Twenty-two commits now sit behind every merged pull request GitHub reports

Start with the rate rather than the total. 2.9 billion commits a month is about 1,120 a second ^[12]. The 130 million merged pull requests work out to roughly 50 a second ^[13]. Divide one by the other and the average merged pull request carries about 22 commits ^[8]. On the teams I have worked with, the pull request is where the human gate sits, so that ratio is the number I would stare at: one approval, 22 changes behind it.

Then look at the step that would actually falsify the claim each commit makes. The New Stack's framing is that generation has gone machine-paced while verification stays human-paced with a capacity curve close to flat ^[14], and the mechanism under that framing is specific: integration and end-to-end tests against a live system funnel through a shared staging environment or wait on a full copy of the stack ^[16]. Everything upstream of that step got faster. AI review triage, test selection, caching and better static analysis are real gains, and none of them runs the change ^[18]. So the arrival rate doubled in four months ^[3] while the one stage with a fixed server count did not, because full-stack duplicates cost enough that teams ration them ^[17]. A queue with fixed service capacity does not degrade in proportion to load; it degrades until utilisation pins near one, and the release valve then is people merging without the check.

Treat GitHub's telemetry the way you would treat someone else's benchmark table. For that curve to describe your repository, your commit mix would have to resemble the platform's, and the account reports aggregates without breaking out automated from human authorship ^[7]. 24 million new repositories a month ^[5] is a reminder that a lot of this volume never reaches anyone's review queue at all. What transfers is the shape rather than the height: commit volume tracked the developer population for years ^[19], and then it stopped.

The outage arithmetic is worth doing as well. 7 hours 47 minutes is 467 minutes, and a 99.9 percent annual target allows 525.6 ^[9]. At August's rate that window spans about 30 million commits of throughput ^[10]. Fedorov's postmortem did not hedge about it: if you were trying to ship software that day, GitHub let you down ^[2]. The answer was 3 million CPU cores, 120 petabytes of high-speed storage and an Azure migration now carrying 58 percent of load ^[4]. That amounts to a purchase order, the good kind of problem to have.

Extrapolate April to August and the curve implies about 20 percent compounding a month ^[11], which would put 2.9 billion near 26 billion a year out ^[20]. I do not believe that number and curves like this bend. The arithmetic is useful anyway, because nothing in a verification path bends upward on the same schedule. The local metric I would instrument is not commits per engineer. It is elapsed time from merge to verified against a running system, plus the count of changes waiting in front of the shared environment while that clock runs.