After 92% of Engineers Use AI, Uber Begins to 'Cap' AI Usage
Uber's 'Zero Growth Stack' initiative decouples capacity growth from business demand by optimizing infrastructure and managing AI lifecycle costs. After 92% of engineers adopted AI tools, AI-related costs skyrocketed, prompting Uber to impose strict per-developer usage limits and develop metrics like 'net code quality ratio' to evaluate AI-generated code efficiency.
Uber's 'Zero Growth Stack' initiative decouples capacity growth from business demand through automated runtime optimization and strict AI lifecycle management. The company developed GOGCTunner, a library that dynamically adjusts GOGC values based on cgroup memory limits and real-time object usage, recovering 70,000 CPU cores across 30 critical services while balancing heap size and memory thresholds to avoid OOM. Uber integrated generative AI into its software development lifecycle, with 92% of engineers using AI agents monthly, 31% of new code written by AI, and Autocover generating over 5,000 unit tests monthly. However, AI-related costs grew six-fold since 2024, reaching $2,000 per developer per month by early 2026. To address this, Uber shifted from unlimited adoption to strict cost governance, capping usage at $1,500 per developer. The company recommends measuring 'net code quality ratio' and 'compute efficiency per feature' to evaluate the trade-offs of AI-generated code, particularly considering the overhead from redundant tests.
- Сокращения
- CPU = Central Processing Unit — центральный процессор
- GC = Garbage Collection — сборка мусора
- OOM = Out of Memory — нехватка памяти
- GOGC = Go Garbage Collection — параметр сборки мусора в Go
- MCP = Model Context Protocol — протокол контекста модели
Source: InfoQ 中国 —
original