Your AI Bill Is Going Up. That's Not the Problem.

The price of AI keeps dropping while the total bill keeps climbing, and most leaders are focused on the wrong half of that.

The cost of a given unit of AI is collapsing. The total amount companies spend on it is going up anyway. Both are true. If you're the one who has to answer for that line item, your instinct is to ask why the bill keeps growing. I'd argue that's the wrong thing to worry about. What you should worry about is whether you can show what all that money actually bought.

Uber Torched Its Entire AI Budget in Four Months

Uber burned through its entire 2026 AI budget in the first four months of the year. The tools didn't get more expensive. Adoption did. The share of Uber engineers using Claude Code jumped from 32% in February to 84% by March, and by spring around 95% of them were using AI tools every month. More people, more usage, a much bigger bill. The CTO said they were back to the drawing board on their budgeting assumptions.

You might be thinking this is a big-company problem. Uber has thousands of engineers and a budget to match, and you don't. But what blindsided them has nothing to do with headcount, and a smaller company is arguably in more danger. Uber at least caught it. They had the scale and the finance scrutiny and the dashboards to spot the overrun inside the same year. A lot of mid-sized companies have AI charges spread across a dozen team credit cards and nobody whose job is to keep an eye on the total.

The price per token had nothing to do with Uber's overrun. The unit cost was fine. What changed was how people used it, and that kind of shift happens at any size.

The Price of Intelligence Is Collapsing

What makes the budget story strange is that AI has never been cheaper. When GPT-3 first came out, running it cost about $60 for a million tokens, call it a million words of text. Today you can get the same quality of output, or better, for around six cents. That's a 1,000x drop in a few years, and it's still dropping. Andreessen Horowitz calls it LLMflation. The price of a given level of intelligence keeps collapsing.

So if the price per use keeps falling, why is everyone's bill going up?

Because Cheaper Never Means Less

Cheaper access doesn't make us spend less. It makes us use more. Economists have a name for this, the Jevons paradox, and AI fits it almost too well. When something useful gets cheaper, we rarely bank the savings. We find new places to spend them.

So we add another agent. We automate another workflow. We let AI handle a little more of the process than it did last quarter. The bill climbs even as the per-unit cost drops, because our appetite grows faster than the price falls.

There's a human side to this too. Almost nobody downgrades once they've used the best model. You get access to the smartest option available, you get used to the results, and you keep reaching for it even on work a cheaper model would have handled fine. That habit is hard to break, and it shows up on every invoice.

The Harder Question Is What You Got for It

A few weeks after the budget story broke, Uber's COO, Andrew Macdonald, admitted something harder. With most of the company's code now AI-generated, he still couldn't draw a clean line from that spending to better outcomes: features shipped, bugs fixed, customer problems solved. He reportedly called it a head-exploding moment.

That's the question your board will eventually ask you, so it's worth having an answer ready. A big AI bill isn't a crisis on its own. Plenty of expensive things earn their keep. The danger is spending at that scale and having nothing convincing to show for it. Uber can eat a budget miss and move on. A mid-sized company that has wired AI into a dozen workflows without measuring any of them is betting a much bigger share of its resources, with far less room to be wrong.

How to Avoid Uber's Spot

A few things I'd tell any leader trying to stay out of that position:

• Decide what you're measuring before you scale. Uber ranked its engineers on internal leaderboards by how much AI they used, so usage is what everyone optimized for. Measure adoption and adoption is what you get. Tie the money to something the business actually feels: hours saved, tickets closed, cycle time, revenue touched. If you can't name that metric up front, you're flying blind.
• Stop sending every task to your most expensive model out of habit. Some work needs the frontier. Most of it doesn't. Claude Fable 5 is available again after its export-control pause, and at $10 per million input tokens and $50 per million output it's worth reserving for the jobs that actually need it (same goes for GPT-5.6 and Gemini 3.5 Pro when they arrive). The companies getting this right match the model to the task, which means teaching your people how to make that call instead of defaulting to the priciest option.
• Revisit old workflows every few months. An agent that barely paid for itself in January can often run today on a model that costs a fraction of what it did back then. Most teams build it once and never look again, so they overpay for months without noticing. When the cost of intelligence drops something like 10x a year, set-it-and-forget-it is a slow leak in the budget.

None of this happens on its own. It takes someone actually watching the meter: pointing work at the right model, shutting down the automations that stopped earning their keep, checking that the spend still ties back to real results. At Uber's scale that's a full-time job. At yours it's probably a responsibility you hand to someone already on the team, as long as you say so out loud and back them up when they push to cut something. Do it before your CFO asks what the AI line bought and nobody in the room has a good answer. The models keep getting cheaper. The mistake is assuming that lets you stop paying attention.