When the AI Bill Comes Due

read on Forbes

Ten months ago, Uber initiated a big AI push across its organization. Thousands of engineers got access to tools like Claude Code and Cursor, and management encouraged them to use the technology aggressively. The goal was to make AI part of the way software actually got built, rather than something people played around with on the side.

It worked. Within a few months, AI tools quickly became part of everyday life at Uber. Engineers were using agents to write code, test it and fix bugs. Some of that code was going directly into the systems Uber uses to match riders with drivers and determine prices. The more capable the tools became, the more people used them. At one point, Uber even had internal leaderboards showing who was using AI the most.

But then the bills started showing up.

By April, the company had blown through its entire AI budget for 2026. As Uber CTO Praveen Neppalli Naga related to The Information, “I’m back to the drawing board.” Uber had spent months trying to accelerate AI adoption and succeeded far beyond its own financial assumptions. Now it had to figure out how much of that success it could afford.

Token Legends

Uber’s experience is an extreme version of something that’s been happening across corporate America. For the last couple of years, leaders have been under enormous pressure to move faster on AI. Boards are pressuring management teams to lean hard into the evolving technology. CEOs worry about competitors getting ahead. Employees are encouraged to experiment with every new tool that appears. Somewhere along the way, the amount of AI budget getting used started to look like evidence that a company was getting more sophisticated.

Meta turned that impulse into an internal sport. Earlier this year, a Meta employee created a dashboard called “Claudeonomics” that ranked AI usage across more than 85,000 people. Employees could earn badges based on how many tokens they consumed, with the heaviest users competing for titles like “Session Immortal,” “Cache Wizard,” and, best of all, “Token Legend.” Over a thirty-day period, the dashboard recorded more than 60 trillion tokens of AI usage.

There was a certain logic to all of this. Most people inside large organizations were still trying to understand what AI could do. Leaders wanted them to dive in and gain proficiency with the tools. A rising token count offered visible evidence that behavior was changing, even if no one could point to real outcomes.

Fast forward a few months, though, and the conversation seems to have shifted. Finance teams are looking harder at AI spending. Execs are proposing limits on expensive tools and moving simpler tasks toward cheaper models. The same organizations that spent the last two years pushing people to use more AI are beginning to ask how much of it they really need. The great AI adoption race has finally reached the finance department.

When the Party Ends

Part of the problem was an issue of incentives. Companies understandably went looking for help from the people who know AI best. Companies like Palantir and Google pioneered the idea of “forward-deployed engineers” who embed directly with customers and build solutions against real problems.

That approach can be extraordinarily valuable. Most large organizations don’t have enough people who understand what the technology can actually do, and an engineer sitting beside a business team can often see possibilities that would never emerge from a vendor presentation.

But there’s an unavoidable incentive problem in that approach. A company that makes money when customers use more AI has a particular perspective on what a good AI strategy looks like. It’s a little like asking Colonel Sanders to develop your chicken strategy.

The abrupt shift in mindset is also a result of learning how AI economics works. Traditionally, enterprise software has been pretty straightforward to budget: a company buys a certain number of licenses at a fixed annual price. AI is different because every additional request consumes real computing power, and an agent can keep calling a model for minutes or hours as it researches, writes, tests and revises. As those costs have become more visible, enterprise pricing has increasingly begun to reflect how much AI people actually consume. In effect, AI companies are starting to charge their biggest customers a little less like a cable company and a little more like an electric company.

There’s a strange irony in Uber finding itself on the receiving end of this particular problem. In its early years, the ride share company used enormous amounts of investor capital to subsidize rides and drive adoption. An entire generation got used to tapping a phone and having a car appear. Even in places like New York, plenty of young adults grew up barely knowing how to hail a cab. Eventually, though, the subsidies had to end. Uber needed the economics of each ride to work on its own, and customers began paying more for a convenience they had come to depend on.

AI is following an eerily familiar path. For the last few years, enormous amounts of venture capital have helped subsidize the cost of getting companies hooked on tokens. Now the providers need those economics to work, too. The difference is that this time the bill is landing on companies rather than consumers. And one of those companies, fittingly enough, is Uber.

Asking the Wrong Question

This kind of correction is healthy. Companies should know what all this AI is costing them. They should know which applications are doing something useful and which ones are just chewing through tokens.

But there’s a danger that we replace one bad idea with another. For the last two years, companies have acted as if using more AI meant they were somehow becoming more AI-enabled. Now the pendulum is swinging back, and spending less may start to look like evidence that management has finally gotten disciplined.

The more powerful question is: What could this technology allow our company to become? Present-focused leaders see AI as a cost line attached to current work. Future-focused leaders see AI as a new ability that might change what work exists, what customers value, and even what business their company is in. That’s a strategy question, not a technology question.

A few months back, I was talking with the CTO of a large automobile company. They were understandably excited about the possibilities that AI offered. Maybe the company could use AI to help customers explore and select from an increased variety of car options.

I could see why that would be appealing. An AI could learn someone’s taste, show them different combinations and offer a recommendation. It could make the purchase experience feel a little easier and more personal. It just won’t change the business very much.

By the time a customer is choosing seat fabrics, the car company has already won the much more important decision. That person has already decided to buy the car. AI might make the last few steps of the current experience a little smoother, while leaving untouched the much larger problems that will determine whether the company succeeds ten years from now.

Compare that with the challenge Volkswagen is facing right now. The company knows it has to get dramatically smaller. Its Future Plan 2030 envisions reducing the workforce by roughly 50,000 positions while simplifying the organization and cutting excess capacity.

A company can approach that kind of challenge as a giant exercise in subtraction. Every function gives something up. Plants lose capacity. Business units cut costs. Headcount comes down until the economics begin to work again. That may produce a smaller Volkswagen. The harder question is whether it produces the Volkswagen the future requires.

This is where AI starts to get interesting. Volkswagen could use it to build a much more forward-looking picture of where demand is actually moving, then use that picture to decide where capacity should come out and where investment still needs to go.

The people question may be even more important. If Volkswagen knows its workforce will be materially smaller ten years from now, AI could help it work backward from the capabilities it will need in 2035. Which skills will matter more? Which ones will matter less? Where should people be retrained now, while there is still time? Those are decisions about the future shape of the company. AI matters because it gives leaders a new ability to make them with more foresight.

What AI Demands of Us

The underlying problem may be that we still think of artificial intelligence as a tool. Tools help us do the things we already know how to do. AI may turn out to be something more consequential because it can change the things we decide to do in the first place. And the more consequential that ability becomes, the more responsibility falls back on us.

This week, Anthropic alignment researcher Evan Hubinger said he personally puts the probability of AI causing human extinction within the next decade at greater than 10 percent. He was responding to another Anthropic researcher who had resigned after growing increasingly concerned about the technology he was helping to build.

Reasonable people can disagree with Hubinger’s estimate. What’s extraordinary is that people working at the frontier of this technology consider the risk serious enough to put a number on it. And that makes the debate over token budgets feel awfully small.

Companies still need to decide what AI is worth spending money on. But those choices sit inside a much larger one. We’re introducing a new capability that could change the companies we build, the work people do, and the role human judgment plays in all of it. Eventually, the question reaches all the way to us. How does AI demand that we evolve as humans?

Technology doesn’t get to answer that question. We do. The future we build with AI will reflect the choices people make about how it is used, where we allow it to lead, and where human beings still need to take responsibility. Future-focused leaders have always had to take responsibility for futures that haven’t arrived yet. AI simply raises the stakes.

When the AI bill comes due, it will include a lot more than money.