The Numbers
Uber gave Claude Code access to roughly 5,000 engineers at the start of 2026 without hard usage caps. By March, 84% of engineers were classified as agentic coding users — up from 32% in February, a jump that happened in a single month. Monthly API costs per engineer reached $500–$2,000 depending on usage intensity. The cumulative result: Uber exhausted its full 2026 AI tools budget by April. The company had planned for twelve months of spending and ran out in four. Uber's COO confirmed the overage and, more significantly, questioned publicly whether the token consumption translated into a proportional increase in consumer-facing features.
Why Agentic Workloads Break Traditional IT Budgets
Agentic coding tools burn more compute than single-turn AI assistance by design. An agent that iterates on a test failure, rewrites the offending function, reruns the tests, and repeats the loop can consume 10–50x the tokens of a single autocomplete interaction. When 84% of an engineering org adopts that pattern simultaneously and usage is uncapped, the bill compounds faster than any prior software rollout finance teams have seen.
The specific shape of Uber's exposure is instructive: agentic workloads are bursty, not steady. An engineer who rarely uses AI tools can spike their consumption 20x during an intensive debugging session. Traditional software procurement — seat licenses, flat monthly SaaS fees — doesn't model this. Enterprises are buying tokens but budgeting in seats. Those two mental models are incompatible at scale.
The COO's Question Is the Right One
Uber's COO did not say Claude Code was worthless. The statement was specific: higher token consumption was not translating into a proportional increase in useful consumer-facing features. That's a meaningful distinction — it suggests a measurement problem, not a tool problem. If 70% of committed code is AI-assisted but consumer-facing velocity didn't accelerate commensurately, the bottleneck is probably in the stages that surround code generation: product decisions, review cycles, deployment, and feedback loops. Token spend scales with model usage; feature velocity depends on the entire product cycle. Agentic coding tools speed up one node in that graph.
The Budget Crisis Reveals an Infrastructure Gap
Uber's situation exposes something now common across enterprises deploying agentic AI: there's no native cost attribution layer between model APIs and team-level budget ownership. Cloud billing dashboards exist. What doesn't exist — and what no major platform does well yet — is a view that maps token spend to business output. Which projects consumed how much compute? Which agentic sessions correlated with shipped features versus abandoned experiments? Without that attribution, finance teams can't measure ROI and engineering teams can't self-govern usage. The spend accumulates until someone reads the April invoice.
What Builders Should Take From This
For teams building agentic coding tools, infrastructure platforms, or enterprise AI products: Uber's situation is the case study your prospective buyers are already reading. Enterprise procurement will now ask for cost controls, usage caps, team-level attribution, and ROI metrics as part of the initial purchase decision. Tools that embed these features natively — not as admin-panel afterthoughts — will clear procurement faster. The agentic AI adoption curve is real; Uber's 84% engineering adoption in a single month proves demand is not the variable. The budgeting infrastructure to govern that adoption at scale is the gap the next layer of the market is about to fill.