Technology
Uber Decouples Growth from Infrastructure: New Runtime Controls Cut AI Costs by Capping Developer Spend
Uber's "Zero Growth Stack" uses automated runtime tuning and AI cost governance to scale services without proportional hardware increases, reclaiming 70,000 CPU cores while limiting developer AI spending to $1,500 monthly.
InfoQ
Uber has unveiled an infrastructure strategy that breaks the traditional link between business growth and computational resource expansion. Termed the "Zero Growth Stack," the approach combines automated runtime optimization with strict AI cost controls to maintain service scaling while shrinking the company's physical hardware footprint.
Automating Runtime Efficiency with Dynamic Garbage Collection#
At the technical heart of this initiative sits GOGCTunner, an open-source Go runtime library that replaces manual garbage collection (GC) tuning with dynamic adjustments. Traditionally, engineers would set static GC parameters for each service—a rigid approach that fails when applications have widely varying memory demands, from 100MB to 1GB per instance.
GOGCTunner monitors live object usage and integrates with container memory limits to continuously recalibrate the GOGC value, which controls how frequently garbage collection runs relative to heap growth. The library maintains heap size at 1.25x actual live objects while enforcing strict 70% memory utilization thresholds to prevent out-of-memory crashes. This automated control loop has reclaimed 70,000 CPU cores across 30 mission-critical services—a substantial efficiency gain that eliminates the need for proportional hardware expansion.
The Double-Edged Sword of AI-Assisted Development#
Uber has deeply embedded generative AI into its software development workflow through a layered architecture: a Platform Layer built on Michelangelo AI as a model gateway, a Context Layer using an MCP Gateway to feed internal code to models, a Specialised Layer deploying autonomous agents like Minion and Shepherd, and a Review Layer with tools such as uReview and Code Inbox.
The adoption has been rapid: 92% of engineers now use these AI agents monthly, and AI now authors 31% of new code at the company. Autocover alone generates over 5,000 unit tests monthly. Yet rapid growth brought runaway costs—AI-related expenses surged sixfold since 2024, with token-based billing driving monthly developer costs to $2,000 by early 2026.
