It is from the ratio of people with antibodies to viral PCR positive tests in selected states/cities, where that ratio was about 6-7 actual infections to positive tests. This ratio has been pretty solidly in the 5-10X range since the first antibody studies were done in NY and Spain. So if there were 14MM PCR positives at the end of Nov, 91MM total infections (6.5X) makes sense. There's uncertainty in that estimate, of course, which is why they also provide the 95% confidence interval of 78-107MM total infections.