Six months ago, a friend of ours, a software engineer at one of the Mag7 companies, told us his performance was being measured by how many AI tokens he burned. So he was building things he had no intention of using, just to keep his boss off his back.
We assumed it was one bad manager. It was the entire industry.
Meta made AI-driven work results a core performance requirement and paid bonuses on it. Employees responded by competing on an internal leaderboard that tracked token consumption, which jumped from 60 trillion to 74 trillion tokens in a single month before Meta quietly pulled the ranking. Amazon shut down its own leaderboard after employees admitted to running pointless tasks just to climb it. Jensen Huang said he would be deeply alarmed if his engineers were not burning tokens worth half their salary. OpenAI handed out physical trophies at DevDay to its biggest token consumers: silver for 10 billion, black for 100 billion, blue for 1 trillion. The winners issued press releases. Nobody was bragging about what they built. They were bragging about what they spent.
The industry had an output measurement problem and solved it by measuring the input.
Here is what the output actually looked like at the same time. METR ran a randomized controlled trial on experienced engineers using AI coding tools. The engineers believed they were 20 percent faster. They were 19 percent slower. And a Stanford study of AI agents found that past a point, more tokens buy nothing. Identical tasks varied 30x in token consumption between runs, and accuracy did not improve with spend.
So token volume went vertical, Google alone went from roughly 10 trillion tokens a month to 3.2 quadrillion in two years, while measured output stayed flat to negative. The growth everyone celebrated was a blend of three things: gamed KPIs, agent architectures that burn 100x the tokens to do the same work, and prices subsidized below cost.
Then the bill arrived. Uber burned through its entire 2026 AI budget in four months. ServiceNow did the same and now monitors usage per employee per day. Meta went from paying bonuses for token consumption to building a dashboard that flags token consumption, in under a year. Demand that disappears the moment someone has to pay for it was never demand. It was a subsidy being consumed by people who were paid to consume it.
The honest version of the problem is that the headline numbers were uninterpretable, a mix of real usage, gamed usage, and inflated usage, and the industry chose to read the whole number as output.
If you sell tokens, or the infrastructure that generates them, this confusion is your business model. You want your customers competing on inputs. But for everyone else, it is like giving awards to portfolio companies for spending the most money instead of earning the most. When enough people get behind a ridiculous idea, it starts to sound reasonable. It never actually becomes reasonable.
