GitHub Scales 30x for AI Coding Boom as Ghostty Exits to Self-Hosted Forge

GitHub just disclosed it scaled its core platform 30x in three years to handle the AI coding boom — and on the same day, Mitchell Hashimoto's Ghostty terminal announced it's leaving GitHub for a self-hosted forge. Two stories, one signal: the platform that runs open source is straining at both ends — capacity at the top, trust at the bottom.
GitHub's engineering blog described request volumes that would've collapsed the 2021-era stack, driven by Copilot, AI agents hammering APIs, and a generation of CI runs spawning more CI runs. Ghostty's exit, meanwhile, points at something quieter: developers who feel GitHub Issues, Actions, and the company's AI direction don't fit how their projects actually work anymore.
What GitHub had to rebuild to absorb 30x
The scaling post is unusually candid. GitHub admits the old monolith couldn't keep up with what AI coding agents do — they don't browse, they brute-force. A single Copilot session can fire dozens of API calls per second; agentic frameworks like Devin or Cursor's background agents do worse. GitHub split read paths, moved hot tables to Vitess shards, rewrote rate-limiting around per-token rather than per-user budgets, and rebuilt its CDN edge to cache logged-in views — something it had resisted for a decade.
The numbers are wild. Git LFS bandwidth is up 9x, Actions minutes are up 22x, and the API surface — the part agents touch — is up 30x. None of that came from human developers typing more code. It came from agents reading the codebase before they write a single line.
Why Ghostty's exit matters more than the size of the project
Ghostty isn't huge. It's a terminal emulator with maybe 35K stars and a small contributor base. But Hashimoto is the founder of HashiCorp, and his projects don't move in random directions. The stated reasons — issue tracker noise, Actions billing surprises, AI-generated PR spam — are the exact friction points other maintainers have been quietly complaining about for 18 months.
The deeper signal: Ghostty is moving to a self-hosted instance of Forgejo (a Gitea fork). That's a deliberate vote against the centralized-forge model GitHub built. If even a few high-profile projects follow, the network effect that's kept GitHub unassailable since 2018 starts to look fragile.
The AI feedback loop GitHub created
GitHub's own scaling problem and the Ghostty exit have the same root cause: AI coding tools changed who the customer is. The platform was designed for humans collaborating asynchronously. Now most of its load — and most of its noise — comes from machines. Copilot creates the demand for the API; bots create the spam in the issues; agents create the PRs that maintainers have to filter.
GitHub's solution is more infrastructure. Maintainers' solution is to leave. Both are rational. Neither is sustainable forever.
My Take
GitHub will absorb the 30x — they're owned by Microsoft and they have the budget. The Ghostty thing is what should worry them. For 15 years GitHub's moat was that everyone was on GitHub. That moat held because the alternative was self-hosting a clunky tool. Forgejo and Gitea aren't clunky anymore, and the AI-generated noise is making the centralized model actively painful for serious maintainers. I think we're 18 months from the first major Linux-adjacent project leaving in a way that hurts. GitHub should take Ghostty more seriously than they probably are.
FAQ
Did Ghostty actually leave GitHub? Yes — Hashimoto announced the migration to a self-hosted Forgejo instance. The repo is being moved, not just mirrored.
Is GitHub at risk of capacity failure? Not in the near term. The 30x post reads as a victory lap, not a distress signal. But the load curve is still going up.
What's Forgejo? A community-governed fork of Gitea, itself a Go-based GitHub alternative. Self-hostable, lightweight, and increasingly capable of running CI without renting Actions minutes.
The Bottom Line
GitHub solved the engineering problem and ignored the sociological one. The 30x scaling story is a flex; the Ghostty exit is the warning. Centralized forges are still dominant — but the cracks are forming where it matters most: among the maintainers who actually keep open source moving.
Related Articles
- 35% of New Websites Since ChatGPT Are AI-Generated
- OpenAI Publishes AGI Development Framework
- A 23-Year-Old Used ChatGPT to Crack a 60-Year-Old Erdős Conjecture