Google porting all internal workloads to Arm, with help from GenAI
Summary
Google has ported roughly 30,000 production packages to the Arm architecture and intends to convert its entire internal workload fleet so services can run on both its homegrown Axion Arm CPUs and traditional x86 processors. YouTube, Gmail and BigQuery already operate on both architectures. The company documented its process in a preprint paper and a blog post, explaining that much of the migration effort involved fixing tests and build/release tooling rather than tackling deep architectural surprises.
To scale the operation, Google extended its automation and created an AI agent called CogniPort. CogniPort analyses build and test errors and can generate migration commits; it succeeded automatically about 30% of the time on tasks like test fixes, platform-specific conditionals and data representation issues. Google still has around 70,000 packages queued for conversion. The goal is to enable Borg (the basis of Kubernetes) to allocate workloads across architectures for improved price-performance and energy efficiency: Google claims Axion delivers up to 65% better price-performance and around 60% better energy efficiency versus x86.
Key Points
- About 30,000 production packages have already been ported to Arm; ~70,000 remain in the queue.
- YouTube, Gmail and BigQuery already run on both x86 and Google’s Axion Arm CPUs.
- Most migration effort focused on brittle tests, complex build/release systems and rollout/configuration issues rather than fundamental ISA bugs.
- Google built an AI tool, CogniPort, to automatically fix build and test failures; it had ~30% success under certain conditions.
- The aim is multi-architecture scheduling via Borg so workloads can move between x86 and Arm for cost and energy gains.
- Google claims Axion machines offer up to 65% better price-performance and around 60% improved energy efficiency vs x86.
Context and relevance
This is a major signal in the industry: a hyperscaler moving large parts of its fleet to Arm at warehouse scale changes economics for cloud infrastructure, chip vendors and enterprise architecture planning. It showcases practical lessons for large migrations — how tests, CI/CD and release tooling become the real blockers — and demonstrates a real-world application of GenAI to automate code-porting at scale. If you run cloud infrastructure, design compilers, or build CI systems, the techniques and trade-offs here matter.
Author takeaway
Punchy and to the point: Google’s exercise is both a cost- and energy-driven play and a blueprint for ISA migration. The surprisingly low rate of deep architectural surprises suggests modern toolchains are robust, while CogniPort shows GenAI can meaningfully speed repetitive engineering work — but it’s not a silver bullet.
Why should I read this?
Want to know how a company the size of Google actually moves tens of thousands of apps between ISAs? This is it — the playbook, the gotchas (mostly tests and tooling), and the GenAI trick they used. If you care about cloud costs, Arm adoption or how GenAI helps real engineers, give it a skim — it’ll save you time and give you a handy map of what to expect.