OpenAI has rebuilt the server side of Git for its monorepo. In an engineering post, it says Pando, “a new backend for Git repositories”, “now powers all development in OpenAI’s monorepo” and “now delivers nearly a petabyte of Git data per day”.
At a full petabyte, that is about 11.6 gigabytes every second, or about 93 gigabits per second around the clock, by our arithmetic. OpenAI president Greg Brockman then posted on X: “Considering providing this to others, let me know (gdb@openai.com) if you’d be interested”.
Business Pill · Jevons Paradox
An evergreen Business Engineer explainer on the Jevons paradox: when something gets cheaper to use, total use can rise. It teaches the general idea only and does not discuss OpenAI, Pando or any company in this story.
The key insight: As we read it, the number that matters is not the petabyte but where OpenAI put the fix: it kept the Git client and protocol and rebuilt the server behind them, so engineers and agents did not have to change how they work. At a full petabyte a day, Pando moves about 93 gigabits per second around the clock, by our arithmetic.
The post says Pando was built “without interrupting any engineering work or changing how our team uses Git”. Engineers and agents keep using Git and its protocol, now through OpenAI’s own Git build; the main change is the server behind them.
It had two goals: “horizontally scalable writes, optimized for concurrent pushes to many branches”, and to “maximize read throughput, so clients can download data fast enough to fully saturate their available network bandwidth”.
OpenAI says it “built Pando with Codex, moving from kick-off to production, in just 10 weeks”. The core was “designed by a single engineer, implemented by teams of agents, and deployed by a handful of engineers”, and Pando “began serving production traffic in April of this year”. It is written in Rust.
OpenAI’s disclosed Pando throughput converted to gigabits per second: nearly a petabyte of Git data a day is about 93 Gbps (11.6 GB/s) averaged around the clock at a full petabyte, and the peak hour of more than 57 terabytes is about 127 Gbps (15.8 GB/s), against the 10 Gbps links OpenAI says clones and fetches saturate. Conversions are our arithmetic, in decimal units. Source: OpenAI, Scaling Git at OpenAI.
What Was Posted on X
We built Pando, a new Git backend for our monorepo, to handle more pushes and fetches without changing how our engineers and agents use Git.
Pando cut push latency by ~80% and now serves nearly a petabyte of Git data a day. We can grow Git capacity alongside agent usage.
OpenAI’s developer account announced Pando on 3 November (UTC): a new Git backend for its monorepo that, it says, cut push latency by about 80% and serves nearly a petabyte of Git data a day. The piece checks these figures against OpenAI’s full engineering post.
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. Considering providing this to others, let me know (gdb@openai.com) if you’d be interested.
OpenAI president Greg Brockman quoted that post, saying OpenAI kept Git so Pando needed no change to how developers work, and that it is ‘Considering’ offering Pando to others. His post names no product, price or date.
Why Codex Broke the Old Setup
The post puts the cause on Codex. More parallel work meant “more branches and pushes, more merges, and more fetches and clones from development environments, research jobs, and CI”.
It adds: “One engineer now has several changes ready for CI at once, and each push can set several jobs fetching the same commit”. From around the end of last year, OpenAI says, it “suffered slow fetches and pushes along with failed merges, with reliability at times falling below 90%”.
How Pando Works
Pando makes two trades. First, it coordinates each branch on its own, so pushes to unrelated branches run in parallel. The cost, in OpenAI’s words: “a reader looking at several branches can see them at different points of time”. The post calls this “causal consistency”.
Second, it reuses packfiles, Git’s compressed bundles of objects, across fetches “instead of generating an individual packfile for each client”. Each push is stored as an immutable blob, and later readers are served the same bytes. A delivery service called Express caches them by region, with “cache hit rates of over 99%”.
To keep those packs reusable, Pando shows a pushing client only a designated trunk branch rather than every branch. With millions of branches, the post says, “the ref advertisement alone can exceed 100 MiB”. Under the trunk-only approach, the median pack was “under 20 KiB” and the 99th percentile “under 3 MiB”.
