I'm somewhat familiar with the Sequoia AVC case in New Jersey, from a Usenix paper by Appel et al.:
We found two design flaws of the AVC Advantage which may cause inaccuracy in counting votes [...] Thus, a voter may mistakenly think she has voted, when she has not; or a voter may vote, and then be invited to vote again by a pollworker who mistakenly thinks her vote was not recorded.
They are also consistent with a 1% undervote observed in the one precinct in which we subpoenaed “voting authority” stubs. In precinct 6 in Pennsauken, NJ on February 5, 2009, there were 283 Democratic voting-authority stubs but the public counters of the 3 AVC Advantage machines added up to only 280, with 280 votes recorded.
The Sequoia machine was used more widely than that, but it seems the paper refrained from estimating the total number of people disenfranchised.
What are some large cases (in absolute numbers of voters disenfranchised) in the US due to problems with electronic voting machines? Of course, numbers can never be entirely certain in such cases, so estimates are likely to be reported. To prevent this question from becoming a conspiracy theory magnet, such estimates would have to be published by reasonably authoritative sources.
Best/accepted answer goes to the largest number reported in an incident, as long I judge the analysis to be reasonably authoritative and evidence supported.