Uber exhausted its entire 2026 artificial intelligence budget by April, four months into the year. Per Forbes, reporting on disclosures to The Information, Claude Code spread across roughly 5,000 Uber engineers faster than the company’s finance models had anticipated, and CTO Praveen Neppalli Naga confirmed Uber was “back to the drawing board” on its assumptions. This is a company whose total research and development spend reached $3.4 billion in 2025. The budget did not collapse because Uber is small or careless. It collapsed because token-based consumption pricing does not behave like the software line items finance teams know how to model, and the gap between what engineers consume and what the budget assumed is no longer hypothetical.
The reflexive lesson founders draw is “AI tools are too expensive.” That is the wrong lesson. The tools are priced like cloud infrastructure and bought like a SaaS seat. The waste lives in the mismatch.
The Uber numbers, because they are instructive
The detail is what makes Uber a case study rather than a headline. Per Forbes, Uber rolled out Claude Code in December 2025, and adoption climbed from 32 percent of engineers in February to 84 percent classified as agentic-coding users by March. By spring, 95 percent of Uber engineers used AI tools monthly, roughly 70 percent of committed code originated from those tools, and about 11 percent of live backend updates were written by agents with no human in the loop. The spend behind that: monthly cost per engineer averaged $150 to $250, while power users ran between $500 and $2,000.
Read that distribution carefully, because it is the whole story. The average engineer is not the problem. The cost is wildly uneven, concentrated in a tail of power users running multi-file refactors and long agent sessions. A budget built on a seat-license assumption (one price, times headcount, renewed annually) cannot survive a cost driver that varies by 10x across the same team. As one widely-shared analysis of the Uber episode put it, the AI budget “should not sit under software subscriptions anymore. It behaves much more like cloud spend: variable, elastic, and very easy to underestimate when the product actually works.”
The timing made it worse. Per Forbes, on May 13 Anthropic announced that paying Claude subscribers would face a separate monthly credit meter for agent tools and third-party tooling, billed at full API rates starting June 15. The line item that finance thought was a flat subscription was about to meter like a taxi.
Sticker price is not real price
The consumer-grade confusion is the same one Uber hit at scale. As MorphLLM lays out in its breakdown of AI coding costs, every tool has a marketing price and a real price: the marketing price is the pricing page, and the real price is “what you pay after token consumption, API overages, agent loops that burn through context.” Claude Pro is $20 a month, but per MorphLLM the Pro plan’s roughly 45 messages per 5-hour window is “gone in 1-2 hours of focused work” with Claude Code.
The gap between plan pricing and raw consumption is enormous, and it cuts in the buyer’s favor only if you stay on a plan. MorphLLM cites one developer whose eight months of daily Claude Code use consumed 10 billion tokens, which at Sonnet API pricing of $3 and $15 per million tokens would be over $15,000, versus $100 a month on the Max plan: a 93 percent saving. But the same 10 billion tokens reveal what the agent is actually doing, which is processing massive context on every call. Cursor went through the public version of this. Per MorphLLM, in June 2025 Cursor replaced fixed request allotments with usage-based credit pools, one developer reported $350 in overages in a single week, and Cursor issued a public apology and refunds in July 2025.
Where the tokens actually go
The uncomfortable finding is that a large share of the spend buys nothing. MorphLLM cites a DEV Community study finding that 70 percent of coding-agent tokens are waste. A coding agent does not just generate code; it reads files, searches the codebase, runs commands, reads the output, and reasons across all of it, and each of those steps reloads context. When the codebase is tangled, every task drags more context than it should, which is the same maintainability problem that slows human engineers, now metered by the token.
The average engineer is not the problem. The cost is concentrated in a tail of power users.
Who burns the most relative to their output is the part founders should sit with. In its 2026 survey of more than 900 engineers, The Pragmatic Engineer found, per the public summary of Part 2, that for less experienced engineers “AI seems to be less helpful and they rack up higher AI token bills,” and suggests they may need more mentorship and space to learn rather than more tokens. The same write-up names an “AI tooling addiction,” where using agents “feels like a slot machine” encouraging “just one more prompt,” and reports that codebase quality “seems to be decreasing, but management at most places does not care.” Higher spend, by the people getting the least from it, producing code that quietly degrades the asset. That is waste in three directions at once.
The trend is already correcting
The market is reacting, and the reaction tells you where the discipline lives. Gergely Orosz’s The Pulse, in a May 28 issue titled a trend of trying to cut back on AI spend, reports “top-down and bottom-up efforts to rationalize AI token spend,” with spending on AI agents “being dampened via per-engineer monthly” budgets at mid-sized and large companies, alongside a related warning about consumption pricing risk: the same issue notes Google Cloud suspended a $2-million-a-month customer without warning. The per-engineer monthly budget is the obvious first control, and companies are reaching for it because the seat-license model already failed them.
A framework for not wasting it
The practical answer is not to ban the tools. Uber’s own numbers show the value is real: 70 percent of committed code from agents is not nothing. The answer is to treat AI spend as the variable, elastic cost it is, and put controls where the variability lives.
Budget per engineer per month, not per seat. The Pulse’s reporting shows this is already the emerging norm. A monthly token budget per engineer turns an unbounded API meter back into a plan, and surfaces the power-user tail before it surfaces in a budget review.
Separate high-value tasks from churn. The Pragmatic Engineer’s survey distinction between engineers who get real output from AI and those who rack up bills without it is the one to operationalize. A multi-file refactor, a gnarly migration, a test-suite backfill: high value, worth the tokens. Re-prompting an agent five times because you would not read the diff: churn, and the slot-machine loop. Knowing which is which is engineering judgment, and it does not come from the tool.
Cut waste at the mechanism. MorphLLM reports that routing and compaction can deliver a 40 to 70 percent cost reduction, because most of the spend is context being reprocessed. Keeping the codebase clean is a cost lever now, not only a maintainability one: a tidy repo means the agent loads less context per task.
Build cost sensors into the workflow, the same way you would build cost alerts on cloud spend. The tooling supports more control than its docs advertise. As André Figueira documents in his read of the Claude Code source at version 2.1.87, settings live in .claude/settings.json and can be committed and shared with a team, and hooks can intervene on what the agent does. The point is not the specific field. The point is that a founder can encode usage policy in the workflow rather than hoping each engineer self-rations.
The honest read of the Uber episode is that the tool worked, adoption exploded because it worked, and the budget broke because nobody had reclassified the line item. A $3.4-billion R&D org ran out of AI money in four months by treating a taxi meter like a subscription. A smaller team does not get four months of runway before that mistake shows up in the numbers.