Imagine throwing a party where every guest orders top-shelf liquor on your tab, except the guests are AI coding assistants and the tab is measured in millions of tokens processed per hour. Now imagine you budgeted for the party to last all year, but by April the bartender is cutting you off. That’s essentially what happened to Uber in 2026, and they’re far from alone.
Uber Burned Through a Year’s Budget in Four Months
Let’s start with the number that made me spit out my coffee: Uber blew through its entire 2026 AI coding budget by April. Not Q3. Not the holiday crunch. April. That’s a company with serious engineering resources and presumably serious financial planning, and they still couldn’t predict how fast their developers would consume AI tokens once the tools were available.
This isn’t a story about irresponsible spending. It’s a story about how fundamentally unpredictable AI usage patterns are when you hand developers powerful coding assistants and say “go nuts.” Turns out, they go nuts.
Microsoft Pulled the Plug on Claude Code
Microsoft’s response was arguably more dramatic. After enabling Claude Code licenses for its developers, the company revoked them months later. Think about that from a workflow perspective. You give your engineering teams access to a tool, they integrate it into their daily process, build habits and dependencies around it, and then you yank it away because the costs spiraled beyond what anyone modeled.
I review AI tools for a living, and I’ve never seen a clearer signal that the industry’s cost models are broken. These aren’t startups burning through VC money with no accountability. These are two of the largest technology companies on earth, and they both got blindsided by the same problem: AI tokens are expensive, developers are hungry for them, and nobody built adequate guardrails before opening the floodgates.
The Governance Vacuum Isn’t Helping
Meanwhile, the regulatory picture is scattered at best. David Sacks confirmed on March 26 that his 130-day term as White House AI and crypto czar expired, and the administration isn’t appointing a replacement. So the one federal role ostensibly tasked with thinking about AI policy at the highest level just evaporated. Labor unions are already raising alarms about what deregulation could mean for workers displaced by these tools.
On the state level, Massachusetts Governor Maura Healey announced the Mass Wins Act on April 16, a $305 million economic development bill designed to attract defense and AI growth. That’s real money aimed at expanding AI infrastructure, but without federal cost governance frameworks, you have to wonder whether those investments will face the same runaway spending problems that hit the private sector.
My Take as Someone Who Tests These Tools Daily
I spend my weeks evaluating AI agents and coding assistants. I measure their output quality, their speed, their reliability. What I haven’t been doing enough of, and what I suspect most reviewers haven’t been doing enough of, is tracking cumulative token costs over realistic usage periods.
A tool that saves a developer two hours a day sounds like a bargain until you realize it’s consuming $400 worth of tokens in the process. The per-query cost looks small. The monthly bill at scale looks like a mortgage payment on a house you don’t own.
The industry needs three things immediately:
- Usage caps with transparency. Developers should see real-time cost dashboards, not discover overruns after the fact.
- Tiered access models. Not every coding task needs the most expensive model. Routing simple completions to cheaper models while reserving premium inference for complex problems is basic resource management.
- Honest benchmarking. Tool reviewers, myself included, need to start publishing cost-per-outcome metrics alongside quality scores.
Where This Goes Next
The 2026 AI cost crisis isn’t a temporary blip. It’s the first real stress test of whether organizations can sustainably adopt AI coding tools at scale. Uber and Microsoft are just the companies big enough and transparent enough for us to hear about their struggles. Hundreds of mid-size companies are quietly dealing with the same shock.
I’ll be updating our reviews at agnthq.com to include long-term cost tracking for every AI agent and coding tool we evaluate. Because a tool you can’t afford to keep running isn’t a tool. It’s a demo.
The party was fun. The tab is due. And the industry is sobering up fast.
🕒 Published: