Money for nothing

In the spring of 2025, Tobi Lütke, Shopify's co-founder and chief executive, posted to X an internal memo he had written to his staff, partly because it was already leaking and he preferred to lose control of it on his own terms. The subject line read like a verdict: "Reflexive AI usage is now a baseline expectation at Shopify." Before anyone could ask for a new hire or a bigger budget, the memo instructed, they would first have to show why artificial intelligence could not do the job instead. AI fluency would be written into performance reviews. Within days the thing had become a genre — Box wrote its own, then Fiverr, then, improbably, the office of the prime minister of Canada.

Mr Lütke had dropped out of school after the tenth grade in Koblenz, apprenticed as a programmer, and followed a woman he met on a snowboarding trip to Ottawa, where he built one of the largest commerce companies on earth. Like much of the cohort that now runs the platform businesses, he had arrived at a conviction: the binding constraint on a company was no longer headcount or capital but the speed at which its people learned to use the machine. The memo's real content, beneath the management prose, was a mood, and the mood was unmistakable — spend, adopt, mandate, and do it now, before the firm down the road does it first. For a year that mood ran the boardrooms. Then someone asked the rude question.

The rude question was whether any of it had worked, and in the autumn of 2025 a group of researchers at MIT's NANDA initiative supplied an answer that nobody at the Lütke end of the conversation wanted to hear. Of some three hundred enterprise AI deployments they examined, 95% had produced nothing their own executives could find on the profit-and-loss statement. The report was titled "The GenAI Divide," and the number traveled exactly as the memo had, screenshotted into every group chat in the industry and mostly read as a single word: bubble. Aditya Challapally, the report's lead author, kept trying to redirect the conversation. The failure, he and his colleagues argued, lay not in the models but in the wiring around them; generic chatbots dazzled an individual and stalled inside an enterprise because they never learned the actual work. The pilots had not failed because the intelligence was weak. They had failed because almost no one could say what the intelligence was worth.

There was a quieter finding in the report, the kind that does not screenshot well. While the official pilots stalled, the same companies were running what the authors called a shadow AI economy — employees paying for ChatGPT out of their own pockets and using it to do their actual jobs, off the books and off the budget. The tools were working. The measurement was not. A company could mandate AI, as Mr Lütke had, and still have no idea, eighteen months later, whether the mandate had bought it anything at all.

The price of everything

This was the part the consensus kept missing, because the cost of AI was the one thing everyone could already see. A token meter is exact; it counts to the fraction of a cent, and the bill arrives, itemized, every month. The difficulty was never the numerator. It was the denominator, the unit of value the spending was supposed to produce, against which any return has to be judged, and which artificial intelligence, alone among the great corporate costs, arrived without. A wage buys an hour of labor. A vendor contract buys a defined deliverable. Each comes with a native unit of output baked in, a thing a twenty-five-year-old in financial planning can count on sight. A token buys "intelligence," which has no native unit at all. A sales team is measured in qualified leads and a help desk in resolved tickets because someone, at some forgotten point, decided those were the units and built the machinery to count them. AI spend inherited no such decision. So finance measured the half that counts itself, called the ratio ROI, and left the half that mattered blank.

Consider the finance chief at a quarterly review in the spring of 2026, working down the operating expenses the way she has for a decade. Payroll resolves itself; she knows the headcount and the comp bands. Software resolves itself; she knows the seats and the contracts. Then she reaches the AI line, which has tripled since the last review, and finds a single figure invoiced by a frontier lab, one number, no breakdown, attached to no unit she has ever been taught to count. She can see, to the penny, what intelligence cost the company last quarter. She cannot see what it did.

