Uber Spent Its Whole Year's AI Budget by April

Here's a sentence that should make every finance leader sit up: one of the most AI-forward companies on earth ran out of AI money four months into the year.

Uber burned through its entire 2026 budget for AI coding tools by April. Not because someone fat-fingered a contract – because the company actively encouraged heavy usage. There was an internal leaderboard ranking engineering teams by how much AI they used. Engineers did what leaderboards make people do. They climbed it. The bill climbed with them.

Then came the part that should concern CFOs more than the overspending itself. On the Rapid Response podcast, Uber COO Andrew Macdonald was asked whether all that usage was actually producing more for customers. His answer: "That link is not there yet." He went further – if you can't draw a straight line from the spend to the features you're shipping, "that trade becomes harder to justify."

The company spent the money, blew the budget, and its own operating chief isn't sure what it bought.

This Is a Token Problem, Not a Software Problem

For years, software was the predictable line on the budget. You picked a tier, multiplied by seats, and you knew your number for the year: stable and boring. Exactly what a CFO wants a cost line to be.

AI coding tools don't work that way. Anthropic's Claude bills on consumption – input tokens, output tokens, cache reads and writes, runtime sessions, tool calls. It behaves less like a subscription and more like a utility bill, the kind that moves every month depending on how much you ran the lights.

OpenAI's Sam Altman has been direct about where this is heading: he sees a future where intelligence is sold "like electricity or water," metered by use. That's a fine vision if you're selling the electricity. It's a harder one if you're the person forecasting the bill.

And the bill can move even when usage holds steady. Anthropic's documentation notes that its newer Opus tokenizer can consume up to 35% more tokens for the same text as earlier models. The same prompt, the same work, can cost more simply because the model under the hood changed. Upgrades that look free on the surface can quietly reprice your whole operation.

Cheaper Tokens Won't Bail You Out

The standard reassurance is that AI gets cheaper every year, so this all sorts itself out. It doesn't.

Gartner projects that running these models will cost AI companies about 90% less by 2030. That sounds like relief – but the same research points to two reasons enterprise bills won't follow those unit costs down. First, the new agentic tools consume far more tokens per task than older chatbots did, so per-unit savings get swallowed by volume. Second, providers aren't going to pass all those savings through to enterprise customers.

The result is the situation Uber landed in: per-unit prices fall, the invoice grows, and both things are true at once. That combination is precisely what catches finance teams off guard.

It's Already Bigger Than People Think

If this still sounds like a Fortune 500 problem, look at what smaller finance teams are quietly discussing with each other. In FP&A forums, people describe checking their Claude costs daily because it's become one of the largest variable cost lines in the business. Individual employees are running $100 to $300 a month. Engineering-heavy users are hitting several thousand.

Even Microsoft – as technical and AI-literate as any company on earth – reportedly pulled back its internal Claude Code licenses after token costs ran up, shifting engineers onto a cheaper alternative. If a company that builds AI for a living couldn't keep the meter under control, a thirty-person startup running its books in QuickBooks is not going to find it straightforward.

So What Does a Finance Leader Actually Do?

The reflex answer is governance – spending caps, usage dashboards, monthly reviews. Those help. But they treat a symptom.

A cap tells you when you've hit a wall. It doesn't tell you why the number moved, whether the spend was worth it, or what to renegotiate before the renewal. A dashboard shows you the line going up. It doesn't sit in the room and defend that line to your board.

People do that work. The companies losing the plot on AI spend usually aren't the ones who adopted too aggressively. They're the ones who layered a new, volatile, hard-to-trace cost on top of a finance team that was already stretched thin – and expected someone to "keep an eye on it" in the margins of their actual job.

What's missing is a person whose job is to pull the usage data, tie it back to teams and outcomes, build a run-rate model that holds up, and walk into the renewal conversation knowing the numbers cold. The kind of mid-level finance operator who turns "I don't know, everyone's using it more" into a one-line answer the CEO will accept.

Uber's COO said the link between the spend and the value isn't there yet. He's right. Someone has to build that link. The question for your business is whether that person exists on your team – or whether you're the one doing it at 11pm during close.