Leadership1 publisher3 min readPublished
Uber exhausted a year of Claude Code budget in four months
Christian Stegh argues that metered coding tools have made AI a variable cost. The funding data in his column puts that meter inside IT's budget at 62% of organizations, with 31% carrying no AI budget line at all.
The Board Room · Leadership desk
What happened
- Uber's CTO said the company's annual Claude Code budget was gone in four months, leaving him back to the drawing board because the budget he thought he would need was already blown away.
- Gartner's poll puts 2026 IT budget growth at 14% overall, an allocation that rising RAM costs are already cutting into this year.
- Stegh's proposal has IT funding a baseline of enterprise tools such as M365 Copilot and ChatGPT Enterprise, with metered add-ons charged back to the business unit that uses them.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- decision Splitting the bill decides who does the rationing: under this model the business unit funds the meter and sets its own limits, while the CIO still answers for governance of what that spending pulls into the estate.
- cost Memory prices and open-ended token consumption are now competing inside the same IT line, so one of them gets deferred, and the deferral stays invisible until an invoice lands mid-year.
- constraint Hard caps are the main control these tools offer, so cost discipline and adoption pull against each other: a cap tight enough to hold the budget is also a cap that pushes work toward tools nobody is monitoring.
- contradiction IBM's 2025 reading of 18% funding from net-new spend sits oddly beside a mid-2026 figure of 6% funding centrally, and with no pollster named for the later number, the direction of travel rests on a single snapshot.
An annual budget spent in four months means consumption ran at roughly three times the budgeted pace [3] [1]. The response Uber reportedly reached for was a ceiling rather than a better forecast: $1,500 per user per month, mainly for coding assistants [4], which if fully used comes to $18,000 per user per year [2]. A cap is the one variable a finance function controls directly, and setting one converts a spending question into a rationing question.
The meter is sitting in the least elastic place available to it. In the mid-2026 poll Christian Stegh cites, 62% of organizations pay for GenAI out of IT's budget, 6% draw on centralized corporate budgets, and 31% have no formal GenAI budget at all [8]; those figures sum to 99%, so they cover almost the whole sample, and central funding is the rarest route of the three [3]. Against that, KPMG's 2025 CEO survey had 69% of chief executives planning to put 10% to 20% of overall budget into AI [6]. Gartner has IT budgets up 14% for 2026, with RAM costs already cutting into the allocation [9]. A line growing 14% that must absorb both memory prices and unbounded token consumption will settle the conflict by deferring something else.
Both directions on the cap dial cost something, and the column says so. Set it low and, in the words of Tim Bachta quoted by Stegh, "If you're not bringing AI to your staff, they're bringing it themselves" [10], which Stegh expects to arrive on CISOs' desks as a governance, risk and compliance problem [16]. Set it high and you get the four-month outcome. His proposed split is that IT funds baseline enterprise tools such as M365 Copilot, ChatGPT Enterprise and Claude Enterprise, and everything metered on top is charged back to the business unit using it [11]. What the split does not divide is the obligation: IT still funds, secures, trains, monitors and governs the base [12], while the business unit finances its own tokens and sets its own limits [13].
A skeptic would say that one CTO's blown budget is an anecdote, and that $18,000 a year is cheap next to a senior engineer. Both points hold, and the record underneath the argument is thinner than the argument: the polls describe where the money comes from, not how many budgets overran or by how much, and the mid-2026 figures arrive without a named pollster in the column [8]. Stegh concedes there are no best practices yet [15]. The material offers no forecast for token prices in either direction, which is why the answerable question this quarter is the ceiling and its owner rather than the price.
So the decision available before the fiscal year closes is smaller than the problem it addresses: a per-user limit, of the kind most pay-as-you-go services already support at both user and organization level [5], and a named signatory for the overage. An organization that sets the first without the second will hold this conversation again next quarter with less room in the line item and a harder case for cutting, since Verdantas, the example Stegh offers, reports cycle time falling from between 10 and 14 days to hours [14].
What to watch
- Whether the mid-2026 funding poll gets a named source or a second reading, which would turn one snapshot into a trend line.
- Whether metered vendors add organization-level forecasting and alerting rather than only hard caps at user and tenant level.
- Whether more named companies disclose the size of an AI budget overrun, which is the evidence this argument currently lacks.