OpenAI is buying Pierre, the Git startup already inside Codex, two days after explaining why it built its own Git backend
OpenAI said on November 4 it will acquire Pierre Computer Company, whose Trees and Diffs libraries Codex already ships. No price, no closing date, and a design detail that sits oddly next to OpenAI's own Git post.
OpenAI put out an engineering post on November 2 about why the usual ways of scaling Git didn’t work for its own code. Two days later, on November 4, it announced it will acquire Pierre Computer Company, a small startup that sells Git storage to companies whose products write code with AI.
Read side by side, the two posts say more than either does alone.
What OpenAI said it needed
The Pando post, written by three OpenAI engineers, describes a monorepo that stopped keeping up once Codex let each engineer run many changes at once. Fetches and pushes got slow, merges failed, and reliability at times dropped below 90%, according to OpenAI. Its fix was Pando, an in-house backend that lets pushes to different branches go through in parallel. OpenAI says Pando cut median push latency by nearly 5x and now serves close to a petabyte of Git data a day.
It also lists two common designs it modeled and didn’t pick. One keeps full copies of a repository on several servers and has them agree on each branch update using three-phase commit, a protocol where a majority has to approve every change. OpenAI’s complaint about that design is that each push needs several round trips, and branches get updated one at a time.
The detail that caught my eye
Pierre’s Code Storage page says the service “leverages Git’s three-phase commit protocol” with a quorum-based architecture. That’s Pierre describing itself, and OpenAI’s post never names Pierre or any other vendor as the design it modeled. So I can’t tell you OpenAI rejected Pierre’s approach. Nothing in either post says so.
My read: it’s a strange pair of announcements, and probably a harmless one. One company monorepo is a different problem from what Code Storage sells, which is storage for many repositories created by AI apps. Still, if I were a Code Storage customer I’d want to know how much of Pando’s thinking ends up in the product, and OpenAI hasn’t said.
What the deal covers
| Item | What’s been said | Who said it |
|---|---|---|
| Buyer and target | OpenAI will acquire Pierre Computer Company, Inc. | OpenAI, November 4 |
| Price | Not disclosed | DevOps.com and Axios |
| Closing | Subject to customary conditions, including regulatory approvals | OpenAI |
| Already in Codex | Trees (file navigation), Diffs (reviewing changes) | OpenAI |
| Code Storage limits | Repositories up to 32TB, 30+ concurrent writes | Pierre’s own page |
| Named customers | Lovable, Poke, Anything, Bolt, Amp, Parahelp | Pierre’s own page |
| Open source | Pierre’s projects stay open source | OpenAI |
OpenAI says Code Storage customers can keep using it after the deal closes while the company works to bring its features into Codex. Until closing, the two stay separate companies. That sentence matters if your product stores its code on Pierre today, because “while we work to bring its features into Codex” doesn’t promise the standalone service lives on.
The stated goal is big. OpenAI wrote that it wants to “advance the future of source control across the industry” and to make Codex the place where developers and agents take on “their most ambitious projects.” Axios framed it more bluntly, as OpenAI targeting Microsoft’s GitHub, which has had reliability problems as AI coding took off.
Who else is in the race
Axios lists the field. Cursor, which Axios says is now owned by Elon Musk’s SpaceX, launched a GitHub rival called Origin in August. Former GitHub CEO Thomas Dohmke’s startup went public with its own rival last month, and I covered how Entire’s Trails replaces the pull request when it launched. In the speed test Entire published for its own launch, Pierre managed 15 pushes a second against Entire’s 418. That’s a competitor’s benchmark of a competitor, so treat it as a rough marker at most.
Microsoft is a complication here. Axios notes it owns nearly a quarter of OpenAI, while GitHub is Microsoft’s property.
What we don’t know
Nobody has said what OpenAI is paying. Nobody has said when the deal closes, whether Code Storage keeps its current pricing, or whether Pando will be sold to outside customers, which OpenAI only says it’s exploring. The deal also still needs regulatory approval, per OpenAI.
My read: the interesting part isn’t the Pierre price tag, which nobody has disclosed. OpenAI already depends on Pierre’s libraries every time someone reviews a diff in Codex, and buying the team is one way to stop depending on someone else. The harder question is whether a company that sells models can also sell the place where the code lives, and whether customers who compete with it will keep their repositories there.
Note on sources
I read OpenAI’s Pierre announcement and its Pando post in full, plus Pierre’s Code Storage page. The Axios report was read on Yahoo Finance’s reprint, and I used it for the competitor details (Cursor’s Origin, the Microsoft stake), which OpenAI doesn’t state itself. The 32TB and 30-writes limits are Pierre’s own claims, not independently tested. The comparison between Pierre’s three-phase commit wording and the design in OpenAI’s post is mine, and neither company says the two are linked.