Where there is a blank on a corporate ledger, there is a market. A category that did not exist two years ago, AI FinOps, borrowed from the older discipline of cloud-cost management, acquired its own vendors almost overnight: Vantage, Finout and CloudZero, firms that plug into the billing of OpenAI and Anthropic and slice the spend by team and feature. The spend-management platforms moved the same way, and faster. In June 2026 Ramp, the corporate-card company that had climbed from a $13 billion valuation to $44 billion in fifteen months, raised another $750 million behind a letter from its chief executive, Eric Glyman, arguing that "token spend" had become the third great category of corporate cost, after people and vendors, and that whoever made it visible would own it. The instinct was sound, as far as it went. You cannot manage what you cannot see. What the pitch declined to dwell on was that seeing the cost is the easy part, that the cleanest token data sits with the very labs whose meters are running, and that visibility into a bill has never, on its own, told anyone whether the bill was worth paying.

The dull dividend

The MIT data, read past its headline, said where the value actually was, and it was not where the money had gone. More than half of corporate AI budgets had flowed into sales and marketing, the glamorous front office, where the returns ran thinnest. The largest measurable gains had landed in the back office, in the unglamorous work of operations and reconciliation and document processing, where the output units were cleanest and a saved hour could be counted. The 5% on the profitable side of the divide were not the firms with the richest models or the flashiest dashboards. They were the ones that had named a unit of value and then managed against it: a deflected support call, a merged pull request, an hour of an analyst's week handed back. Companies that bought a tool and integrated it succeeded about three times as often as those that built their own. Measurement, it turned out, was a management act, a decision about what counts as work, long before it was anything a vendor could sell you.

The blank is about to get wider. The spending that finance cannot yet measure is the spending humans authorize; the next wave, already arriving, is autonomous. Agents have begun to hold corporate cards, source vendors and clear payments without a person in the loop, which means the meter will soon run on decisions no employee consciously made. Tokens are dollars and agents are hires, the spend-management firms like to say, and the line is a good one. It is also a confession. A company that cannot yet say what a token bought will shortly be asked what a thousand autonomous agents bought, in a quarter, across every function at once.

There is a respectable objection to all of this, which is that ROI was the wrong lens from the start. No one ran a payback study on electricity before wiring the factory, or on the spreadsheet before buying the first copy of Lotus 1-2-3; general-purpose technologies earn out through a slow reorganization no quarterly model can see, and a company that demands a clean number in year one will strangle the projects most worth keeping. The objection is fair, and the firms that succeeded did so mostly by integrating patiently rather than chasing a fast return. But a downturn does not wait for reorganization, and the budget line with no number beside it is the first one a board crosses out. The absence of a unit is not neutrality. It is a vacuum, and a vacuum hands the defining of value to whoever ships the dashboard, the vendor selling the meter or the lab whose meter is the only instrument in the room.

The lag is older than the technology. Robert Solow, the economist, remarked in 1987 that the computer age was visible everywhere except in the productivity statistics, and the statistics stayed flat for years afterward. Paul David, an economic historian at Stanford, explained why in a celebrated 1990 paper about the electric dynamo. Factories at the turn of the century had electrified by bolting a single motor where the steam engine used to sit, and the productivity gains did not arrive until the 1920s, decades on, when a new generation of managers finally redesigned the factory floor around the idea that each machine could carry its own motor. The dynamo was never the bottleneck. The reorganization was, and so was the long interval in which nobody yet knew what to measure. The unit of value, in each case, was invented rather than discovered. What the companies now squinting at their token bills are attempting is that same invention, run at speed, under a chief financial officer's quarterly deadline rather than an economist's leisurely hindsight.

Oscar Wilde's cynic, in the old line, knew the price of everything and the value of nothing. Two years into the spending, the modern company has the first half cold; the meter was never the hard part. The verdict is. Mr Lütke's memo, and the genre it spawned, ordered the spending before anyone had built the scoreboard to keep it, which is the ordinary way of these things, and was the ordinary way in 1920, too. The firms that come out on the right side of the divide will not be the ones that spent the most, or the least. They will be the ones that decided, before the bill came due, what a token was supposed to buy, and then had the discipline to check.