Modeled at three hundred and eighty trillion tokens a day. The AI business has stopped being a demo and started being an economy — and the unit is an index, not a barrel.
AiSignal staff · 7 September 2026 · 11 min read
Sometime this summer the meter broke. Not a single meter — there is no single meter — but the idea that anyone, including the labs, could tell you how much intelligence the world was producing before the number had already moved. Tokens Per Day, a stubborn little index that refuses to pretend it is a utility bill, put the disclosed floor at 300 trillion tokens and the modeled total near 360 trillion in mid-July. We carry 380 trillion as a September run-rate. Treat every digit after the first two as weather.
A token is a shard of a word, a piece of an image patch, a slice of a tool call. It is also a unit of cost, of power, of time. Fireworks AI, an inference company most civilians have never heard of, was already pushing tens of trillions of tokens a day for Cursor, Uber, Samsung, Notion. One vendor. The rest of the iceberg is ChatGPT — OpenAI publicly reported more than 2.5 billion messages per day; Epoch later modeled ChatGPT closer to ~4 billion, which is an estimate, not the same kind of disclosed number — plus Gemini in search, Claude in the enterprise, Grok on X, and a thickening fog of open-weight endpoints that do not file 8-Ks.
The unit the century picked
Oil had barrels. The web had pageviews. Advertising had impressions. This industry priced tokens, and it is slightly embarrassing how well the unit works as money — and how poorly it works as physics. Tokenizers differ. Input, output, reasoning, and multimodal tokens are not the same object. Even current inference trackers warn that aggregated token counts are a directional index, not a barrel of oil. We still sit the meter. We do not pretend the units commute.
The most persuasive economic document of 2026 is not a 10-K. It is a token counter.
Jensen Huang has been saying out loud what the factories imply: a gigawatt of IT capacity that used to emit a couple of million tokens a second is on a path to hundreds of millions. Vera Rubin is the named vehicle. The grid is the constraint. If the chips land and the substations hold, 2027 is not a continuation of 2026. It is a different animal. Median scenarios that looked like 5 trillion tokens a day at the start of the year are already off by two orders of magnitude on the disclosed floor alone.
Who is buying all this thought
Not, mostly, the movie-poster version of AGI. ChatGPT crossed a billion monthly active users in June and a shocking share of those sessions are homework, recipes, and 'what is this rash.' OpenAI's own research put roughly 70% of usage outside work. That is not a bug in the business. That is the business. Consumer demand is what lets a lab raise at a trillion-dollar track while the enterprise still argues about procurement.
The serious money, the money that shows up as 29.6 gigawatts of AI data-center capacity, is a bet that those casual sessions become agents — software that does not stop at the reply. Today the physics of that bet are unkind. Convert the global token factory into full-time knowledge-worker equivalents and you get something on the order of 1,300 fully autonomous agents. Not 1.3 million. Thousand. The jump to a million is a 2028 story, and only if the chips, the power, and the product all show up.
Until then the flood is real and the job is not to mythologize it. It is to meter it.