Google now processes more than 3 quadrillion AI tokens a month. Two years ago the number was around 10 trillion. That is a 300-fold jump, nearly 30,000%, in 24 months. We have gotten extraordinarily good at spending tokens. We have barely started asking what each one buys.

We optimized one side of the equation and ignored the other

Every AI conversation for the last three years has been about consumption. How many seats, how many calls, which model, how many tokens. That is one side of an equation. The other side, the return, has gone almost entirely unmeasured.

MIT studied that gap and found that 95% of enterprise GenAI pilots produced no measurable P&L impact, despite $30 to $40 billion spent. Read that as an accounting failure, not a technology one. The spend gets tracked (very primitively). The yield does not. We built a habit of buying intelligence and never wrote down what it earned. That worked while tokens were a rounding error. It stops working the moment they become the largest variable cost on the page.

Tokens behave like money. So treat them like money.

A token is the smallest unit of AI work, the way a cent is the smallest unit of a dollar. That comparison used to be a metaphor. It is turning literal.

The oldest test of whether something is money, “The Barter System”; Can you exhange it for something with real utility (utility is subjective btw).

• Entertainment. An evening of aimless web surfing is now an evening of conversation with a model. The attention that went to feeds goes to prompts.

• Productivity. Anyone can spin up a tool, a draft, an analysis, on demand. Work that needed a hire or an agency now costs a few thousand tokens.

• Interaction. Advice that came from a consultant, a tutor, or a colleague now comes from a chat window. Professional and casual both.

• Search. The question that went to ten blue links now goes to one answer. Even Google’s own surface is shifting from serving links to spending tokens.

• Expertise. The second opinion, the legal read, the research hours. Knowledge that used to carry an hourly rate now comes by the prompt.

• Creativity. Design, copy, music, and video that needed a studio or a freelancer now render from a sentence.
  and other countless things….

When one thing can be exchanged for that many others, it is behaving as a currency. And the people building the supply already talk this way. Jensen Huang calls every company a token factory and argues the metric that matters most is tokens per watt. He went further: he said he would be deeply alarmed if a $500,000 engineer didn’t consume $250,000 of tokens a year. That is not a usage stat. That is paying people, partly, in a new denomination. Sam Altman describes selling intelligence “on a meter, like electricity”, with the cost per token falling roughly 10x a year. And Satya Nadella reminded everyone of Jevons paradox: when the unit gets cheaper, we don’t use less of it. We use vastly more.

Cheaper units plus exploding volume is exactly how a currency takes over an economy. Which raises the uncomfortable question. If tokens are money, where are your books?

The new rule: return per token, not tokens per month

Most enterprises run a token economy with no accounting. They can tell you, roughly, how much they spent. Almost none can tell you what a token earned.

That is the shift. Call it AI(eco)nomics: the unit of measurement moves from tokens consumed to value produced per token. Tokens per ticket resolved. Tokens per deal closed. Tokens per engineer-hour saved. A currency without a ledger is just leakage, and right now most AI budgets are leaking in the dark.

This isn’t a finance technicality. It’s the difference between two companies spending the identical amount on AI, where one compounds the return and the other funds a habit. When the unit of value is invisible, every decision downstream is a guess: which tools to renew, which teams to fund, which use cases actually moved a number. The companies that win the next phase won’t be the ones who spend the most. They’ll be the ones who can price what each token returned, and redirect the spend toward whatever earns.

What keeping the books actually looks like

The enterprises pulling ahead aren’t burning the most tokens. They’re the ones with a ledger under the spend. Tie tokens to outcomes and a vague AI line item turns into real numbers: cost per support ticket, cost per deal closed, cost per engineer-hour saved. Same tokens, but now you can see which ones earned their keep and which ones just ran. That is the gap that decides who compounds and who simply spends, and it is why measuring AI return has to stop being a quarterly scramble and become the operating metric.

Every currency eventually gets its accounting

Gold got assayers. Dollars got ledgers. Tokens are next, because their price keeps falling while their volume keeps climbing, which is precisely when an asset stops being a curiosity and starts being an economy. The companies that build the books first will write the rules of that economy. The rest will keep spending a currency they can’t count.

The question for your next board meeting isn’t how many tokens you burned last quarter. It’s how much each one earned.


FAQ

What does "AI tokens are the new currency" mean? It means AI tokens have started to function like money: they are metered, spent, and increasingly substituted for things people used to pay for in dollars or time, from entertainment to professional work. Some companies have even begun allocating token budgets to employees on top of salary.

What is the AI token economy? The AI token economy is the emerging system in which tokens, the smallest units of AI work, become the primary unit of both cost and value in an enterprise. Spend is measured in tokens, and return is increasingly measured per token rather than per project.

How should companies measure AI token ROI? By shifting the unit from tokens consumed to value produced per token: tokens per support ticket resolved, per deal closed, or per engineer-hour saved. The goal is a ledger that ties every token of spend to a business outcome, not a monthly usage total.

Why are AI tokens compared to money? Because they pass the oldest test of currency: substitution. Tokens now buy entertainment, productivity, and advice that previously cost dollars or hours. Industry leaders reinforce the framing, describing "token factories," intelligence sold "on a meter," and engineer compensation paid partly in token budgets.

Are AI tokens actually replacing dollars? Not as legal tender, but as a unit of economic exchange inside organizations. As the cost per token falls roughly 10x a year and usage compounds, tokens are becoming the largest variable cost in many enterprises and the truest measure of AI value.