Thursday Thoughts: GitHub Is Cracking Under the Weight of AI

A few people have asked me lately what I think the future of GitHub looks like. Cursor releasing their new forge product, Origin, brought this question back into focus. It’s a genuinely interesting one, and I don’t think most people have sat with it long enough.

GitHub was built for humans, entirely for humans. The architecture assumes a relatively slow rate of change, a manageable number of commits, and a person deciding at every meaningful step. When GitHub added webhooks, Actions, and CI/CD over the years, it didn’t rethink that foundation. It just bolted new pieces on.

That worked fine when humans wrote code. It’s not working as well now.

GitHub’s Architecture Was Never Designed for This Volume

We’re in the middle of the same shift that happened when companies moved from on-premises infrastructure to the cloud. Back then, plenty of organizations just lifted their old workflows into cloud environments and called it a transformation. It wasn’t. The tools looked different but the thinking hadn’t changed, a pattern I saw again from the AI-native side.

That’s what happened with GitHub and agents: human workflows with agents dropped in. GitHub wasn’t built for this velocity, and the numbers say so. Pull requests opened by AI agents jumped from roughly 4 million to 17 million between September 2025 and March 2026. Weekly commits reached 275 million, on pace for roughly 14 billion in 2026, about 14 times 2025’s total. An analysis by LeadDev counted 257 incidents between May 2025 and April 2026, 48 of them major, and GitHub’s own CTO, Vlad Fedorov, has said the platform “wasn’t built for the scale it’s now being asked to handle” and must design for 30 times today’s load.

Cracks means outages. And although it’s fun to have a “digital snow day” as a dev, GitHub outages are productivity killers.

What Origin Is Actually Trying to Do

Origin is a rethink of what a forge looks like if you start from agents rather than bolt them on afterward. Cursor unveiled it at its Compile conference on June 16, 2026, with demo numbers that were the whole pitch: 22.6 commits per second in a single repository, hundreds of thousands of clones per hour, sub-400ms global sync. Agent-scale numbers, not human-scale ones.

And it’s not hypothetical anymore. Cursor shipped Origin’s early beta on August 17, 2026, the same week GitHub had a major outage that hit its website, API, Actions, pull requests, authentication, and Copilot all at once, timing Cursor couldn’t have bought with launch copy.

A few things stand out beyond the throughput claims.

It’s designed around stacked commits, lining up with Cursor’s acquisition of Graphite, the stacked-review startup it bought in December 2025. Agents produce lots of small, rapid changes and spawn parallel workstreams, a pattern stacked commits fit far better than a traditional PR-centric model.

It also gives agents a native interface instead of routing everything through expensive, high-latency webhooks. Latency and overhead compound fast at agent scale.

And importantly, Origin is still built on Git. The early beta even mirrors GitHub rather than replacing it outright. GitHub stays the system of record at first, and a repository can be detached later to make Origin authoritative. That’s a smart call. One of the real barriers to enterprise adoption of anything new is the switching cost hidden inside familiarity, and staying compatible with GitHub during the transition lowers that barrier considerably.

Why This Might Actually Work, Despite the Odds

In normal times, I’d say displacing GitHub is essentially impossible. Every enterprise has years of workflows, integrations, and institutional muscle memory baked into GitHub, dating back to Microsoft’s 2018 acquisition of it for $7.5 billion. Switching costs aren’t theoretical. They’re painful and real.

But these aren’t normal times, and Microsoft’s vulnerability goes deeper than server architecture. Copilot’s own developer share is sliding. The 2026 Stack Overflow survey put it at 51%, down from 67% the year before, while Cursor and Claude Code went from a standing start to double digits, and JetBrains’ 2026 survey found the three roughly bunched at 29%, 18%, and 18%. Copilot isn’t the clear leader in the category Microsoft invented anymore.

Part of that is architecture, not just model quality. Copilot shipped as a plugin bolted onto VS Code and GitHub rather than a rethink of either, and developers noticed. Microsoft also used its grip on the editor to squeeze rivals: in 2025 it broke closed-source extensions like Pylance and C/C++ IntelliSense for VS Code forks, a move devs called anti-competitive, and one that only mattered because a fork had already taken real share. Cursor is proof that even VS Code, with 70%-plus share and Microsoft’s full backing, wasn’t safe from a startup willing to rebuild the editor around agents. If VS Code was vulnerable, GitHub isn’t obviously immune either.

Meanwhile, SpaceX’s $60 billion buyout of Cursor’s parent company, Anysphere, the largest venture-backed startup acquisition on record, has made them genuinely well-funded and positioned to compete in the enterprise space in a way most startups simply aren’t. They have resources and momentum, plus a vertically integrated stack now: Composer for editing, Graphite for review, Origin for hosting.

The period when switching costs are the most material barrier is closing. But the better product doesn’t automatically win that fight, a lesson I keep coming back to from Claude Code vs. Lotus 1-2-3: distribution has beaten quality before.

The Question That Will Decide Everything

The one thing I keep coming back to is whether Origin will be open or proprietary. That’s not a small detail. It’s probably the whole game.

If Origin is built openly, it has a real path into the enterprise. Enterprises are cautious, and open standards give them the confidence that they’re not locking themselves into a vendor with a single point of failure. It also means the broader ecosystem can build on top of it, which accelerates adoption in ways that a closed system simply can’t replicate.

If Origin ends up being closed, designed to keep developers inside the Cursor and xAI orbit, the enterprise story gets much harder. Enterprises have been burned by lock-in before and remember it, the same pattern I saw with Anthropic and AWS.

I don’t know which direction they’ll go. But I think that decision, more than anything else about the product itself, will determine whether Origin is a footnote or a genuine inflection point.

---

We’re still figuring out what AI-native tooling looks like at every layer of the stack. The forge is just one piece of it, but a consequential one, because whoever controls where code lives has a lot of leverage over everything downstream.

What would it take for you to move your organization off GitHub?

By the Numbers

• 22.6 commits per second — the throughput Cursor demoed for Origin in a single repository at its June 16, 2026 launch.
• 17 million — pull requests opened by AI agents on GitHub in March 2026, up from roughly 4 million in September 2025.
• 257 incidents — GitHub outages LeadDev counted between May 2025 and April 2026, 48 of them major.
• 51% — Copilot’s share in the 2026 Stack Overflow Developer Survey, down from 67% the year before.
• 70%-plus — VS Code’s market share when Cursor forked it anyway and started taking share.
• $60 billion — the all-stock deal SpaceX used to acquire Cursor’s parent company, Anysphere, the largest venture-backed startup acquisition on record.
• 2018 — the year Microsoft bought GitHub for $7.5 billion, the incumbent advantage Origin is trying to unseat.
• 3 — layers of Cursor’s now-vertically-integrated stack: Composer for editing, Graphite for review, Origin for hosting.