Uber burned its entire 2026 AI coding budget in 4 months

In April 2026, Uber's CTO told colleagues the company had exhausted its 2026 AI tooling budget. The cause wasn't waste — it was an adoption curve nobody was watching at the right resolution.

What happened

Per AI Magazine and The Information's Applied AI newsletter, Uber CTO Praveen Neppalli Naga told the engineering org that the year's AI tooling budget was already gone — four months in. Uber's engineers had moved aggressively onto Anthropic's Claude Code (alongside Cursor), and adoption climbed from roughly a third of the engineering org to ~84% inside the same period. Roughly 70% of committed code is now AI-generated, and per-engineer monthly API spend reportedly ranged from $500 to $2,000.

For context: Uber's 2025 R&D spend was roughly $3.4B, up 9% year over year. The CTO described the team as "back to the drawing board" on AI budgeting and is now testing OpenAI Codex to diversify provider risk.

Why a budget number alone wasn't enough

Most enterprises set an annual AI budget the same way they set their cloud budget in 2014 — a single number, reviewed quarterly. That works when consumption changes 5–10% a quarter. It fails when:

• Per-call costs vary 10x across models (Haiku vs. Opus, GPT-4.1-mini vs. GPT-4.1).
• A new "agentic" feature can run a tool-use loop dozens of times per user request.
• Adoption can double inside a single sprint when the tool genuinely works.

The single-number budget doesn't tell you which team is responsible for the spike, which model is the cost driver, or whether you can survive another month at this rate. By the time the quarterly view is in, you're "back to the drawing board."

What would have changed it

1. Per-team monthly budgets with email alerts at 80% and 100%. The platform team and the rider growth team should have separate budgets, separate alerts, separate accountability.
2. Per-model spend visibility. When the daily burn jumps, the dashboard should answer "is this Claude Opus calls, Haiku calls, or Cursor's daemon" without an engineer pulling SQL.
3. Per-engineer caps as a safety net. Not as a productivity ceiling — as a "this single account is doing something weird" sentinel.
4. Provider-side metadata tagging. Pass X-Inventoria-Team headers through your AI gateway so usage rolls up to the team automatically. Anthropic, OpenAI, and Vertex all support tagging.
5. A 6-hour polling cadence on the Admin APIs — not a monthly export. The fastest way to catch a 16x curve is to catch it on day three, not day ninety.

The honest read

Uber didn't fail at AI strategy. They succeeded at it — 84% adoption is a number their competitors will spend two years chasing. They failed at AI cost telemetry, which is a different discipline and was, until recently, no one's job.

That's the gap Inventoria's AI Spend module closes: per-team and per-model token-and-dollar visibility across OpenAI, Anthropic, Azure OpenAI, Vertex/Gemini, and GitHub Copilot. Six-hour polling, monthly budgets, automatic alerts. It's the boring plumbing that makes a $3.4B R&D org's AI rollout sustainable instead of a "back to the drawing board" memo.