# Uber’s Zero Growth Stack: Scaling Services, While Optimising Infrastructure and AI Cost > Source: > Published: 2026-07-28 07:07:00+00:00 Uber's "Zero Growth Stack" offers a scalable infrastructure solution that decouples capacity growth from business demand, allowing the company to reduce its physical hardware footprint while maintaining service scaling significantly. By prioritising automated runtime optimisation and rigorous AI lifecycle management, this strategic shift mitigates the overhead of traditional resource scaling. Central to this approach is a systematic transition toward dynamic system control, ensuring that infrastructure performance remains efficient even as internal software complexity increases. A central pillar of this initiative is the systematic optimisation of Go runtime performance. Recognising that static garbage collection (GC) tuning often failed to address services with highly varied memory footprints—ranging from 100MB to 1GB—Uber engineered a dynamic solution: [GOGCTunner](https://github.com/uber-go/gogctunner). This library automates the adjustment of the GOGC value, which dictates the frequency of GC cycles relative to heap growth. By integrating directly with cgroup memory limits and monitoring live object utilisation, the tool enables dynamic runtime adjustments. While this automation successfully reclaimed 70,000 CPU cores across 30 mission-critical services, it replaces static settings with a control loop that must balance heap size (maintained at 1.25x live objects) against strict 70% memory utilisation thresholds to avoid OOM events. Beyond infrastructure, Uber has integrated generative AI into its software development lifecycle to accelerate developer velocity. The architecture layers include a Platform Layer (utilising [Michelangelo AI](https://www.uber.com/blog/michelangelo-machine-learning-platform/) as a model gateway), a Context Layer (employing an [MCP Gateway](https://modelcontextprotocol.io/introduction) to inject internal source code), a Specialised Layer (deploying background agents like [Minion](https://www.infoq.com/news/2026/03/stripe-autonomous-coding-agents/) and Shepherd), and a Review Layer (using tools like uReview and Code Inbox). While 92% of engineers utilise these agents monthly—with 31% of new code being AI-authored and Autocover generating over 5,000 unit tests per month—the rapid adoption introduced significant economic friction. AI-related costs increased sixfold since 2024, as token-based billing models led to budget exhaustion and individual developer costs reached $2,000 monthly by early 2026. To address these challenges, Uber transitioned from unconstrained adoption to strict cost governance, implementing usage caps limited to $1,500 per developer. The company is now auditing the efficacy of AI-generated code, seeking to quantify the direct impact of automated throughput versus infrastructure overhead. To evolve these metrics, teams are shifting from aggregate cost tracking to granular indicators. Rather than relying solely on revenue per unit of labour and AI compute spend, the company suggests evaluating the "net code quality ratio", comparing the frequency of post-release hotfixes in AI-authored versus human-authored code. Establishing a "compute efficiency per feature" metric allows engineering leads to determine whether the overhead of AI-generated code—particularly when factoring in redundant tests generated by tools like [Autocover](https://github.com/SitoCH/autocover)—aligns with long-term infrastructure sustainability goals, ensuring that infrastructure optimisation remains synchronised with actual feature delivery.