At Google, the future is multiarch; AI and automation are helping us get there (cloud.google.com)

Google announced progress toward a multi-architecture cloud by deploying its first Arm-based Axion CPUs and running key services (YouTube, Gmail, BigQuery) on Arm and x86 simultaneously. In a paper, the team analyzed 38,156 commits (≈700K changed lines) from their monorepo to document how automation and AI enabled migration of more than 30,000 applications, with Arm hardware now fully subscribed. The practical payoff: up to 65% better price-performance and up to 60% improved energy efficiency for customers, plus reduced operational friction for Google’s fleet as new apps are made multiarch by default. Technically, Google found most migration effort was “boring” work—tests, build/release configs, and rollouts—rather than deep ISA semantics, thanks to modern compilers and sanitizers. They leveraged existing automation (Rosie for mass commits, sanitizers/fuzzers, CHAMP for rollout monitoring) and built an LLM-driven agent, CogniPort, to fix build/test failures. CogniPort uses three nested agent loops (an orchestrator plus build-fixer and test-fixer agents) that iteratively call an LLM and tooling to modify code until builds/tests pass. On a 245-commit benchmark it auto-fixed failing tests ~30% of the time, particularly for platform conditionals and data-representation issues. The work shows LLM-based automation can scale ISA migration across massive codebases, lowering the barrier for heterogeneous deployment and offering a blueprint for other large-scale multiarch transitions in AI/ML infrastructure.