Earlier this year, Meta ranked around 85,000 employees by how many AI tokens they used. In 30 days the dashboard logged more than 60 trillion tokens, and the top user alone averaged 281 billion. Two days after the press picked it up, the leaderboard was gone.
I come from an information systems background, so this pattern feels familiar. Every technology wave starts by counting usage, and the grown-up phase begins when we ask about economic impact. Cloud went through it, and FinOps was the answer. Now tokenomics is basically FinOps for AI.
That leaderboard stuck with me, and it became the starting point for a short paper I wrote with Boris Otto at Fraunhofer ISST: “Stop Counting Your Tokens”. The full paper is here. This post is my personal take on it.
Why the scoreboard bothers me
Tokens are the call volume of AI. They are easy to count, they are fully automatic, and they say very little about results. Picture a sales team that is measured on calls per week: everyone can hit the target, and it tells you nothing about revenue.
The evidence points the same way. A study that linked ChatGPT usage with the financials of more than 1,500 listed companies found only a weak relationship between token use and revenue per employee. In coding-agent experiments, identical tasks differed up to 30-fold in token use between runs. I find that number humbling. If the same task can cost thirty times more on another day, then tokens tell me more about the system than about the work.
Where the cost really hides
Agents don’t chat, they loop. Every step resends the context built up so far, so an agentic task can use anywhere from a few times to, in agentic coding, around a thousand times the tokens of a plain chat, mostly on the input side. Providers already cache repeated context on their side, which makes re-read input much cheaper. Even so, prices per token are falling while consumption is rising faster, and the bills go up anyway. Uber reportedly used up its annual AI budget in about four months.
What interests me more is the reaction. Capping tokens feels like the logical answer, but saving 500 input tokens while burning 5,000 on rework just moves the cost to a place no dashboard shows, usually a developer’s afternoon.
To me the real levers are less about volume and more about architecture: whether the same context gets paid for twice, whether the biggest model handles steps a small one could, and whether a step needs a model at all. In one analysis of 24 workflows, handling the deterministic steps with ordinary logic cut model costs by about 70%. That is an engineering result, not a budget policy.
What I care about
I care about outcomes: what shipped, which rework disappeared, which vulnerabilities got closed. These are the same delivery metrics engineering teams have always tracked. I would love to see AI spend connected to them. A few companies already do this, and one of them ranks people by what they get per dollar of tokens instead of by raw volume. That feels like the grown-up version of the conversation to me.
And once capable models are available to everyone, the model itself stops being the advantage. The system around it is: context management, routing, governance and measurement. That is the part I find most interesting, and I expect to keep coming back to it here.