Quick Facts
- Uber exhausted its entire 2026 AI coding budget within four months of deploying Claude Code to roughly 5,000 engineers in December 2025.
- Monthly per-engineer AI tool costs at Uber ran between $500 and $2,000, five to twenty times what most enterprises budget for a developer seat.
- Uber responded by capping context windows at 400,000 tokens, setting reasoning effort to medium by default, and building a 24-million-node knowledge graph to cut prompt bloat.
Uber gave approximately 5,000 engineers access to Claude Code in December 2025. By April 2026, adoption had climbed from 32% to 84% and the company had burned through its full annual AI coding budget. CTO Praveen Neppalli Naga told The Information the company was "back to the drawing board."
During one two-hour hands-on session he ran himself, Neppalli Naga spent $1,200 in tokens. That figure helped trigger a company-wide conversation about what AI coding tools actually cost at scale.
The Cost Problem
Internal leaderboards ranked engineering teams by AI tool usage, which pushed adoption higher and spending with it. AI-related costs at Uber rose roughly six-fold since 2024. The company's R&D expenses hit $3.4 billion in 2025, up 9% year over year.
COO Andrew Macdonald called the CTO's disclosure a "head-exploding moment" during a May 23, 2026 appearance on the Rapid Response podcast. "It's very hard to draw a line between one of those stats and, 'Okay, now we're actually producing 25 percent more useful consumer features,'" Macdonald said. He added that the company would need to weigh token consumption costs against headcount tradeoffs. CEO Dara Khosrowshahi separately said Uber was slowing hiring to offset growing AI spending.
Four Technical Fixes
Uber's engineering team built a wrapper called AIFX, short for AI Foundations, to manage settings across coding tools including Claude Code, Codex, and OpenCode. When an engineer switches between tools, their settings, skills, and plugins carry over. AIFX enforces four cost controls across all of them.
First, context compaction kicks in automatically at 400,000 tokens, even for models capable of handling one million tokens. Because an agent resends its full context to the model on every request, a larger context raises the cost of every step in a session.
Second, reasoning effort defaults to medium rather than high. The deeper analysis setting drives up costs without a guaranteed improvement in output quality for most coding tasks.
Third, Uber stopped pre-loading Model Context Protocol tool schemas at session startup. With more than 100 MCP tools installed, pre-loading added 50,000 to 70,000 tokens of schema overhead to every initial prompt, and that overhead was resent on each context turn. Uber now projects all 1,000-plus MCP tools from its internal gateway as command-line tools, so the model resolves and invokes them dynamically rather than carrying their schemas in the session context.
Fourth, Uber extended its prompt cache time-to-live from five minutes to one hour for interactive sessions. Engineers often leave sessions idle for more than five minutes, which previously invalidated the prefix cache and forced full-price context rebuilds. Extending the TTL means subsequent reads cost 0.1 times the standard input token rate. Sub-agents retain the original five-minute TTL because their tasks are short-lived.
The Context Graph
Beyond the four defaults, Uber built what it calls the AI Context Graph: a structured knowledge network with 24 million nodes, 80 million edges, 86 node types, and 117 edge types covering production services, employees, pull requests, design documents, incidents, and Jira events. Agents receive the graph schema at the start of a session and traverse the rest themselves, pulling relevant documents and source code as needed rather than loading everything into the prompt at once.
The graph grounds Uber's internal Cortana assistant and reduces the size of agent traces. It replaces the brute-force approach of stuffing maximum context into every prompt with a system that retrieves only what an agent actually needs for each step.
What This Means for Other Companies
Uber's experience is a preview of what many engineering organizations will face as agentic coding tools move from pilot to company-wide deployment. The per-token cost structure that made individual developer productivity gains look compelling changes shape fast when thousands of engineers run multi-hour sessions daily. The defaults that ship with most AI coding tools are optimized for capability, not cost. Uber had to build its own layer to change them.
Uber now runs roughly 30,000 skill executions per day across its internal skill marketplace, portable across harnesses through AIFX. The company's answer to the budget problem was not to cut access but to engineer the defaults so that high adoption no longer guarantees runaway spend.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.
