GitHub Plans 30X Capacity Expansion as AI-Driven Commits Overwhelm the Platform

A Platform Under Pressure

GitHub is grappling with severe infrastructure strain as AI-driven development reshapes how code gets written. In a blog post on 27 April, CTO Vladimir Fedorov disclosed that the world's largest code-hosting platform now needs roughly 30 times its current capacity to keep up with demand. The numbers are staggering: GitHub is processing 275 million commits per week, on track for around 14 billion commits in 2026, a fourteen-fold jump from the 1 billion logged across all of 2025. Weekly compute minutes on GitHub Actions have already doubled to 2.1 billion. Pull requests opened by AI agents leapt from about 4 million in September 2025 to more than 17 million by March 2026.

Outages Pile Up

The growing pains are obvious. GitHub suffered eight major outages in February and March 2026 alone, missing the 99.9% availability promised to enterprise customers. On 23 April a merge queue regression hit 658 repositories and 2,092 pull requests. On 27 April an overloaded Elasticsearch cluster, likely targeted by a botnet, disrupted search across the platform, and another degradation appeared on 29 April. Reliability engineer Lorin Hochstein, formerly of Netflix and AWS, identified a core weakness: GitHub's responders lack granular controls to selectively shed load during crises, forcing a binary choice between blocking everything or watching the system buckle.

Scaling for the AI Era

GitHub's response is sweeping. Engineers are migrating webhooks off MySQL, redesigning session caches, rewriting authentication flows in Go instead of Ruby, and pushing toward a multi-cloud architecture beyond its Azure footprint. Critical services like Git and Actions are being isolated from other workloads, and merge queues are being tuned for repositories handling thousands of pull requests daily. Separately, on 27 April GitHub announced that Copilot will shift from request-based to usage-based billing on 1 June, conceding that the current model is unsustainable as inference costs balloon. "Our priorities are clear: availability first, then capacity, then new features," Fedorov wrote, signalling a fundamental rethink of how the platform supports an AI-driven era.