Home / Technology / Uber's AI Strategy: Defaults Control Soaring AI Expenses
Uber's AI Strategy: Defaults Control Soaring AI Expenses
20 Nov
Summary
- Uber exhausted its 2026 AI budget early due to surging costs.
- Default settings limit AI model context to 400,000 tokens.
- Cost-saving measures reduced AI spending per request by 34%.
Uber has implemented stringent default settings across its AI tools to manage escalating costs, having reportedly exhausted its 2026 AI budget by early this year. These measures aim to curb unexpected expenses that have also impacted other organizations, with 38% delaying AI projects due to higher-than-expected costs.
Key among Uber's strategies is setting automatic compaction at 400,000 tokens for its AI models, even those with larger context windows. Reasoning effort is defaulted to medium, not high, to reduce processing. Prompt caches are limited to one hour for main agents and five minutes for sub-agents. Tool definitions are not loaded at startup to save context tokens.
These optimizations have yielded significant savings, reducing cost per 1,000 requests by nearly 34% and cost per session by 52% from their peaks. Uber is also focusing on making agent workloads model-agnostic, increasing flexibility. The company is giving agents a "mental map" of its internal systems, a context graph with nearly 200 million nodes, to retrieve information as needed rather than pre-loading vast amounts of data.
Uber measures the success of its AI initiatives by maximizing engineer throughput and maintaining quality guardrails, tracking metrics like design documents, pull requests, and deployments. The company also monitors incidents per code change and code reversion rates, gradually building trust in AI-assisted fixes through automated processes.