We are used to talking about artificial intelligence in terms of what it can do for us: write a function, fix a bug or help us finish an application. However, whenever that work translates into shared change, an infrastructure we rarely see. Versions must be recorded, requests must be addressed, and human and automated system tasks must often be coordinated at the same time. Extending agents is making the less visible part of development as worthy of attention as code.
Four million additional cores. A look at what GitHub has added to its infrastructure so far in 2026 shows the scale of that work According to information released by the platform, it has added four million CPU cores, 160 petabytes of high-speed storage and 50 terabits per second of network capacity. We’re not talking about four million physical processors or the total capacity of GitHub, but rather additional resources to meet growing operations. Companies place AI agents among the factors that explain this pressure.
An activity that has exploded. To understand why GitHub requires so many resources, we first need to look at how much work is flowing through its system. According to Octoverse 2026, the platform grew from recording 12.9 billion commits annually to 34.3 billion in the twelve months ending September 2026, including public and private repositories. Last month alone, GitHub recorded 7.38 billion commits. Each represents a saved version of a project’s changes, although it may contain changes of varying sizes. This is an important distinction: we’re talking about logged activity, not generated lines of code.
Agents change the rules. Part of this acceleration can be seen in how the work of programmers is changing. According to GitHub, over 40% in September pull requestProposals to incorporate changes into a project, starting with a draft prepared by an AI agent that can recognize the platform; In March they were 17%. The company also observed that new repositories with agent configuration files saw an average of eight times more commits than those without these files. This is a significant association, although the latter may use AI and the comparison alone does not show an equivalent improvement in productivity.
The problem isn’t just storage. Let’s think of an agent that accepts a task, reviews a repository, suggests changes and, after making new checks, revises it again. Even if the end result is a single improvement, syncing that work to GitHub can include multiple pushes, reads, and automated tests. If many agents work in parallel, the platform not only has to store more information, it has to transfer and process it faster. This is the least intuitive part of the phenomenon: the challenge is not just the size of the code, but the frequency with which it is processed.
The limits have already been mentioned. Increased activity also revealed capacity issues. On August 17, 2026, GitHub suffered a seven-hour, 47-minute outage that affected several of its services. As explained by the company, the incident began when traffic reached a new peak and a critical component of its data center in the United States failed to respond to demand. GitHub acknowledged that it didn’t expand certain resources early enough, though it didn’t directly attribute that drop to AI agents.
More power will not be enough. GitHub is changing the architecture that supports its repository in order to prevent the increase in activity from a permanent problem. The company is moving to a distributed infrastructure across multiple Microsoft Azure regions and is working to separate storage from compute so that both resources can scale independently. It tries to reduce activities that need to be coordinated with each other. In internal tests, the new architecture was able to multiply the write capacity by up to 35, although this result still does not represent the overall performance of the platform.
An infrastructure for another way of working. GitHub’s experience allows us to observe a less obvious consequence of the proliferation of artificial intelligence. Making code proposals easier doesn’t mean that all work can be automatically included in a project. Developers still need to review changes, test whether they work and decide which ones are worth keeping, while the services they use need to be able to handle increasingly automated activities. So transformation doesn’t end when an agent writes code: it also reaches systems that allow these changes to be transformed into functional and available software.
Featured Image: Markus Winkler
Engadget ｜ Training large models requires more hardware. China responded with a supercluster of 100,000 national cards
