Sam Altman said it plainly earlier this year: intelligence is becoming a utility, like electricity or water, and people will buy it from OpenAI on a meter. If that vision holds, the token, the small chunk of text or data an AI model reads or writes, may become the defining unit of the AI age in the same way the kilowatt-hour defined the electrical one. It is already the currency AI companies use to bill customers, and increasingly it is the metric economists, analysts, and enterprises are using to measure how fast artificial intelligence is spreading through the economy.
THE METER RUNNING IN THE BACKGROUND
Every time a model reads a prompt or writes a response, that work is measured in tokens. AI providers still tend to sell flat-rate subscriptions to everyday consumers, but businesses and developers increasingly pay by the token, and as AI usage inside companies has soared, many have discovered just how expensive that metering can get. A period of unrestrained internal AI use, dubbed “tokenmaxxing” by some tech workers, gave way this year to companies like Uber and Amazon capping employee use of coding tools and tightening the guardrails around AI spend, a corrective some have called “tokenminimizing.”
That same metering creates something new: a digital paper trail. A recent National Bureau of Economic Research working paper by Nicola Borri, Aleh Tsyvinski, and Yukun Liu analyzed 380 trillion AI tokens processed through OpenRouter, a platform that lets developers access hundreds of models through one interface, to build what they call an “AI Factor,” a broad measure of how AI consumption is changing over time. They then tested which companies’ stock prices move most closely with that factor. Companies markets see as the biggest AI beneficiaries earned what the researchers term an “AI premium,” outperforming perceived AI laggards by about 0.64 percentage points a week. Tellingly, that premium showed up well beyond tech, spanning airlines, utilities, industrial manufacturers, retailers, and banks, suggesting Wall Street expects AI’s economic effects to ripple far past Silicon Valley.
GROWTH THAT KEEPS OUTRUNNING THE FORECASTS
If tokens are becoming the meter of the AI economy, the meter is spinning far faster than anyone predicted. Dell’s Chief Operating Officer Jeffery Clarke said in October 2025 that his company had modeled inference hitting one quadrillion tokens by 2028, then revised that estimate to 57 quadrillion, and still suspected it was too low. Current global token processing is already running at roughly 135 quadrillion tokens a year, more than double even that revised figure, with years still to go before 2028 arrives.
Every major forecaster has been caught behind the curve in similar fashion. Goldman Sachs now projects monthly token processing will rise more than 70 times over between 2025 and 2030, with agentic workloads eventually accounting for the large majority of that volume. Tirias Research’s 2023 forecast for 2024 token output missed the actual figure by more than 33 times. Google disclosed it processed 3.2 quadrillion tokens in a single month in 2026, a 330-fold increase over two years earlier. Anthropic set out to grow tenfold in 2026 and instead saw usage climb 80-fold on an annualized basis in the first quarter, straining its own compute capacity in the process.
The reason keeps coming back to agents. Multi-agent systems can consume up to 15 times more tokens than a simple chat exchange, and coding agents specifically have been measured using roughly a thousand times more tokens than a comparable chat-based interaction. Reasoning models compound the effect further: research analyzing 100 trillion tokens on OpenRouter found that the share processed by reasoning models jumped from essentially zero at the start of 2025 to around 60 percent by year’s end, with prompt tokens per request quadrupling and completion tokens tripling over the same period. Users are not just sending more requests. Each request is doing dramatically more work.
A GENUINE ECONOMIC UNIT, NOT JUST AN ACCOUNTING LINE
Researchers are now formalizing what practitioners have sensed for a while: tokens function simultaneously as units of computation, memory access, energy consumption, and monetary exchange. A working paper from NYU’s Quanyan Zhu frames the emerging field of “AI tokenomics” as the study of how tokens are generated, priced, valued, allocated, and governed, arguing this moves the conversation well beyond ordinary financial operations and cost tracking.
That framing matters because token demand behaves unlike conventional software costs. It is nonlinear, shaped by task difficulty, context size, and the uncertainty inherent in a given piece of reasoning, and much of it is invisible even to the user paying the bill. Advanced reasoning models generate internal chains of thought that consume tokens without ever appearing in the visible response, and providers price output tokens at roughly four to six times the rate of input tokens because generating text requires sequential computation in a way that reading it does not. A short customer service exchange might use a few hundred tokens. A multi-step autonomous workflow, chaining together planning, retrieval, tool calls, and verification, can run past a million.
