OpenAI Acquires Pierre Computer to Bolster Codex Storage

OpenAI Acquires Pierre Computer to Bolster Codex Storage

by

OpenAI’s internal Git backend now delivers nearly a petabyte of code data per day, and switching to it reduced median push latency by nearly 80%. On Tuesday, Axios reported that OpenAI will acquire Pierre Computer, a developer-tools startup whose technology is already used inside Codex, with plans to integrate Pierre’s Code Storage product directly into the AI coding system used for generating and changing software.

The acquisition, detailed by unite.ai, places OpenAI among the companies challenging GitHub, the Microsoft-owned platform that has long served as the central repository and coordination point for software development. The move comes as AI agents take on more coding work, dramatically increasing the volume of software that repositories must store and serve. Pierre’s open-source projects will remain open source after the deal closes, preserving the underlying components that other software can assemble and use.

OpenAI Accelerates Internal Repository System

OpenAI’s push into storage infrastructure reflects mounting pressure on its own systems. In a November 2, 2026 engineering post, the company described Pando, its backend for Git repositories that powers all development in its monorepo. Pando maintains compatibility with the Git protocol and clients, enabling engineers and agents to adopt it without altering their existing workflows.

It went from kickoff to production in 10 weeks with the help of Codex and started handling production traffic in April 2026. Before Pando, Codex allowed engineers to do more work in parallel, producing more branches, pushes, merges, fetches, and clones than earlier monorepo systems could handle. Those older systems experienced slow fetches and pushes and failed merges, with reliability at times falling below 90%. That connected the growth of agent-assisted coding directly to the demands placed on source-control infrastructure.

Pando coordinates updates to each branch independently, allowing writes to unrelated branches to proceed in parallel. Additionally, it reuses packfiles—Git’s compressed object collections—across fetches rather than building a distinct packfile for each client. The architecture serializes updates to single references—the pointers for branches and tags—yet does not provide someone reading multiple branches with one consistent point-in-time snapshot of the repository. That tradeoff differs from Code Storage’s quorum-based architecture, which Pierre says guarantees consistency across replicas.

The distinction matters for reliability at scale. While Pando prioritizes throughput by allowing parallel writes to different branches, Code Storage’s quorum model ensures that every read sees a consistent, durable state across all replicas. For AI agents working simultaneously on shared codebases, that difference determines whether an agent sees its own changes reflected immediately or risks building on stale data.

The numbers illustrate how dramatically the workload has shifted. Nearly a petabyte of Git data per day means OpenAI’s internal systems now move more code than many hosting platforms handle for all their customers combined, and the 80% latency reduction shows the gap between human-speed editing and agent-speed commits.

OpenAI Acquires Pierre Computer to Bolster Codex Storage

Code Storage Brings Machine-Oriented Design to Codex

Pierre Computer’s Code Storage addresses that consistency gap with a quorum-based architecture built on Git’s three-phase commit protocol. The distributed storage system features API ergonomics designed for machines, capable of holding generated code and other files that change over time, including agent memory files, design artifacts, and legal documents that accumulate as AI systems work.

Applications can create and manipulate repositories through the native Git protocol or software development kits. Code Storage offers Git clone, push, and fetch endpoints within customers’ own applications, supplies TypeScript, Python, and Go SDKs, supports webhooks, and syncs with GitHub, GitLab, and Bitbucket. Those features let coding platforms connect repository storage to build systems, bots, and agents without rebuilding storage from scratch.

Pierre names Lovable, Poke, Anything, Bolt, Amp, and Parahelp among the service’s users. Its documented specifications cover 30 or more simultaneous writes, repositories up to 32TB in size, and Git LFS support, highlighting the product’s focus on heavy machine workloads and large stored files. The company describes itself as a builder of low-level, composable infrastructure for machine work, listing Diffs, Trees, and DiffsHub as open-source projects alongside the Code Storage product.

For developers and businesses building AI coding platforms, the practical stakes are immediate. As agents take on more software work, repositories must store and serve far more data than human-only workflows required, and the infrastructure beneath that code becomes part of the coding product itself. The proposed deal hands OpenAI a storage solution already handling machine-focused workloads next to Codex, whereas Pando tackles the expansion of parallel development within its monorepo. OpenAI is assessing scenarios in which many agents share a single Git repository and investigating methods to offer the technology to users. What has not been answered is how quickly Code Storage will surface inside Codex for external developers, or whether OpenAI will sell it as a direct competitor to GitHub’s enterprise storage services.

More on this

Read next