Uber Burned Through Its AI Budget for 2026. We Don’t Have AI Limits Either

Photo by Erik Mclean on Unsplash

I don’t trust Uber (after their workplace culture…issues), yet I’m still one of their customers. That’s because my values have a price, and someone coming to pick me up at the bar.

So I’ve never (and will never) apply for work there. I feel they’ve given engineers a lesson in workplace culture and management practices without ever needing to see it themselves.

So this week I was surprised to hear that they’re giving another lesson to developers and the wider software development community for free. If you give engineers unlimited access to tools, demand that they increase productivity and don’t measure how much of the “intelligence meter” they use you’re asking for trouble. And trouble Uber have earned.

Uber Lyft Off

Uber reportedly exhausted its annual AI budget in around six months. According to reports, the company is now reconsidering how it manages AI usage as token costs continue to mount, and perhaps even choose to not reward high token usage.

The same is happening at my workplace, but for us we had token limits of $4000 a month which were put into place last month.

So I’m not surprised about the path Uber have been on, because I’ve been on it too.

Uber Uber

Talking to a mid level developer today they said they were bumping up to the limit of their token usage.

I asked they why? What have they needed to do that would cost thousands of dollars in tokens? The answer is quite simple. They put their Claude instance on the super-max mode for all tasks “because it’s the best”. Rather than picking the best model for the task at hand or discriminating between tasks that require “max” effort.

So no wonder our collective bill for using AI increased. Yet is it any wonder? We are expected to make productivity gains using AI, so by definition we need to use the systems and approaches available to us.

Developers Must Use What You Give Them

I work with software developers every day. Give us a faster compiler and we’ll compile more code.Give us better hardware and we’ll run larger builds. We use the tools that we are given to ship, to improve our work and improve the features that are in the hands of customers.

This isn’t abuse. It’s what we are paid to do.

For years software companies have invested millions trying to remove friction from development. Then they seem surprised when removing friction actually increases demand. This means that Jevons Paradox can be fully applied to software development.

When something becomes cheaper and easier to use, people don’t necessarily consume less of it.

They consume more.

We are certainly getting much more code, building more and creating more. Of course there are costs in doing so, anyone who throughout this wouldn’t be true is surely misguided.

The Calculator Problem

I sometimes hear people argue that developers should use AI less, which feels rather like asking accountants to stop using spreadsheets.

Once people become accustomed to a productivity tool they rarely volunteer to give it back, as they need to keep working and discovering new ways of doing things.

Perhaps we have passed the inflection point since AI is now embedded in how many software teams work. We can’t simply go back and pull the plug at this point.

Need a regular expression? Ask AI.

Need a unit test? Ask AI.

Need to understand a codebase you’ve never seen before? Ask AI.

The individual requests might seem insignificant.

The cumulative effect is enormous.

Nobody Knows What Good Usage Looks Like

One challenge for companies is that AI spending is still new, which might well explain the rate which companies are happy to use money to power their software development systems.

It feels like Most organizations understand cloud costs. They can compute office costs, and salary costs. AI costs are different because they are directly connected to curiosity, and software developers should indeed be curious people (the good ones at least). Rather than asking a single question we ask twenty, exploring alternatives and generating examples. When you can’t quite trust the answer you might ask a question in many different ways. Pretty soon you’ve spent thousands of tokens solving a problem that previously involved searching Stack Overflow for half an hour. Whether the result is better or not seems to depend on whether the comparative Stack Overflow question was any good or not.

What Happens Next?

I suspect Uber won’t be the last company to have this problem.

As a company our budget caps have been introduced. Dashboards tracking token consumption, that have been introduced to measure rather than create a sense of competition between software developers. Slack channels about AI optimization are popping up..

But I don’t think usage will fall significantly, since the underlying demand hasn’t changed. Developers have discovered a tool that helps them work faster, they’re going to use it. Our velocity has been untouched, and we can’t justify going back to the old world and developing more slowly.

The real question is whether companies are prepared for what happens when every employee suddenly becomes more ambitious once again. And the token costs are reviewed by the providers.

Conclusion

Changing the way we work has turned out to be a much more expensive problem than anyone ever expected.

About The Author

Professional Software Developer “The Secret Developer” can be found on Twitter @TheSDeveloper.

The Secret Developer asked Claude what the time is.