Google has embarked on one of the most ambitious projects in its history — a full-scale migration of its internal infrastructure to the Arm architecture. According to the company, approximately 30,000 software packages have already been adapted, including key services such as YouTube, Gmail, and BigQuery. Ultimately, Google aims to complete the migration across all systems, enabling applications to run seamlessly on both x86 processors and the company’s new in-house Axion chips.
The technical foundations of this initiative are outlined in a preprint titled “Instruction Set Migration at Warehouse Scale” and on Google’s official engineering blog. The publication’s authors — research engineer Parthasarathy Ranganathan and developer relations specialist Wolf Dobson — admit that at the outset, the team anticipated numerous challenges: discrepancies in floating-point computations, variations in multithreading behavior, and complications arising from platform-dependent operations and performance tuning. Yet, as practice has shown, modern compilers and testing frameworks have already resolved many of these issues.
In the early stages, engineers manually ported Google’s largest internal services — F1, Spanner, and Bigtable. The greatest difficulty lay not in architectural incompatibilities but in the sheer labor of maintenance: fixing x86-dependent tests, modernizing outdated build and release systems, resolving deployment errors, and ensuring the stability of mission-critical services. In total, around 30,000 applications have been processed — an immense volume of code that prompted Google to employ its full suite of automation tools and develop a new AI-based assistant known as CogniPort.
As the project’s authors explain, CogniPort analyzes build and testing errors. When a library, binary, or test fails to compile, the agent automatically attempts to correct the issue. In its Blueprint Editing mode, the system generates commits requiring complex, multi-layered modifications beyond simple code edits. Experimental data shows that CogniPort achieves an efficiency rate of around 30%, particularly in fixing test cases, platform-specific conditionals, and data representation errors.
Although the current success rate remains modest, Google continues to scale up the process — with roughly 70,000 additional packages yet to be adapted. The ultimate goal is to achieve complete cross-architecture compatibility, enabling the Borg orchestration system — the internal precursor to Kubernetes — to distribute workloads between servers with maximum efficiency.
The transition carries significant economic implications as well. Google estimates that servers powered by Axion processors deliver up to 65% better performance-per-dollar compared to x86 platforms and approximately 60% higher energy efficiency. These gains promise to reduce infrastructure costs and lower the overall energy consumption of Google’s vast global data center network.
