...and biased upwards.

That's also not the 2% number being used. 2% of total infections and/or total exposures would be the number that should be used for the type of "total deaths" projections the OP is seeking to use. Nobody knows the total infections number and we won't even have a very good estimated until the reliable antibody testing is conducted on a robust random sample.

For example, suppose we have tested some crazy low number of a total population like 0.09% (not 9%, 0.09% or less than one one hundredth of one percent, which is where we are in the US) and so, given the low testing availability, you only test people who are highly symptomatic, hospitalized, or dead. If you try to extrapolate from an extremely small sample of the total population with severe known selection bias, then you get numbers like the one's you stated. Plugging that death rate in and applying it to the total population would give you some ridiculous numbers that are so far out of line with the actual projections it's silly to even talk about.