Uber’s CTO just handed every CFO in tech a warning they did not ask for. Praveen Neppalli Naga publicly confirmed that his company burned through its entire 2026 AI budget in four months. The culprit was Claude Code, Anthropic’s agentic coding assistant, and the story of how it happened tells you something important about how AI adoption actually plays out at scale.

What Happened at Uber

Uber rolled out Claude Code access to its 5,000-strong engineering team in December 2025. By February, adoption had climbed to 63% of engineers using it monthly. By March, that number reached 84%. The tool spread through the organization faster than any software Uber had deployed in recent memory.

The financial consequences followed directly from that adoption curve.

Monthly API costs per engineer ran between $500 and $2,000, depending on usage intensity. Multiply that across thousands of engineers and you get a bill that looks nothing like a traditional enterprise software licence. By April, Naga confirmed the obvious: the budget he had set for the full year was gone.

“I’m back to the drawing board because the budget I thought I would need is blown away already,” he said.

Today, 95% of Uber engineers use AI tools at least monthly. Seventy percent of committed code at the company now originates from AI. Eleven percent of live backend updates are written by AI agents with no human in the loop.

Those are not pilot numbers. That is a production-scale shift in how Uber builds software.

Why the Budget Math Was So Wrong

The budgeting failure was not really about Claude Code being expensive. It was about a fundamental mismatch between how enterprise software is priced and how AI tools are consumed.

Traditional software licences are predictable. You pay a flat fee per seat, cap the number of seats, and the cost is bounded. You might forecast growth, but the cost per unit is fixed.

Claude Code does not work that way. It runs on token consumption, which means the invoice is a direct function of how much the model processes across all engineer sessions. The more engineers use it, the more aggressively they use it, and the more complex the tasks they give it, the higher the bill. There is no per-seat ceiling. If your engineers are running it for hours every day on large codebase tasks, the costs can scale faster than adoption metrics alone would suggest.

Uber also made an internal management choice that accelerated the burn. The company tracked and ranked engineer usage of Claude Code on internal performance dashboards. That kind of visible adoption metric creates cultural pressure. Engineers who used Claude Code more appeared, implicitly, to be performing better. Engineers who used it less stood out.

The result was exactly what you would expect: adoption accelerated, usage intensity increased, and the token-based bill grew faster than the headcount growth alone would predict.

What This Means for Business

Uber is not a cautionary tale about AI tools being too expensive. If 70% of committed code comes from AI, and 11% of backend updates run without human review, the productivity gain is real. The CTO still thinks the tools are worth it. The problem is not the technology, it is the budget model that was applied to it.

For any business leader thinking about rolling out AI coding tools or AI productivity tools more broadly, the Uber story surfaces a few practical questions worth answering before you start:

Is your pricing model per-seat or consumption-based? Most AI tools today are consumption-based. The implication is that your costs scale with usage, not just headcount. Budget accordingly and build in meaningful headroom.

What cultural signals are you sending about AI adoption? If you make AI usage visible on performance dashboards, you will get more usage. That is probably the right outcome, but it should be a deliberate choice, not an accidental one, because it directly affects spend.

Do you have visibility into usage at the individual engineer level? Without that data, you cannot distinguish between high-value usage (complex tasks that save hours of work) and low-value usage (running AI on tasks that don’t justify the token cost). Neither Uber’s success nor its budget surprise would have been legible without real usage data.

What is the actual productivity gain, and is it worth the cost? Uber has not said the answer to this is no. Naga is “back to the drawing board” on the budget, not on Claude Code. That distinction matters.

The broader lesson here is not unique to Uber or to Claude Code. It is about the gap between piloting AI tools and scaling them. Most companies are still operating in pilot mode, which means their budget models reflect pilot assumptions. When adoption hits the whole engineering organisation, or the whole sales team, or the whole finance function, those assumptions fail.

The companies that navigate this well are the ones that treat AI budgeting as a first-class operational discipline, not an afterthought to a software rollout.

The Next Phase of AI Adoption

Uber’s budget story is a symptom of where the market is, not where it is going. The AI tools that drove this spending are genuinely transforming how software gets built. The pricing models will likely evolve as enterprise procurement teams push back and as vendors compete on cost.

Anthropic has strong commercial incentives to develop enterprise pricing that gives large customers more predictability. Per-seat options, consumption caps, and volume tiers are the obvious mechanisms. That pricing evolution will happen on a timeline set partly by what enterprise customers like Uber demonstrate is necessary.

For now, Uber’s budget blow-out is the most concrete public data point on what it actually costs to fully adopt AI coding tools across a large engineering organisation. It will not be the last.

---

For a deeper walkthrough of tools like this and how they fit together, the free Working With Claude field guide covers the ecosystem end to end. Get the guide.