The Black Cloud

AI - OpenAI Pando and why AI companies are rethinking Git

AI companies are now generating so much code that they are rebuilding the systems that store and serve it. I find this a fascinating turn of events: first the race was to write the code, and now the infrastructure is struggling to keep up 🤣

OpenAI’s Pando is the latest example, but Cursor, Anthropic and Google all have relevant stories here. Some are improving Git, some are replacing it internally, and some are competing with GitHub. Those are different things, but the wider pattern is getting rather interesting.

OpenAI Pando

Pando is a new Git backend powering OpenAI’s monorepo. The important part is that OpenAI kept the Git protocol and familiar workflows rather than asking everyone to learn a different version control system.

Official announcement

Scaling Git at OpenAI

What is new?

OpenAI says growing Codex usage meant more concurrent branches, pushes, fetches and CI downloads. Pando is its answer to that infrastructure pressure:

That petabyte figure is traffic served, not a petabyte of unique new source code written every day. Still a rather impressive amount of Git traffic!

There are trade-offs. Individual branch updates remain ordered, but reading several branches does not guarantee one repository-wide point-in-time snapshot. Reusing packfiles can also mean transferring objects a client already has. This is a deliberate engineering choice, not a magic removal of every scaling limit.

Built with Codex in ten weeks

Alex highlights the development timeline:

OpenAI’s more precise account is that one engineer designed the core, teams of agents implemented it, and a handful of engineers deployed it. The project went from kick-off to production in ten weeks.

The agents helped create the infrastructure problem, then helped build the infrastructure to deal with it. There is something quite amusing about that 🤣

Keeping Git was the point

Greg Brockman puts the compatibility decision nicely:

Pando is a Git backend which can scale to very large monorepos. We took the hard road of continuing to use Git, which means we could enable Pando with no change to devex.

Keeping familiar workflows does not mean there was no migration work: OpenAI also shipped its own Git distribution and migrated integrations. It describes upstream contributions and a public Git fork, which should not be confused with releasing Pando itself as open source.

NOTE
As of 3 November 2026, Brockman is considering providing Pando to others. That is an expression of interest in outside access, not an announcement of a generally available service that we can all sign up for.

Cursor / SpaceXAI: Origin is more than a GitHub rival

I previously covered Cursor launching Origin as a GitHub alternative. That was already a sign that AI companies wanted more control over the infrastructure around their agents.

Origin hosts repositories, pull requests and agent workflows. For Origin-hosted repositories, GitHub is not in the push path; for synced GitHub repositories, GitHub remains the source of truth.

But the connection to Pando goes deeper than hosting. Cursor’s Git at any scale article explains its Git storage design: a write-ahead log in S3, with replicas following that log and sharing compacted data.

Cursor deliberately keeps a globally consistent view of each repository, whereas Pando relaxes repository-wide consistency to let independent branches scale. Different trade-offs, but both are rethinking the server behind Git rather than simply putting another interface in front of it.

Anthropic: Scion actually replaced Git internally

I have written about Claude Code desktop and parallel sessions using Git worktrees, but Scion is a much bigger change.

In its 13 October article, Redesigning source control at Anthropic, Anthropic says it moved its monorepo from Git to its own version control system in five months.

The scale explains the motivation:

Those are Anthropic’s results for its own workload, not a promise that everybody moving away from Git would get the same improvement.

Its own caveat is worth keeping:

Scion is by no means a suitable replacement for git for general use.

This is a bespoke system for enormous, always-online workloads. Anthropic says making it broadly useful elsewhere would be difficult, because adoption also means changing code review, search and CI. Not quite a case of changing your Git remote and carrying on 😉

Google: Piper came first, Antigravity improves the workflow

Google’s earlier work includes Piper, an in-house, centralised version control system. It is described in Software Engineering at Google, and Google’s large-scale monorepo approach was documented in a 2016 research paper.

That history matters. Piper is not a new reaction to the current coding-agent boom; it is earlier work that Anthropic explicitly acknowledges as an influence on Scion.

More recently, I covered Antigravity’s Git controls and embedded terminal. That update improves reviewing, staging and committing changes, including changes made outside the agent’s own editing tool.

That is a Git workflow improvement, not another backend or a replacement for Git. Useful, but a different part of the same development process.

My take

I think the interesting story is bigger than another round of predictions that Git is dead.

OpenAI and Cursor are rebuilding Git infrastructure. Anthropic has replaced Git for its internal monorepo. Google provides both older large-scale precedent and newer improvements to how people work with agent-generated changes.

The AI race is clearly spilling beyond models and coding assistants into hosting, version control, review and CI. Generating more code only helps if we can reliably store, inspect, test and merge it.

The supposedly boring infrastructure is suddenly becoming one of the most interesting parts of the AI race 😊

<< Previous Post

|

Next Post >>

#Blog #AI #OpenAI #Pando #Git #Cursor #Origin #SpaceXAI #Anthropic #Scion #Google #Piper #Antigravity #Open-Source