GitHub Adds 4 Million CPU Cores as AI Agents Strain Platform

GitHub has added four million CPU cores, 160 petabytes of high-speed storage and 50 terabits per second of network capacity to its infrastructure so far in 2026, according to figures the company has published. These are additional resources to handle growing activity, not the platform's total capacity, and GitHub lists AI agents among the main drivers of the pressure.

Activity has nearly tripled

According to Octoverse 2026, GitHub went from 12.9 billion commits a year to 34.3 billion in the twelve months ending in September 2026, counting public and private repositories. September alone saw 7.38 billion commits. A commit is a saved version of changes to a project and can range from tiny to large, so the figure measures recorded activity rather than lines of code generated.

Agents explain part of the jump. In September, more than 40% of pull requests started with a draft written by an AI agent that GitHub could detect, up from 17% in March. New repositories with agent configuration files logged on average eight times more commits than those without them. GitHub notes that the comparison does not prove an equivalent gain in productivity, since repositories without those files can also use AI.

Frequency, not just size

The strain comes less from storing code than from how often it is touched. A single agent task can involve multiple pushes, reads and automated checks, and many agents working in parallel multiply that traffic.

The limits have already shown. On August 17, 2026, GitHub suffered an outage lasting seven hours and 47 minutes. The company said traffic hit a new peak and a critical component in a US data center failed to scale, and it admitted it had not expanded certain resources early enough. It did not directly blame AI agents.

A new architecture

GitHub says more capacity alone will not be enough. It is migrating to infrastructure distributed across several Microsoft Azure regions and separating storage from processing so each can scale independently. It also aims to cut operations that must be coordinated with each other. In internal tests, the new design raised write capacity up to 35 times, though that does not reflect overall platform performance yet.

The episode shows a less obvious side effect of AI coding: generating proposals is easy, but developers still have to review, test and decide which changes to keep, while the services behind them must cope with increasingly automated traffic.