool to save money on the tool, while spending salary to do it, is the same category error as tokenmaxxing. Just pointed the other way.
The takeaway
Tokenmaxxing will be a footnote. The thing to take from it is the shape, so you recognize it next time: a number that’s easy to count gets promoted to a target, people optimize the number, and the number stops meaning what it meant.
When someone offers you a tidy figure that proves your team is adopting AI well, the question to ask is whether it can be moved by simply doing more, and whether doing more is what you actually wanted. Usage is not adoption. Spend is not adoption. What’s still running in thirty days is the closest honest answer any of us have.
Sources: the quotes from Jensen Huang, Sam Altman, Satya Nadella, Alex Karp, Vincent Gusdorf (Moody’s Ratings), Jue Wang (Bain & Company), Hassan El Mghari (Together AI) and Raffi Krikorian (Mozilla), the Meta internal competition, and the definition of a token are all from the Associated Press, “Workplaces look for cheaper AI as ‘tokenmaxxing’ fades as a corporate fad” (opens in a new tab) (2026). The Nicholas Arcolano quote and the point about output tokens costing more than input are from TechTarget (opens in a new tab). Two interests worth naming: Nadella was promoting Microsoft’s approach while criticising rivals’ data handling, and El Mghari’s employer sells cheaper open-source model access. The three replacement metrics are mine, not anyone’s research.
Related: Buying AI Tools Isn’t Adopting AI, and the Bill Comes Due Later is the same mistake one step earlier, when the spend is on licences instead of tokens. AI Productivity Gains: Everyone’s Using It, Almost No One Is Saving Time is what happens when usage goes up and nothing else does.