The Pulse
OpenAI Says Pando Served 57TB of Git Data in an Hour
OpenAI says its Pando Git backend served more than 57 terabytes of Git packs in one hour and nearly a petabyte of data per day. The system now powers development in OpenAI’s monorepo, where it supports concurrent work by engineers and Codex
AI.info Team · Written with AI · fact-checked ·
Pando moved 57TB of Git packs in one hour
OpenAI says its Pando backend served more than 57 terabytes of Git packs in a single hour, with more than 99% of those bytes coming from local disk cache. The system delivers nearly a petabyte of Git data per day and now handles all development in the company’s monorepo, according to an engineering post published November 2.
OpenAI built Pando to keep its existing Git workflows as its engineers and Codex agents create and move more code in parallel. The company says Pando cut median push latency by nearly 80% and improved merge reliability. Its public description presents Pando as an internal system; OpenAI says it is exploring ways to share the technology with users, but gives no release plan.
Codex increased pressure on OpenAI’s monorepo
The company’s existing source-control systems began struggling around the end of 2025, the post says. More concurrent work meant more branches, pushes, merges, fetches and clones, with several changes from one engineer sometimes ready for continuous integration at once. A single push could prompt several jobs to fetch the same commit.
OpenAI says merge reliability at times fell below 90% as the old system struggled with slow pushes and fetches. The team chose to replace the backend rather than change how developers use Git. “It was clear that we needed to revamp our source control system to scale to the level of concurrent work that Codex makes possible,” write Aaron Friel, Taylor Blau and Ted Nyman, OpenAI members of technical staff. Aaron Friel, Taylor Blau and Ted Nyman, OpenAI
Independent branches can update in parallel
Pando coordinates updates branch by branch, allowing pushes to unrelated branches to proceed without waiting on one repository-wide update. It stores incoming Git packfiles as immutable blobs and records branch versions and their dependencies in a metadata store. When clients fetch, Pando selects stored packs that contain the history they need instead of generating a separate packfile for each reader.
That design comes with a consistency tradeoff. A client reading multiple branches may see them at different points in time; Pando does not promise a single, synchronized snapshot of every repository reference during a clone or fetch. Updates to an individual branch remain serialized, but readers cannot assume that separate branch reads reflect one shared moment.
Pack reuse and regional caches speed fetches
Pando relies on Git’s packfile-uri feature to let clients download stored packs separately. OpenAI says its regional caches recorded hit rates above 99% for pack delivery, and that standard cloud virtual machines could saturate 10-gigabit-per-second network links during clone and fetch operations.
Background compaction combines smaller packs into larger covers, reducing the number of files clients need to retrieve and helping the server plan fetches as repositories grow. OpenAI says its monorepo uses packfile covers through level L4. The design keeps packs at multiple levels available to serve different fetches rather than discarding older pack versions after compaction.
OpenAI built Pando in 10 weeks
The project moved from kickoff to production in 10 weeks and began serving production traffic in April 2026, according to the post. One engineer designed its core, teams of agents implemented it, and a handful of engineers deployed it. OpenAI wrote the backend in Rust and tested it with a Git conformance suite and deterministic simulations aimed at finding race conditions and regressions.
The company says agents used GPT-6 Astra to write Lean and TLA+ formalizations that helped find bugs and prove system properties. OpenAI also contributed improvements to Git’s pack downloads and related tooling upstream. For now, the clearest boundary is practical: Pando runs OpenAI’s monorepo, while the company says it is still exploring how to share the underlying technology.