AI Did Not Remove the Bottleneck It Moved It

The capacity warning

In April 2026, developers pushed 1.4 billion commits to GitHub in a single month. By August: 2.9 billion.

Then, on August 17th, GitHub went down for nearly eight hours. No bad deploy. No configuration change.

Their CTO wrote: “Our existing operational practices did not keep up.”

GitHub doesn’t say what doubled the volume. I think you can guess.

AI did not remove the bottleneck in software delivery. It’s moved it.

Generation outruns review

For decades, writing the change was the slow part.

Now an agent reads the repository, edits a dozen files, writes the tests, and opens the pull request while you are still in the meeting.

Google says 75 percent of its new code is AI generated, then approved by engineers.

Anthropic says its engineers merge eight times as much code per day as they did in 2024. And in the same essay: “human code review has become a new bottleneck.”

They call it a signature of Amdahl’s law. Speed up one part of a system, and the part you left alone sets your limits.

Every change is a claim

Here’s the part we left alone. Every generated change is a claim.

It claims to match the intent. It claims the tests mean something. It claims nothing downstream will break.

Generation multiplies the claims. It doesn’t settle a single one.

So the queue comes back: in review, in QA, in the incident channel.

One vendor’s telemetry, across 22,000 developers: as AI adoption rose, throughput went up by a third. Time in review tripled.

Justified confidence

Ask an engineering leader what’s hard right now, and nobody says, “We can’t write code fast enough.” They say, “I can’t tell what’s safe to merge.”

That’s the new constraint. Not the production of software. The production of justified confidence.

We spent twenty years optimizing the left side of delivery: write, build, integrate, deploy.

Now the right side has to catch up: understand the impact, verify the outcome, authorize the action, prove what happened.

So on Monday, don’t ask how much code your agents can write. Ask how much authority you can safely hand them. That number is your real velocity.

Generation is getting cheaper. Confidence is becoming the constraint.

That’s the Reliability Thesis.