OpenAI introduces Pando: A Git backend built for AI-driven coding

OpenAI has introduced a custom Git version control backend called Pando, designed to handle the heavy demands of widespread AI-assisted coding. The new infrastructure maintains full compatibility with standard Git tools, processes about a petabyte of data daily, and reduces push latency by nearly five times, solving reliability issues tied to traditional servers.

• Pando remains fully compatible with standard Git clients and protocols.
• The system processes roughly a petabyte of Git data daily and significantly accelerates data retrieval.
• Development took 10 weeks, with architecture designed by a single engineer and implementation executed by Codex agents in Rust.
• Preview access rolled out to corporate partners in November 2026, with commercial availability planned for Q1 2027.

The Scaling Challenge of AI Development The need for Pando arose as OpenAI experienced a massive surge in version control traffic driven by automated coding tools like Codex. According to the company, quarterly commits to its monorepository were projected to increase by over 2,400% by year-end. Traditional Git servers were engineered for human developers generating dozens of changes daily, whereas continuous AI usage escalated operations to thousands of commits and merges per hour. Because of this extreme load, the reliability of the legacy infrastructure dropped below 90%, causing slow clone speeds, merge errors, and strict repository size caps. Rather than abandoning the Git ecosystem or protocol, engineers concluded that the bottleneck lay in traditional server architecture, inspiring the creation of Pando after just 10 weeks of development.

Architectural Innovation and Performance In designing Pando, engineers discarded two classical Git server requirements: maintaining a globally consistent state across all branches simultaneously and generating a custom packfile for every client request. Pando processes writes to each branch independently, meaning pushing changes to one branch doesn't require waiting on unrelated branches. Incoming packfiles are stored directly as BLOBs without unpacking, and branch tips are tracked in a dedicated metadata store, reducing CPU and disk I/O overhead. To fetch code, Pando utilizes the packfile-uri protocol mechanism, while the Express caching layer achieves a cache hit rate above 99%, allowing clone speeds to saturate standard 10 Gbps network connections. Background compaction processes merge smaller packs over time without deleting data, avoiding heavy garbage collection routines. Developers continue using standard Git commands, while OpenAI has also contributed several client-side fixes back to the upstream Git project.

Commercial Outlook and Monitoring Pando is not being distributed as open-source software. Following its official announcement on November 2026, the system became available in a restricted preview for select enterprise partners. Commercial access is slated to expand in the first quarter of 2027, featuring a pricing model tied to repository volume and operation counts. The commercial tier will also include specialized monitoring tools designed to track metrics such as the proportion of AI-generated code, merge conflict frequencies, and overall repository growth velocity, helping large organizations manage automated development pipelines.