WHAT THIS LOOKS LIKE INSIDE A COMPANY
Enterprise technology teams have started responding to that volatility with a discipline that looks a lot like financial planning. Oracle describes AI tokenomics as the practice of measuring and managing how tokens are used, priced, and optimized, built around a central question: does the value generated by an AI workflow justify the token volume and infrastructure cost behind it. Companies that get this right map token consumption across an entire workflow rather than a single prompt and response, since a single customer request might silently trigger a database query, a summarization pass, and an internal model evaluation before the user ever sees an answer.
The practical steps look consistent across most guidance. Enterprises are setting token budgets by application and team, routing simple tasks to smaller and cheaper models while reserving frontier models for genuinely hard problems, trimming unnecessary context and conversation history from prompts, and separating pilot programs from production workloads so that early experimentation doesn’t get mistaken for a stable usage pattern. Some companies have gone as far as listing generous token allowances alongside health insurance and retirement matching as a recruiting perk, a sign of how central token access has become to how people actually work.
A NEW KIND OF INDUSTRIAL METRIC
Some observers are already reaching for the electricity comparison in earnest. Writing for the European Commission’s Apply AI Alliance community, strategist Daniel Zivica points to Google’s disclosure that it processed more than 3.2 quadrillion tokens in a single month, alongside the emergence of hundreds of enterprises consuming over a trillion tokens annually, as evidence that token volume has become a genuine macroeconomic indicator, one he argues should be treated the way earlier eras treated electricity output or steel production. His broader argument is that AI is not simply software to be licensed and installed but something closer to a new digital power grid, and that public investment should follow accordingly, building sovereign compute capacity rather than continuing to fund technology procurement models built for an earlier, more static era of enterprise software.
Whether or not every part of that framing holds up, the underlying instinct tracks with what the market data shows. Token consumption is growing at a pace forecasters keep underestimating, it is starting to double as a measurable signal of economic activity that researchers can study almost in real time, and it is forcing a level of financial discipline inside companies that didn’t exist even two years ago.
THE OUTLOOK
None of this guarantees the optimism embedded in today’s stock prices will be vindicated. Markets can be wrong, forecasts get revised again within months of being issued, and much of the data available today, including the OpenRouter usage records used in recent economic research, skews toward sophisticated users actively shopping for the cheapest model rather than the average subscriber on a flat monthly plan. But the direction is hard to miss. Tokens have gone from an obscure detail of how transformer models process language to the accounting unit for an entire economic system, one where usage, pricing, and value creation all have to be reasoned about together rather than separately. If Altman’s utility metaphor proves right, the businesses that learn to read that meter accurately, and act on what it says, will be the ones positioned to convert AI’s raw growth into something more durable.
References
Rosalsky, G. “AI Tokens Could Become the Kilowatt-Hour of the AI Age.” NPR, Planet Money, July 28, 2026. https://www.npr.org/sections/planet-money/2026/07/28/g-s1-135475/ai-tokens-could-become-the-kilowatt-hour-of-the-ai-age
Kindig, B. “AI Token Demand Is Shattering Forecasts.” I/O Fund, July 30, 2026. https://io-fund.com/ai-stocks/ai-token-demand-shattering-forecasts
Zhu, Q. “AI Tokenomics: The Economics of Tokens, Computation, and Pricing in Foundation Models.” arXiv:2606.24616, June 10, 2026. https://arxiv.org/html/2606.24616v1
Garey, L. “AI Token Economics: Measuring & Optimizing the Cost of AI.” Oracle, July 27, 2026. https://www.oracle.com/cloud/ai-token-economics-tokenomics/
Zivica, D. “The Token Economy and the Real Meaning of Modern Productivity.” Futurium, Apply AI Alliance, May 20, 2026. https://futurium.ec.europa.eu/en/apply-ai-alliance/community-content/token-economy-and-real-meaning-modern-productivity
