GitHub Is Running on AWS Now. Here’s Why AI Coding Broke the Platform.

GitHub's infrastructure buckled under a surge of AI-generated code, with commits on pace to hit 14 billion in 2026, up from 1 billion in 2025. The platform suffered 19 service-degrading incidents in April and May, and Microsoft-owned GitHub quietly added AWS capacity because Azure alone could not keep up. AI agents like Claude Code now drive the majority of traffic, overwhelming a system designed for human developers.

GitHub commits are on pace to hit 14 billion this year. In 2025, the number was 1 billion. That is not organic growth — it is AI agents committing code around the clock, and it has pushed the platform built for human developers to its architectural limits. The fallout: 19 service-degrading incidents across April and May, agent session wait times spiking to 54 minutes, and the detail that says more than any official statement — Microsoft-owned GitHub quietly adding Amazon Web Services capacity because Azure alone could not keep up. The Numbers Behind the Meltdown The scale of what happened to GitHub in 2026 is genuinely hard to absorb. Claude Code alone accounts for 2.6 million commits per week — up from roughly 100,000 in September 2025, a 25x increase in six months. That single tool now represents 4.5% of all public commits on the platform. Add every other agent, Copilot session, and automated workflow, and the pattern is clear: AI agents are now the primary driver of GitHub traffic, not humans. GitHub Actions minutes tell the same story. Weekly usage went from 500 million minutes in 2023 to 2.1 billion by early 2026. AI agent pull requests grew from 4 million per month in September 2025 to 17 million by March 2026 — a 325% increase in six months. GitHub’s infrastructure team had planned for a 10x capacity expansion back in October 2025. By February 2026, they realized they needed 30x. The platform’s own SVP, Jakub Oleksy, put it plainly: “We’re now serving 40 percent of monolith traffic from Azure, up from 8 percent in February” — a migration sprint prompted by a crisis, not a plan. The AWS Plot Twist Here is the detail that cuts through the messaging. Microsoft — the parent company of GitHub, the owner of Azure, the cloud platform that competes directly with AWS — had to contract Amazon Web Services capacity to keep GitHub functional. The explanation given is “multi-cloud resilience,” and that framing is not wrong, but it obscures the real story: Azure could not absorb the load fast enough on its own. GitHub has been racing to migrate its infrastructure to Azure since Microsoft’s acquisition. That migration is real and ongoing. But the AI surge outpaced the migration timeline. When your own cloud cannot serve your own platform, you go to your competitor’s cloud. That is what happened here. The Register’s reporting on the outage timeline https://www.theregister.com/software/2026/06/12/github-outages-persist-as-ai-coding-drives-traffic-surge/5255125 and the Business Insider investigation into the AWS contracts https://www.investing.com/news/stock-market-news/microsoft-taps-amazon-to-ease-github-aidriven-strains--business-insider-4743699 document how this unfolded. What the Outage Data Actually Shows GitHub’s official status page reports approximately 99.9% uptime. The unofficial “Missing GitHub Status Page” project, which tracks real user-facing degradation, put the 90-day figure at 87.26%, with monthly availability ranging between 78% and 94%. The gap between those numbers reflects how GitHub defines “incident” — slow CI, elevated error rates, and partial degradation often do not appear in official counts. April was the worst month: ten incidents, including 84% session start failure rates that peaked at 97.5%, and agent wait times that hit 54 minutes. May had nine incidents. The most dramatic failures cascaded — one outage took code search down for 8.7 hours, degraded Copilot’s cloud agent for four hours, and knocked the Copilot backend offline for 2.7 hours simultaneously. Why AI Traffic Breaks Platforms Designed for Humans The core problem is not volume — it is pattern. A human developer commits a few times a day, opens a PR, waits for CI, then responds. That is a sequential, low-frequency workflow. An AI agent operates continuously, commits hundreds of times per session, and triggers downstream processes the moment each change lands. One AI agent PR does not create one unit of compute load. It spawns a CI pipeline, which triggers webhooks, which activate code review bots, which generate Copilot suggestions, which create more commits. The cascade multiplies each action into dozens of downstream events. Open source maintainer Xavier Portilla Edo estimated that only 1 in 10 AI-generated PRs is legitimate, meaning the human review burden exploded even as human commit rates stayed flat. The technical analysis of this cascade effect https://zenvanriel.com/ai-engineer-blog/github-ai-agent-commits-infrastructure-crisis/ is worth reading to understand why GitHub’s architecture is strained in this specific way. What GitHub Is Doing — and What You Should Do Now GitHub has declared its operational priority order: availability first, then capacity, then new features. Concretely, that means isolating Git and GitHub Actions from other workloads, accelerating the Azure migration, and leaning on AWS for overflow. Usage-based billing, live since June 1, is partly an infrastructure play — if agentic sessions cost real money per token, developers will be more selective about when to run them. The individual plan sign-up pause https://github.blog/news-insights/company-news/changes-to-github-copilot-individual-plans/ is part of the same picture: GitHub stopped accepting new Pro, Pro+, and Student subscribers while it rebalances capacity for existing users. In the meantime, there are things your team can control: Schedule CI during off-peak hours. Early morning UTC sees significantly lower queue times. Run local validation before pushing. Pre-commit hooks and local test runs reduce unnecessary CI triggers. Batch commits strategically. High-frequency micro-commits from AI agents are a primary load driver — add checkpoints before auto-push. Monitor the unofficial status page rather than GitHub’s official one for a more accurate picture of real availability. Add human-in-the-loop gates for AI agents. Do not let agents auto-commit without a review checkpoint. The Harder Question GitHub’s infrastructure response — more Azure, AWS overflow, usage-based pricing — addresses the immediate crisis. It does not address the structural question underneath: was the branch-PR-CI-webhook model designed for a world where agents commit thousands of times per hour? Billing changes can moderate demand at the margins. But if agent-scale development is the new baseline, GitHub may need to rethink the collaboration model itself, not just the data centers behind it. For now, expect more incidents, budget for CI delays, and maybe appreciate the irony: the AI tools making you more productive are also making the platform they run on measurably less reliable.