Here’s what our cloud hosts had to say about this week’s episode: 358: AI Spend Limits Because Frontier Models Aren’t Free Therapy
Justin Brodley
Watching the slow train wreck at github over the last year has been quite sad. Between May 2025 and April 2026 they logged 48 major outages, about one a week, including a February failure that knocked out Actions, pull requests, notifications, and Copilot all at once, and an April merge-queue bug that silently reverted commits across 658 repositories. They blame exponential AI growth. CTO Vlad Fedorov apologized in April, said none of it was acceptable, and laid out the fix: a capacity plan that jumped from 10x to 30x, isolating Git and Actions so failures stop cascading, moving the worst paths off the old Ruby monolith, and going multi-cloud.
I think it’s more complacency. The Pragmatic Engineer pulled the real numbers and found load grew about 3.5x over two years, not 10x in a month, and the chart github used to make its case didn’t even have a Y axis. The worst failures weren’t capacity anyway. The merge-queue bug was an incomplete feature flag that nothing caught because the service still looked up. An expired TLS certificate took down every Git operation back in November. That’s not AI demand, that’s stuff a careful team hardens years earlier. Everyone thinks of github when they think git, and that grows some ego in an organization. It doesn’t help that nobody replaced Dohmke as CEO and the whole thing got folded into Microsoft’s CoreAI org.
Now that reputation is getting squandered. I hear people talking about gitlab, Zig already moved to Codeberg over Actions bugs, and Epic Games just shipped Lore, an open-source alternative built for the large files git is notoriously bad about. Mitchell Hashimoto is even pulling Ghostty and its 52,000 stars off the platform. Once developers vote with their feet it’s hard to get them back. The pricing doesn’t help either, ridiculous money for advanced security, $30 and $19 per committer for code scanning and secret protection, structured to push you onto an enterprise contract right when Mythos-class models are compromising code faster than ever. Security should be the floor, not the upsell. I like github a lot, and I want it to win. But the way back is simple: recover stability, build features people actually want, and don’t forget about the developers who got you here.
Leave a Reply