Sometimes, I feel like concepts such as these need to be translated into terms the average hackernews reader will understand.
Let's say your key metric you're trying to maintain or improve is P50 request latency. Its relatively easy to exhibit an acceptable P50 to 10 users, from 1 VM. Its actually harder to deliver the same P50 to 100 users, from 10 VMs. Even harder for 1000 users and 100 VMs, snd so on. You run into previously fine bottlenecks you didn't know existed. Issues which were invisible at ten users become cascading failures at thousands.
It just keeps getting harder the more users you add, even though you're adding proportionally more underlying VMs. There is no end to how difficult it gets, until eventually, it becomes impossible.
Testing is a problem with difficulty that scales with population size. Of the top ten countries by population size, there are only two countries that have higher testing per capita than the US: Russia and China. Those three countries are pretty close; no one else, anywhere in the top ten, is remotely close to the US, Russia, and China.
(Kind of funny how the top two testers in the top 10 countries by population are dictatorships, who have every reason to maintain a strong, powerful image. Funny how it works out like that, isn't it? Do you believe their numbers?)
I run the fastest website on the planet. No one but my mom visits it, but its still the fastest. Lithuania (the world's top daily tester per capita yesterday) is that website. The US is, uh, lets say Google. Should you bemoan Google for taking longer to load?