(Which, yes, has implications for energy use/climate change too for sure).
It doesn't look like i currently have access to the usage data on any of the lots-of-runners-lots-of-PRs projects I currently work on (which are still probably way less than some large companies).
Any "large companies" don't give a shit about things at this cost level. They spend more on the time it takes you to open the door. The number of CI minutes could be astronomical and it still wouldn't rate above the threshold of caring. The time people in this thread have spent wringing their hands is way more expensive.
On my larger organization, we have on average 20 to 30 *active* runners during business hours. Assuming 5 on the off-hours, my napkin math says it comes down to about 10 fully-utilized-runners per month, so about 864$/mo. For the size of my organization that is honestly totally acceptable.
This is assuming 0.002$ per minute of job being actively executed. If it turns out to be 0.002$ per minute of *runner being registered* on the control plane, it would increase quite a bit. We are still using the old HorizontalRunnerAutoscaler with actions-runner-controller, with quite a pool of prewarmed runners idling to pick up a job. It would be a strong reason to use the new RunnerScaleSet (to take advantage of the reactive webhook-based scaling) and keep a very lean pool of prewarmed runners.
I get the logic of it, they have to have some sort of task running on their side when the runner is working. If it's only build time, then we don't care.