Two OpenAI announcements this week, two days apart, and together they say something louder than either one alone: OpenAI is building its own version control plumbing, and it does not plan to rent it from Microsoft forever.
First came Pando. OpenAI’s in-house Git backend, quietly put into production in roughly ten weeks, with Codex agents handling a lot of the migration work. The claimed numbers: median code-push latency down nearly 80%, and nearly a petabyte of Git data delivered daily. A petabyte a day. That is what happens when millions of agent sessions are cloning, diffing, and pushing real repositories around the clock. Agent-driven code movement has a different traffic profile than human-driven, and the humans who built GitHub’s backend in 2008 were not designing for it.
Then came the acquisition. On November 4, OpenAI agreed to buy Pierre Computer Company, Inc., the startup behind Code Storage, and fold it into Codex. Terms undisclosed. Code Storage is a distributed storage system built on Git with, and this is the phrase that got me, “API ergonomics designed for machines.” Repositories up to 32TB. More than 30 concurrent writes. Git LFS. TypeScript, Python, and Go SDKs. Webhooks. Sync with GitHub, GitLab, and Bitbucket.
That phrase, “designed for machines,” is the tell. Traditional source control treats the repository as a place humans read and write, with tools helping. AI-first infrastructure inverts it: the primary writer is an agent, the human shows up at review time. Concurrency requirements go through the roof because a hundred agent sessions work in parallel, and the API surface matters more than the web UI because nothing on the other end has eyes.
The customer list of Code Storage reads like a map of where agent-native development already lives: Lovable, Poke, Anything, Bolt, Amp, Parahelp. These are the products where the human writes a prompt and the machine does most of the commits. They are already off the traditional plumbing, at least partially, because traditional plumbing was not built for their write patterns.
Pierre’s open-source projects, Diffs, Trees and DiffsHub, stay open source after the deal, which is the part that keeps this from feeling like a pure land grab.
Now the awkward framing nobody at OpenAI is stating out loud. Microsoft owns GitHub, the dominant source control platform on earth, and also owns roughly a quarter of OpenAI. Pulling source-control plumbing in-house means Codex gets storage, diffs, and review tooling that do not route through a Microsoft property. For over a decade, source control was a solved problem nobody would touch, because why rebuild Git hosting. The answer in 2026: because the write traffic changed, because agents need machine-ergonomic APIs, and because whoever owns the layer underneath AI-written code owns the telemetry, the review loop, and eventually the economics of every product built on top.
Watch what happens to Code Storage’s existing customers after integration. The bet OpenAI is making is that agent-written code volume keeps compounding, and that a Git backend tuned for agents beats one tuned for humans at that game. If the compounding is real, and the petabyte-per-day number suggests it is, then source control stops being a solved problem and becomes contested infrastructure, the same way browsers became contested infrastructure once the web got valuable.
For builders, the takeaway is not “buy OpenAI stock”. It is that the layer you assumed was boring and stable, the place your code lives, is being re-litigated specifically for agent workloads. If your product writes code through agents, ask where it stores, how it diffs, and who sees the traffic. The answers are going to change under you over the next two years, and the teams who treat source control as a commodity will be the ones surprised by it.