Uber's CTO let one line slip to The Information. "The budget I thought I'd need has been blown away already."
January through April 2026 — four months — and a full year's AI budget was gone at a company that spends $3.5 billion on R&D. The culprit: Claude Code. But this isn't a story about AI being expensive — it's a story about how incentive structures create token runaway. And every other company is heading down exactly the same path.
What Is It?
Uber CTO Praveen Neppalli Naga officially confirmed to The Information that the company's entire 2026 AI budget was already exhausted as of April. Key operational numbers came out with it —
- 95% of Uber engineers use AI coding tools every month, with 84% classified as "agentic coding users"
- 70% of committed code is AI-generated, and 11% of live backend code changes are written autonomously by AI agents
- 1,800 code changes per week coming from AI
- AI costs up roughly 6x vs. 2024, with 2025 R&D spend at $3.4B (YoY +9%)
- Cursor adoption plateaued — Claude Code has become the de facto internal standard
Here's the thing — R&D spend only grew 9% over that same period. If 70% of your code comes from AI but output only grew 9%, where did all the other tokens go? Om Bharatiya's take: some shipped, some was exploration, and some was "code that didn't need to be written — but once friction dropped to zero, engineers wrote it anyway."
Why Governance Is the Real Problem
The single most damaging decision Uber made was building an internal leaderboard — ranking engineers by AI usage volume.
Here's why that's a problem — it's like evaluating your sales team on "number of calls made" instead of "deals closed." When you optimize for consumption, consumption explodes, not value. Around the same time, Meta built an internal dashboard called "Claudeonomics" to track token usage across 85,000 employees. When it surfaced that one person had used 281B tokens (~$1.4M) in 30 days, they killed the dashboard in two days. Not because the data was wrong. Because it was too embarrassing.
| Category | Uber (consumption-optimized) | Realistic Governance |
|---|---|---|
| Budget model | Single company-wide shared pool | Per-team, per-use-case allocation |
| Model routing | Always the most capable model | Route by task (code review = Sonnet, boilerplate = Haiku) |
| Metrics | "% of engineers using AI" | PRs per sprint, AI code bug rate, deploy time |
| Incentives | Token usage leaderboard | Drop the leaderboard + evaluate on outcomes |
| Cost levers | None | Prompt caching (−90%) + Batch API (−50%) |
As of April, Uber said it's shifting to an "agent engineers" model — where AI handles coding, testing, and deployment end-to-end, with humans acting only as orchestrators. They're also evaluating OpenAI Codex. But the bigger issue is elsewhere —
Getting Started
- Kill the leaderboard immediately
Stop measuring "% of engineers using AI tools." It's a vanity metric — the moment you track it, it gets gamed. - Per-team token budgets + alerts
Separate pools by team, not a single company-wide bucket. Auto-alerts at 50%/75%/90% thresholds. Same playbook as cloud cost allocation. - Add a model routing layer
Code review: Sonnet. Boilerplate: Haiku. Architecture reasoning: Opus only. 5–15x cost reduction, with nearly identical quality on simpler tasks. - Enable prompt caching and Batch API
Anthropic offers prompt caching at −90% and Batch API at −50%. Enterprise workloads have a lot of repeated context — this alone can cut costs 30–50%. - Replace your metrics
Swap "% using AI" for "PRs merged per sprint," "bug rate in AI-generated code," and "production deploy time." Grade on business outcomes.
FAQ
(Rendered automatically by FAQSection — not included in content)
Deep Dive Resources
Project Flux — Blown by April Operational data breakdown from The Information's reporting — 84% adoption rate, 11% of live code auto-written, and more projectflux.ai
Om Bharatiya — Uber Blew Its AI Budget in One Quarter Infrastructure and governance perspective — leaderboard pitfalls, model routing, and caching strategies ombharatiya.com
Yahoo Finance — Uber's Anthropic AI Push Hits a Wall CTO Naga quotes and internal tool comparison (Claude Code vs. Cursor) finance.yahoo.com
Hacker News Discussion Thread What engineers think the real cause is — including estimates like "$1,250 mid-point spend per engineer" news.ycombinator.com
