GitHub Added 4 Million CPU Cores to Absorb AI-Generated Code Flood

GitHub added 4 million CPU cores, 160 petabytes of storage and 50 terabits per second of networking capacity this year to absorb a flood of AI-generated code. The company's own technical report puts a hard number on what developers have suspected for months: machine-written code is now a measurable infrastructure problem.

Monthly code commits grew sixfold year over year, reaching 7.4 billion in September. Merged pull requests, which finalize a feature or bugfix, jumped fourfold. The growth pattern matters more than the multipliers. GitHub counted 1.4 billion commits per month in April. That figure doubled to 2.9 billion by August, then nearly tripled again to 7.4 billion in September. The company logged 17.9 billion code pushes over the past 12 months — more than all of 2024 and 2025 combined.

GitHub attributes the surge to agentic AI. The report states that by September, over 40 percent of pull requests were initiated by an agent, up from 17 percent in March. Repositories that begin with AI show roughly eight times the activity of repositories built by hand. These are GitHub's own figures. No independent verification exists.

What does the code flood do to infrastructure?

Every commit and every merged pull request triggers a chain of automated work in modern development pipelines. The report lists the usual suspects: checking for leftover to-do items, linting code for stylistic consistency, and running code and integration tests to confirm new changes do not break existing functionality or interfaces with other programs, services and APIs.

Multiply that chain by 7.4 billion monthly commits, and the hardware bill writes itself. The strain falls hardest on CI/CD pipelines, the automated systems that build and test every change. GitHub says the load grew so quickly that it reengineered its backend to decouple storage from compute as much as possible. That is an admission of architectural stress, not a routine optimization.

How much of this code is useful?

GitHub's report measures volume, not value. Commit counts cannot distinguish a one-line fix from a thousand lines of machine-generated boilerplate, and the company offers no breakdown of how much agent-written code survives review or reaches production. An eightfold activity multiplier for AI-seeded repositories sounds impressive, but activity is not output. More pushes mean more test runs, more storage and more networking — GitHub earns the cost either way.

The report is also silent on failure rates. If agents open 40 percent of pull requests, the natural question is how many of those merge successfully. GitHub does not say. Until it does, the sixfold commit figure says more about infrastructure demand than about software quality.

What comes next?

The trajectory points one direction. Six months took commits from 1.4 billion to 7.4 billion per month, and the report's own framing — the need to both upgrade and rethink infrastructure — signals that the current build-out buys limited headroom. Given how fast demand is rising, GitHub is likely to have to upgrade again sooner rather than later, and the next round of capacity additions will land on an even larger base.