Background workers merge small packs into larger ones in levels, as in log-structured merge-trees; the monorepo “uses up to L4 packfile covers”. OpenAI says that on standard cloud virtual machines it has seen “clone and fetch operations saturate 10 Gbps network links”.
The Numbers OpenAI Gave
Switching to Pando “improved merge reliability and reduced the latency of the median push by nearly 80%”, which the post also describes as “nearly 5x”.
The post says “Pando can deliver a petabyte of Git data per day with just a handful of CPU cores” and “has served more than 57 terabytes of Git packs in a single hour, with more than 99% of those bytes from local disk cache”.
By our arithmetic, 57 terabytes in an hour is about 15.8 gigabytes a second, or about 127 gigabits per second. That is about 1.4 times the average hour at a full petabyte a day, and the daily average alone equals about nine fully used 10 Gbps links.
Kept Git, or Replaced It
Anthropic described the same kind of strain in its own monorepo and went the other way. In a 13 October engineering post, it said it “built Scion, an in-house version control system, and fully migrated our monorepo from git in five months”, with “A team of seven people”.
Brockman framed OpenAI’s choice this way: “We took the hard road of continuing to use Git, which means we could enable Pando with no change to devex”.
The two companies also differ on sharing. Anthropic wrote that “it’s difficult to offer Scion in a form that would be broadly useful to other organizations”. OpenAI’s post says: “We’re currently exploring ways to share the technology that makes Pando possible with our users”.
As we read the two posts, keeping Git is what makes an outside offer plausible: a customer could keep using Git and its existing tools, where Scion would require migrating “every piece of the development lifecycle”, as Anthropic put it.
As we read the post, agents changed the shape of the load as much as its size: many parallel branches, each push setting off several jobs fetching the same commit. Pando answers with reuse, serving the same stored bytes to many readers.
As we read Brockman’s post, the offer is a question, not a launch. Keeping Git is what would let an outside team adopt it without moving every tool, which is the step Anthropic said makes Scion hard to share.
Greg Brockman, president of OpenAI, on X, 3 November 2026
“We took the hard road of continuing to use Git, which means we could enable Pando with no change to devex”
Three Implications
FOR OPENAI Brockman wrote that OpenAI is “Considering providing this to others”, and the post says it is exploring ways to share the technology. No product, price or date is given.
FOR TEAMS RUNNING AGENTSOpenAI says one engineer now has several changes ready for CI at once, and each push can set several jobs fetching the same commit. Source control is one of the systems that load lands on.
FOR GIT ITSELFOpenAI says several of its client improvements are already in upstream Git, and others are in its public repository. Its approach keeps the Git protocol and client.
The framework’s thesis on software is that “One super-harness absorbing the full softwarelife cycle is already happening inside OpenAI”.
As we read it, Pando is that absorption reaching source control: a system OpenAI says was implemented by teams of agents to carry agents’ pushes and fetches. The framework was published in July 2026 and does not discuss Pando.
Brockman wrote “Considering”, and the post says OpenAI is “exploring ways” to share the technology. Neither gives a product, a price, a date or a list of customers, and neither says whether an offer would be hosted by OpenAI or released as software.
The post does not name the system Pando replaced or who ran it, and gives no daily count of pushes, commits or agents. It gives the latency cut as a share, not in seconds. OpenAI says it is not keeping its Git client changes to itself: “Several of these improvements are already in upstream Git”.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
OpenAI says Pando, a Git backend built with Codex in 10 weeks, now delivers nearly a petabyte of Git data a day and cut median push latency by nearly 80%. That daily volume is about 93 gigabits per second around the clock, by our arithmetic. Brockman says OpenAI is considering offering it to others.
A note on sourcing. Pando facts and quotes are from OpenAI’s engineering post, read in full; Greg Brockman’s words are from his post on X. Scion facts are from Anthropic’s engineering post of 13 October. Conversions to gigabytes and gigabits per second, the peak-hour ratio and the link count are our arithmetic, in decimal units and at a full petabyte.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
