AI Coding Agents Are Straining GitHub, and Microsoft Is Said to Be Calling AWS

The machines that write code have started to overwhelm the place that stores it. Through May and June, GitHub, the world's largest home for software and a Microsoft property since 2018, has been buckling under a surge of automated traffic that its own infrastructure was never sized for. The cause is not a hack or a botnet. It is the very AI coding agents that GitHub and its peers spent the past two years encouraging developers to adopt.

The scale of the shift is striking. AI agent pull requests, the proposed code changes that bots now file on their own, climbed from around 4 million in September 2025 to about 17 million by March 2026, a more than fourfold jump in six months. GitHub has been processing on the order of 275 million commits a week, putting 2026 on pace for roughly 14 billion, against about 1 billion for all of 2025. Automated builds tell the same story. GitHub Actions burned through 2.1 billion compute minutes in a single week early this year, up from 500 million in 2023.

When the platform starts to wobble

That load has had consequences users can feel. GitHub logged nine service-degrading incidents in May alone, and its availability has slipped below the 99.9 percent mark that enterprise contracts quietly assume. For a paying customer running a release pipeline, a few hours of downtime spread across a month is not an abstraction. It is shipped features that slip and engineers sitting idle.

According to Tech Times and Windows Central, Microsoft has responded by routing some GitHub traffic through Amazon Web Services, a rival cloud, to hold the platform together while a longer migration to Azure catches up. It is worth being precise about what is confirmed here. The AWS arrangement rests on reporting attributed to anonymous sources, and neither Microsoft nor GitHub has publicly acknowledged it. If it holds up, the optics are remarkable. Microsoft, which has spent years pulling its services onto its own cloud, would be leaning on Amazon's to keep a flagship running.

The bigger squeeze

GitHub's troubles are one visible symptom of a problem rippling across the industry. Demand for computing power is outrunning the world's ability to build and connect it. Goldman Sachs now projects cumulative global spending on AI infrastructure of around $7.6 trillion between 2026 and 2031, and reckons US data centers already face a power shortfall of more than 11 gigawatts, a gap it expects to widen past 40 gigawatts by 2028. Lead times for the GPUs that train and serve these models have stretched to somewhere between 36 and 52 weeks.

The shortage is reshaping who buys compute from whom. Google has agreed to pay SpaceX around $920 million a month for access to xAI's Colossus data centers, framing it as short-term bridge capacity for surging Gemini Enterprise demand. When one of the richest companies on earth has to rent a competitor's hardware to meet its own demand, the constraint is no longer theoretical.

There is a quiet irony running through all of it. The industry built tools that let software write itself, sold them hard, and is now finding that those tools generate work faster than the underlying plumbing can absorb. We have written before about the strain on developers as GitHub shifted Copilot to usage-based billing and about OpenAI's push to put coding agents in the cloud. The capacity crunch is the bill for that success arriving all at once.