Commits on GitHub have doubled in four months. Verification capacity has not.

GitHub has seen a doubling of commits over the past four months, with the volume now at 2.9 billion per month. This growth broke the platform, leading to a 7-hour and 47-minute outage on August 17 after a core infrastructure component failed to scale with the increased traffic. To address the issue, GitHub increased its computing resources by over 3 million CPU cores, 120 petabytes of high-speed storage, and migrated to Azure, which now serves 58% of the platform's load.

The critical issue, however, is not just capacity but verification. Every commit carries an implicit claim that the change works, but most of the systems involved in producing, transporting, and merging the code did not check that claim against a running system. Verification, the process of proving a change does what was intended without breaking what already worked, is still human-paced, and its capacity curve is close to flat.

This gap between the rapid code generation and slow verification is the defining infrastructure problem for the next three years.

AI-generated code has become machine-paced, and its volume is growing exponentially. Verification, on the other hand, has not kept pace, and thus, the distance between these two curves represents a significant infrastructure problem. The commit curve, which is now public, is a proxy for the development activity of thousands of engineering teams. The surge in commit volume is not due to the growth of the developer population, but rather the emergence of agent-generated code.

While GitHub's problem is primarily capacity-related, the internal issue for organizations is confidence. The verification process, which is crucial for ensuring the correctness of changes, has not kept up with the pace of code generation. Verification was designed for the human pace of development, and the rise of agents has overwhelmed its capacity. The two curves that previously tracked each other have now diverged, and the gap between them is filled with unverified merges.

Teams are responding to this widening gap in various ways. Some are trying to speed up the review process using AI code review tools, which can catch real defects in diffs but cannot guarantee the behavior of the change against its dependencies. Others are throttling the agents to limit the amount of generated work. However, these coping strategies are not sufficient, as the bottleneck lies in the verification process, which has not evolved to keep pace with the rapid generation of code.