To be fair, he doesn't really:

> And because it's a conservative heuristic we're underestimating the real number, it's probably going to be at least twice as much.

The actual measurement is 5%. The 10% figure is entirely made up, with zero evidence or reasoned argument except a hand-wavy "conservative".

Edit: actually, the claim is even less supported:

> out of these ~25000 crashes have been detected as having a potential bit-flip. That's one crash every twenty potentially caused by bad/flaky memory

"Potential" is a weasel word here. We don't see any of the actual methodology. For all we know, the real value could be 0.1% or 0.01%.

It depends on how the data are distributed.

I wouldn't be too surprised if that 5% all come from a few particular bad machine.