Uber’s AI Budget Lasted Four Months. That’s a Problem.
Uber rolled out Anthropic’s Claude Code to its 5,000 engineers in January 2026. By April or May, the company’s entire annual AI budget was gone. Now Uber president Andrew Macdonald is publicly saying what every enterprise CTO has been thinking: the money doesn’t add up.
“That link is not there yet, right?” Macdonald said on the Rapid Response podcast. “I think maybe implicitly there is more that is getting shipped, but 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.’”
🔍 THE BOTTOM LINE
The first major tech executive to publicly question AI coding ROI just admitted the connection between token consumption and product value doesn’t exist — and it’s the company that gave Claude Code to every engineer.
Token Tsunami
The numbers tell an uncomfortable story. Uber spent $3.4 billion on research and development in 2025 — 9 percent more than the previous year. Then in 2026, it gave every engineer access to Claude Code and watched token consumption explode.
When individual developers use AI coding assistants, they generate enormous volumes of tokens daily — writing code, debugging, testing, iterating. Under the new variable consumption pricing models that AI providers have shifted to, those daily interactions compound fast.
Uber CTO Praveen Naga confirmed the company completely exhausted its allocated 2026 AI budget within just four to five months. The original subsidy era is over: AI providers spent 2024 and early 2025 heavily subsidising inference costs, sometimes spending thousands to support a single $20/month user. Now those true costs are landing on enterprise balance sheets.
What is token-based pricing? AI companies like Anthropic charge per “token” — roughly a word or word-fragment — for every interaction with their models. When 5,000 engineers are chatting with Claude Code all day, those tokens add up fast. Under traditional software, you’d pay a flat fee. Under token pricing, costs scale with usage, and there’s no ceiling.
The ROI Problem
Macdonald’s candour is striking because he’s not a sceptic — he’s the president of a company that went all-in on AI. Uber didn’t dip a toe in; it rolled Claude Code out company-wide. And still, he can’t draw the line from spend to value.
“We’re going to have to start talking about token consumption and the associated cost versus headcount,” Macdonald said. “So if you’re not actually able to draw a direct line to how much useful features and functionality you’re shipping to your users, that trade becomes harder to justify.”
Uber CEO Dara Khosrowshahi framed it differently: the company is compensating for rising AI costs by hiring fewer humans. But that’s not the same as proving AI creates value — it’s just saying the cost is offset by headcount reduction, which is a very different proposition.
The Enterprise AI Bubble?
Uber isn’t alone. Microsoft is reportedly dealing with similar internal budget blowouts. As providers shift from subsidised pricing to real token costs, the math gets brutal: individual token prices are dropping, but total volume is tripling, leaving overall enterprise spend higher, not lower.
The pattern is becoming familiar:
- Subsidy phase (2024-early 2025): AI providers lose money on every user to drive adoption
- Reality phase (mid-2025 onwards): Providers demand real revenue; enterprise budgets blow out
- Questioning phase (now): Executives ask whether the spend translates to anything customers notice
Nolan Lawson’s viral essay “Using AI to Write Better Code More Slowly” hit 945 points on Hacker News this week — arguing that AI’s real value is in code quality, not velocity. But that argument doesn’t help a CFO staring at a $3.4 billion R&D line item with nothing measurable to show for the AI portion of it.
What This Means
If Uber — a company that went all-in on Claude Code and replaced human hiring with AI spend — can’t justify the ROI, the signal is clear: the enterprise AI spending narrative is cracking. This isn’t a company that tried AI and failed. This is a company that succeeded at adoption and still can’t prove it mattered.
The next six months will tell. Either AI coding tools start producing measurable feature velocity, or 2026 becomes the year the enterprise AI bubble found its ceiling.
❓ Frequently Asked Questions
Q: Does this mean AI coding tools don’t work? Not at all. The issue isn’t whether Claude Code works — engineers are using it heavily. The issue is whether the cost scales proportionally to value delivered. Uber’s president is saying that link doesn’t exist yet.
Q: What does this mean for NZ businesses? NZ companies watching this should be cautious about enterprise AI contracts with token-based pricing. The “try it and see” approach that worked during the subsidy era doesn’t work when real costs land. Start small, measure rigorously, and negotiate volume caps.
Q: Will AI providers lower prices? They already are, on a per-token basis. But total consumption is growing faster than per-token prices are dropping. The fundamental problem isn’t price — it’s that nobody can measure what they’re getting for the money.
🔍 THE BOTTOM LINE
When the company that gave AI to every engineer can’t prove it’s worth the spend, the question isn’t whether AI works. It’s whether anyone can afford to keep pretending the ROI is obvious.
SOURCES
- The Verge
- Business Insider
- Rapid Response Podcast
- Blazetrends
