GitHub unveils AI-native storage system for high-concurrency code collaboration

On November 7, GitHub’s official channel released an internal evaluation report detailing the performance of its new Git storage architecture, capable of handling up to 5,000 AI agents concurrently working on a single repository, with 7,200 push operations per second and consistently low-latency read handling.

Decoupling storage from computation to enable scalable AI workflows

The architecture, codenamed Project Rubicon, was initially designed in May 2026 and features a fundamental shift: separating data storage from computational logic. Git objects are now stored in Azure Blob Storage, while reference states are managed by a dedicated Ref-CAS service that uses compare-and-swap mechanisms for consistency. Compression tasks are offloaded to independent background nodes, enhancing overall efficiency.

Public debut at GitHub Universe, limited private preview underway

Authored by Karthik Raman and Brian Celenza, the technical overview was first presented at GitHub Universe. The system is currently undergoing a restricted private preview phase and is scheduled for broader rollout in early 2027.