When AI Coding Agents Overwhelmed GitHub’s Infrastructure Capacity
GitHub’s AI capacity crisis is a breakdown in which explosive demand from AI coding agents outpaced GitHub infrastructure capacity forecasts, causing repeated outages and forcing Microsoft to rent extra cloud resources from its main rival Amazon Web Services to keep the developer platform online. What began as a standard cloud migration to Azure turned into an emergency capacity scramble once AI coding tools changed how much code developers produce and how often agents touch repositories. Kyle Daigle, GitHub’s COO, wrote that GitHub commits were on pace to reach 14 billion in 2026, up from 1 billion in 2025, a shift that no traditional growth curve could explain or absorb. That spike translated into GitHub Actions queues backing up, repositories lagging, and routine workflows failing, exposing how poorly suited older planning models were for agentic workloads.
Why Microsoft Turned to AWS Despite Owning Azure
Microsoft’s decision to add GitHub capacity on AWS highlights a blunt operational reality: uptime mattered more than cloud rivalry. GitHub historically ran its own data centers, with a plan to complete migration to Azure by 2027, but outages in 2026 showed that schedule no longer matched AI coding agent demand. According to Business Insider, Microsoft is now “adding extra computing capacity via Amazon Web Services” after repeated AI-driven disruptions. A Microsoft spokesperson confirmed that GitHub is using multiple cloud providers to secure “future capacity, compute elasticity and horizontal scale required to support continued growth.” In practice, that means Azure remains the strategic home, while AWS acts as overflow capacity when agentic workloads spike beyond what current Azure and legacy infrastructure can handle. The servers follow demand curves, not corporate slide decks, and GitHub’s users care more about push and CI reliability than cloud brand loyalty.
Agentic Workflows and the GitHub Actions Scaling Squeeze
The main pressure point is not classic Copilot autocomplete but agentic workflows running through GitHub Actions and related services. A single AI coding agent session can read large repositories, plan changes, modify code, run tests, and open pull requests in loops, multiplying the load on compute, storage, and networking. GitHub’s product direction has accelerated this shift: Agent HQ for Copilot Pro Plus and Enterprise plugged agents like Anthropic’s Claude and OpenAI’s Codex directly into GitHub, GitHub Mobile, and Visual Studio Code. Each agentic workflow turns what used to be a few manual commits into many automated operations across CI pipelines, issue tracking, and pull request checks. When thousands of teams run such flows in parallel, GitHub Actions scaling becomes a systemic bottleneck. The result has been dozens of major outages in 2026 and month-long stretches where degraded performance slowed repositories and build pipelines across the platform.
Unpredictable AI Demand and the Risks of Moving Too Fast
GitHub’s strain is part of a wider pattern: AI workloads grow in bursts that capacity planners struggle to predict. SpaceX and Google disclosed a deal in which Google will pay SpaceX USD 920 million (approx. RM4.24 billion) a month for AI compute capacity from October 2026 to June 2029, even as Google Cloud sells AI capacity to Anthropic. These cross-cloud deals show how fast AI usage can outrun original data center plans. GitHub’s own experience displays the downside of pairing rapid agent feature launches with infrastructure that assumed human-paced usage. Developers felt this directly; HashiCorp cofounder Mitchell Hashimoto said GitHub was “no longer a place for serious work if it just blocks you out for hours per day, every day.” The tension is clear: shipping new AI features draws users in, but without resilient GitHub infrastructure capacity behind them, cloud infrastructure strain becomes the story.
From Outages to Usage-Based Billing: GitHub’s Next Phase
Infrastructure pressure is reshaping GitHub’s business model as well as its cloud layout. On April 27, Chief Product Officer Mario Rodriguez announced that all Copilot plans would move to usage-based billing from June 1, 2026, replacing premium request units with GitHub AI Credits tied to token consumption. Base prices for Copilot Pro, Business, and Enterprise stayed the same, but the message was clear: long, multi-step agentic sessions have different costs than quick chat prompts. Some users shared projected bills hundreds of dollars above previous months, with one Reddit user seeing an estimated jump from a USD 39 (approx. RM180) plan to USD 847 (approx. RM3,904). That kind of shock will push teams to cap spending, tune model settings, or consider alternatives like Cursor, Claude Code, or Amazon Q Developer. GitHub must now prove that its new Microsoft AWS partnership and Azure push can keep AI coding agent demand on a stable, predictable footing.
