Mitchell Hashimoto — co-founder of HashiCorp, creator of Vagrant, Terraform, and some of the most important infrastructure tools in modern computing — just quit GitHub after 18 years. His reason: "GitHub is no longer a place for serious work." His Ghostty terminal emulator project is moving elsewhere. The GitHub status journal shows outages "almost every day." The cause? AI. The same AI tools that depend on GitHub are generating the load that's breaking it.
The AI Load Paradox
Here's the vicious cycle:
-
AI coding agents need GitHub — They clone repos, read code, create branches, submit PRs, run CI/CD. GitHub is the central nervous system of AI-assisted development.
-
AI generates massive GitHub load — AI-generated PRs jumped from 4M/month to 17M/month in six months. 90% of those PRs are noise. Each one still consumes compute, storage, and bandwidth.
-
The load breaks GitHub — GitHub's CTO published his second availability update in six weeks, citing "record acceleration" in usage. The platform can't handle the volume.
-
Broken GitHub breaks AI tools — When GitHub goes down, every AI coding agent that depends on it fails. The tools that caused the problem become victims of it.
Hashimoto's departure is the most visible symptom of this cycle. When one of the most respected infrastructure engineers in the world says a platform is too unreliable for serious work, it's not a hot take — it's a credibility crisis.
The Numbers
- 4M → 17M PRs/month in 6 months (325% increase, mostly AI-generated)
- 90% of AI PRs are noise — automated spam flooding the review queue
- GitHub outages "almost every day" per Hashimoto's status journal review
- GitHub CTO: "record acceleration" in platform usage — second availability post in 6 weeks
- Five major incidents in 48 hours in early April
Why This Matters Beyond GitHub
GitHub is the canary in the coal mine for infrastructure collapse under AI load. The same pattern is playing out across every platform that AI depends on:
npm, PyPI, and Package Registries
AI agents generate and publish packages at increasing rates. Registries face the same flooding problem that GitHub faces with PRs.
CI/CD Systems
Every AI-generated PR triggers CI/CD pipelines. Compute costs for testing are exploding as pipelines run on code that 90% should never have been submitted.
API Providers
Anthropic, OpenAI, and Google are all struggling with capacity. Claude had another major outage April 28 (6,500 users affected). The compute-cost wall is real.
Cloud Infrastructure
Enterprise GPUs at 5% utilization ($37B wasted), but compute availability for AI training is still constrained. The system is simultaneously over-provisioned and under-available.
What This Means for Development Teams
1. Don't Depend on a Single Platform
If GitHub can go unreliable, anything can. Build platform-agnostic development workflows:
- Use git (the protocol), not GitHub (the platform) as your dependency
- Mirror repositories to multiple providers (GitLab, Bitbucket, self-hosted)
- Design CI/CD pipelines that work with any git hosting provider
2. Filter AI-Generated Contributions
If your team uses AI agents that submit PRs, implement quality gates before they hit the main repository:
- Require AI-generated PRs to pass automated validation before submission
- Set limits on AI PR volume per day/week
- Auto-close PRs that don't meet quality thresholds
3. Prepare for Platform Migration
Hashimoto is moving Ghostty off GitHub. Your team should have a migration plan too:
- Document all GitHub-specific features your team uses (Actions, Packages, etc.)
- Identify alternatives for each feature
- Test migration on a small project first
- Keep the plan current — don't wait until GitHub is down to figure it out
4. Invest in Self-Hosted Infrastructure
The most reliable infrastructure is the infrastructure you control:
- Gitea or Forgejo for self-hosted git
- Jenkins or Woodpecker CI for self-hosted CI/CD
- Poolside Laguna XS.2 for local AI coding (68.2% SWE-bench on a Mac)
5. Reduce Your AI Footprint on Shared Platforms
Be part of the solution, not the problem:
- Don't let AI agents submit PRs to public repositories without filtering
- Use local AI tools (Laguna XS.2, DeepSeek V4-Flash) for iteration instead of cloud-dependent agents
- Batch AI operations to reduce API calls
Honest caveat: Self-hosted infrastructure requires maintenance, security updates, and expertise. Not every team has the capacity to run their own git hosting and CI/CD. The trade-off is between convenience (cloud platforms) and reliability (self-hosted). For teams doing critical work — like Hashimoto's Ghostty — reliability wins.
The Financial Impact
Cost of GitHub Unreliability
| Impact | Cost | |--------|------| | Developer downtime during outage (50 devs × $100/hr × 2hrs) | $10,000 per outage | | Failed CI/CD pipelines from outages (100 pipelines × $5) | $500 per outage | | Context switching and recovery time | $5,000 per outage | | Total per major outage | $15,500 | | Estimated annual cost (3-4 major outages/month × 12) | $558,000-744,000 |
Cost of Migration to Self-Hosted
| Item | Cost | |------|------| | Gitea/Forgejo setup and configuration | $5,000-10,000 (one-time) | | CI/CD migration | $10,000-20,000 (one-time) | | Ongoing maintenance (0.25 FTE) | $20,000-30,000/year | | Total first-year cost | $35,000-60,000 |
ROI of self-hosting: 9-12× return in the first year based on outage cost avoidance.
Closing Thoughts
When Mitchell Hashimoto — the person who literally wrote the book on infrastructure automation — says a platform is too unreliable, the industry should listen. This isn't a casual complaint from a frustrated user. It's a professional assessment from someone who has spent decades building reliable systems.
The AI load paradox isn't going away. AI agents will generate more PRs, more API calls, more compute demand. The platforms that weren't designed for this volume will continue to buckle. The teams that recognize this and build resilient, platform-agnostic workflows will keep shipping. The teams that don't will keep refreshing the GitHub status page.
GitHub was built for human-scale collaboration. AI-scale collaboration is a fundamentally different problem. Until the platform catches up — if it ever does — the smart move is to reduce your dependency on it.
Worried about platform reliability? Book an Infrastructure Resilience Assessment — we'll audit your development workflow dependencies, design platform-agnostic alternatives, and build a migration plan that keeps your team shipping regardless of which platform goes down next.