Uber Technologies set an artificial intelligence budget for 2026 that was supposed to last the year. It lasted four months. Andrew Macdonald, the company’s president, said publicly that Uber exhausted its AI budget for the year by the end of April, and that connecting the enormous volume of tokens the company consumes to actual shipped product features has become increasingly difficult to justify. The confession, delivered in an unguarded moment on a public stage, set off a vigorous debate on Hacker News and The Verge, where it was read as a signal about the whole industry’s AI spending math.
The scale of Uber’s AI usage makes the admission significant. The company runs operations that touch millions of people every day, and it has been wiring models into everything from customer support and trip dispatch to advertising and internal software development. At that volume, token consumption becomes a line item that finance teams cannot ignore, and the bills grow with every feature that leans on a model. Mr. Macdonald’s remarks suggest that even a company with Uber’s resources is now asking a question that smaller enterprises are just beginning to face: what, exactly, is all this compute buying?
The question is harder to answer than it sounds. Some AI spending produces visible results: a support chatbot that deflects calls, an ad system that prices better, a dispatch model that shortens wait times. Much of it, however, goes to experimentation, prototyping, and models that improve incrementally without producing a single shippable feature. The gap between token consumption and product output is where the industry’s financial anxiety lives, and Mr. Macdonald’s phrasing, that the spending is getting harder to justify, captured the mood of a generation of technology executives.
The reaction to his comments was immediate. On Hacker News, engineers debated whether Uber’s problem was wasteful engineering or a structural feature of the AI era; on The Verge, the story was framed as a reality check for a market that has priced in unlimited AI upside. The fact that the remarks came from a senior executive at a company with annual revenue in the tens of billions of dollars gave the discussion an authority that anonymous complaints lack. If Uber cannot map tokens to products, the reasoning went, what chance does the rest of the economy have?
The industry’s response has been to attack the cost side of the equation. Model makers have raced to release cheaper, faster models, and enterprises have discovered that much of their AI traffic can be routed to smaller models or served from cached responses. Reasoning models that think for seconds per query are being reserved for the hardest problems, while routine work moves to cheaper tiers. The efficiency wave is real, but it is also a response to the problem Mr. Macdonald described: the default way of using AI, big models for everything, is too expensive for a business that has to show a profit.
The investment side of the ledger is no simpler. Uber has said it plans to spend heavily on AI in the years ahead, and its investors have generally supported that direction, betting that the company’s data and distribution give it an edge in deploying models at scale. But the same investors are increasingly asking for evidence that the spending compounds into durable advantage rather than recurring expense. Mr. Macdonald’s remarks, whatever their intent, gave them a data point: the company’s own executives are not certain the equation works yet.
Mr. Macdonald’s remarks also highlight a governance question that boards across the industry are starting to confront. AI spending at large companies has grown faster than the processes designed to oversee it, and the budgets, often assembled bottom-up from team requests, can lack the discipline applied to other capital outlays. Some companies have responded by centralizing AI procurement, requiring business units to justify model usage against measurable outcomes. The approach is straightforward in principle and politically difficult in practice, because every team believes its usage is essential.
The company’s own response to the squeeze is still taking shape. Uber has said it is routing more of its traffic to cheaper models and building systems to reuse responses rather than recompute them, the standard toolkit of the efficiency wave. Mr. Macdonald’s public candor may also serve an internal purpose: by acknowledging the budget problem in public, he has made it harder for teams inside the company to treat token consumption as an unlimited resource, and easier for finance to impose discipline. In that reading, the confession was less a warning to investors than a signal to the organization.
For the rest of the industry, the episode is a preview. Every company that has adopted AI at scale will eventually face the same accounting, the same question about whether tokens translate into value. The ones that answer it well will be the ones that built measurement into their AI programs from the start, tracking not just usage but the revenue and savings each use case produces. Uber’s four-month budget is an early, honest snapshot of a problem that is only going to get bigger as the models get more capable and the bills get larger.
