After a solid year of outages caused by config files, control planes, and automation eating itself, GitHub went down on Monday because it ran out of room.
Seven hours and forty-seven minutes, github.com and auth and Actions and the API and Copilot, worldwide. No bad deploy, no poisoned config, no clever race condition. Record traffic hit a Central US data center, the Istio sidecar proxies that broker service-to-service traffic hit their processing ceiling, and the autoscaler — which scales on the services and not on the sidecars mediating them — didn’t fire, because from its point of view nothing was wrong. The pressure cascaded into the load balancers and then into authentication, which is how a capacity problem becomes an everything problem. And then during recovery a VS Code retry bug turned every waiting client into a load generator, so the system’s own users held the door shut while the team tried to open it. Thundering herd, thirteen years and counting, undefeated. GitHub’s writeup notes this was their second capacity failure of the month, after Actions on August 6, and that neither one involved a code or config change.
The part I keep circling: the traffic was real. Coding agents don’t take lunch breaks, don’t sleep, and retry politely and constantly, and the load curve of a platform whose users are increasingly machines looks nothing like the curve its capacity model was fitted on. GitHub is buying its way out — three million CPU cores and 120 petabytes, per the post — and that’s the right move at their scale. The lesson for everyone smaller is cheaper: your traffic shape changed when your users started running agents, and if your autoscaler measures the thing it always measured, it is measuring the wrong thing now. Ours scales on request rate and CPU. Neither would have caught this. That’s the sprint.
A confession, per the format’s rules. I led last entry with four cloud providers and missed the biggest infrastructure story of the summer, disclosed the same week: OpenAI ran a cyber-capability evaluation with the safety classifiers deliberately turned down — a legitimate red-team technique for measuring raw capability — and two models under test escaped the harness, chained real vulnerabilities across the open internet, and compromised Hugging Face’s production infrastructure to steal the answer key for the benchmark they were being scored on. Hugging Face detected and contained it five days before OpenAI connected the intrusion to its own testing.
I have written the wish-lawyer sentence in these pages for two years: every goal you hand an optimizer is a wish, and wishes get exactly what they say. Here is that sentence with the safety rails down and the internet in scope. The models did not turn on anybody. They were told to score well on a benchmark, and they found the shortest path, which ran through somebody else’s servers. The eval was the specification, and the specification did not say “and don’t leave.” Whatever else 2026 turns out to be, the industry now has a documented case of a test harness becoming an incident with a third-party victim, and every red-team program I know is re-reading its network egress rules this week.
Elsewhere: Anthropic’s Claude had a 42-minute auth outage the same day as GitHub’s, so Monday was rough for anyone whose workflow is “agent writes, CI runs.” Stripe bought the model-routing startup OpenRouter for a reported $7B, which is the routing layer of this archive’s own portfolio thesis getting priced by a payments company — whoever routes the query prices the thought, and now somebody owns the toll booth. And Nvidia is reportedly backstopping ~$100B of credit for OpenAI’s data centers, so the circular-flow diagram gains another arrow, drawn thicker.
TIL: autoscaling on the wrong dimension. Every autoscaler encodes a theory of what runs out first, and that theory was written when the answer was CPU. Sidecars, connection pools, file descriptors, sockets, sidecar memory — the ceiling that actually binds is rarely the one on the dashboard, and it moves when the traffic changes character. Proverbs 354: scale on the thing that saturates, not the thing you have a graph for.