I'm not sure if I'm interpreting the paper correctly, but I don't think those numbers quite match what the authors found. I find their presentation confusing, though, so possibly I'm the one that's wrong.
I think you understand this but in case it's not clear to others (I wasn't sure until I got to the end of the paper) all of the machines in their study are using ECC memory. This allows them to measure both the number of Correctable Errors (single bit) and Uncorrectable Errors (more than single).
I think you are right that the chances of a single bit flip for a randomly chosen machine are about 25%. In Table 1, they say CE incidence for a randomly chosen machine is 32% per year. But they also show that a machine that has one correctable memory error is much more likely to exhibit many more.
I think that "CE Median Affct" (also in Table 1) says that among the machines that have been affected by at least one correctable error, the mean number of errors is 277. Thus rather than one crash per 10 machines over 3 years, it's likely that you'll get hundreds of crashes, but with one (or zero, or two) of those 10 machines responsible for all of them.
The Uncorrectable Error rate they found was about 1/20 of the correctable rate, a little over 1% per year. Assuming the same "use time" factor applies, this would mean that with ECC none of the 10 machines would be likely to ever crash due to bit flips, reducing the number of crashes by hundreds.
But this doesn't tell the full story either. Maybe it's best to view it as 1000 employees, each of whom is issued a laptop. If you buy them machines without ECC, the initial cost is less but 100 of them will experience frequent crashes during the lifetime of the laptop, leading to frustration, possible lost work, and early machine replacement.
If you pay more up front for ECC, the number of affected employees will be reduced from 100 to 5. How much extra one should be willing to pay for this reduction depends on circumstances, but I think the impact is quite different than merely avoiding 100 equally distributed random crashes over the course of 3 years.