Uber gave 5,000 engineers access to Anthropic's Claude Code in December 2025. By mid-April, the company's entire annual AI budget was gone. Not reduced. Not under pressure. Exhausted. Four months into a twelve-month budget cycle, Uber's CTO was, in his own words, "back to the drawing board."
Then, on May 26, Uber's president went further — telling a podcast audience that the company can't actually connect the spending to better products.
This is the most important AI story of the month, and it has nothing to do with model capabilities.
What happened at Uber
The rollout was, by every adoption metric, a resounding success. Usage of agentic coding features surged from 32% in February to 84% by March. By April, 95% of Uber's engineers were using AI tools monthly. Roughly 70% of code commits involved AI assistance, and around 11% of live backend updates were being written by AI agents with no human in the loop.
CTO Praveen Neppalli Naga confirmed to The Information that the full annual AI budget was exhausted by mid-April. "I'm back to the drawing board, because the budget I thought I would need is blown away already," he said.
Per-engineer monthly API costs ballooned to between $500 and $2,000. Naga himself reportedly spent $1,200 in a two-hour personal demo session. For context, Uber's total R&D spending hit $3.4 billion in 2025, up 9% year-over-year. AI-related costs have risen approximately six-fold since 2024.
The kicker: Uber had set up internal leaderboards ranking engineers by AI tool usage — a management choice designed to drive adoption that worked exactly as intended. In a token-based billing environment, that incentive structure directly translates into budget acceleration. The people who designed the leaderboard were almost certainly not the people responsible for the AI services budget line.
"That link is not there yet"
On May 26, Uber President and COO Andrew Macdonald gave the quiet part a voice. In an interview with the Rapid Response podcast, he said the company cannot connect rising token consumption to measurably better consumer products.
"It's very hard to draw a line between one of those stats and, 'Okay, now we're actually producing 25 percent more useful consumer features,'" Macdonald said. "I think over the coming quarters and years, maybe that will become clearer, but I think today it's hard, even if some of the underlying metrics are trending in a really astronomical direction."
Macdonald described the moment Naga's budget disclosure went public as a "head-exploding moment" that triggered a company-wide recalibration. CEO Dara Khosrowshahi had already said on an earnings call that Uber was slowing hiring to fund AI investments — a trade-off that becomes significantly harder to justify when the AI spending itself can't be linked to output.
The term circulating inside Uber for this dynamic? Tokenmaxxing — rising AI token consumption without proportional gains.
Microsoft is doing the same maths
Uber isn't alone. On May 14, Microsoft told thousands of engineers in its Experiences and Devices division — the teams building Windows, Microsoft 365, Outlook, Teams, and Surface — to stop using Claude Code and switch to GitHub Copilot CLI by June 30.
The official reason is "toolchain unification." The deadline is the last day of Microsoft's fiscal year. As The Next Web noted, "somewhere in a Microsoft conference room earlier this spring, someone looked at a Claude Code invoice and did the arithmetic against a Copilot CLI roadmap, and made a decision."
Microsoft EVP Rajesh Jha framed it diplomatically: "Claude Code was an important part of that learning. At the same time, Copilot CLI has given us something especially important: a product we can help shape directly with GitHub." The tool wasn't cancelled because engineers disliked it. It was cancelled because they used it too much.
This is the same pattern playing out at two of the world's most sophisticated technology organisations simultaneously. It signals the end of the experimental phase — the period where large companies absorbed arbitrary token costs in exchange for learning. The learning is done. The invoices are real.
Why token pricing catches businesses off guard
The core issue isn't that AI tools are expensive. It's that they're expensive in a way that traditional software budgeting doesn't account for.
A seat-based SaaS licence costs the same whether an employee uses it for ten minutes or ten hours a day. Claude Code — and most agentic AI tools — run on token consumption. The invoice is a function of how many tokens the model processes, which scales not with headcount but with ambition. A developer running autocomplete suggestions consumes negligible tokens. A developer running Claude Code as an autonomous agent across a monorepo, refactoring an API layer and generating tests in parallel, can consume thousands of dollars in a single afternoon.
This mirrors the early cloud cost problem. When enterprises first adopted AWS at scale in the early 2010s, bills routinely arrived at triple what finance had modelled. It took a decade of FinOps tooling — reserved instances, tagging frameworks, cost anomaly alerts — to bring cloud spend under control.
The FinOps Foundation's 2026 State of FinOps report confirms the industry knows this is coming: 98% of respondents now manage AI spend, up from 31% just two years ago. AI cost management is the number one skillset teams are seeking to add. But for most organisations, the tooling and governance frameworks haven't caught up to the spending.
What this means if you run a business
You don't need to be running 5,000 engineers to hit this wall. The dynamics scale down.
If you're paying for Claude, ChatGPT, or any token-based AI tool across your team, the same maths applies. A team of five using agentic AI tools aggressively could easily run up a bill that surprises you at the end of the quarter. The difference between a $50/month tool bill and a $2,000/month tool bill is often just how much autonomy you give the AI — and how many people are using it simultaneously.
The practical lessons from Uber's experience:
Set budgets before you set leaderboards. Uber incentivised adoption without capping spend. If you're encouraging your team to use AI tools, make sure someone is watching the meter. Most AI providers offer usage dashboards — use them weekly, not quarterly.
Understand the pricing model. Token-based pricing means your bill is driven by usage patterns, not user counts. A single power user running complex agentic workflows can consume more budget than twenty casual users combined. Know who your heavy users are and what they're doing.
Measure output, not adoption. Uber had 95% adoption and couldn't link it to better products. The metric that matters isn't "how many people are using AI" — it's "what measurable improvement has AI delivered." At Heygentic, we've seen this pattern repeatedly: the businesses getting real value from AI are the ones who defined success criteria before they started spending.
Consider the pricing trajectory. Microsoft's move suggests large buyers will increasingly push AI providers toward flat-rate or committed-spend models. Until that repricing happens across the market, token-based billing remains the default — and the risk sits with the buyer.
What to watch
Uber's CTO has signalled the company will trial OpenAI's Codex as part of a multi-vendor strategy, likely motivated by competitive pricing leverage. Microsoft's forced migration to Copilot CLI will test whether an internally controlled tool can match the capability that drew engineers to Claude Code in the first place.
The broader question is whether Anthropic — and the rest of the inference providers — will face enough enterprise pressure to shift their pricing models. Microsoft has already done this with Copilot, which recently crossed 20 million paid enterprise seats: a flat per-seat model that limits vendor upside but gives enterprise finance teams the predictability they need. As Claude Code usage scales, Anthropic will face increasing pressure from procurement teams to offer similar structures.
For businesses of any size, the takeaway is simple: AI tools deliver real value, but "adopt everything, measure later" is not a strategy. It's how you end up back at the drawing board in April.
Sources
- Uber president says AI spending is getting 'harder to justify' — The Verge
- Uber CTO Shows How Claude Code Can Blow Up AI Budgets — The Information
- Uber Burned Its Entire 2026 AI Budget in Four Months — Startup Fortune
- Microsoft and Uber Are Running Into an Expensive New AI Problem — Techloy
- Microsoft Shifts Engineers from Claude Code to GitHub Copilot CLI — WinBuzzer
- Microsoft's quiet Claude Code retreat and the real cost of enterprise AI — The Next Web
- State of FinOps 2026 Survey — FinOps Foundation / Linux Foundation
