In some fairly crude testing on a test file of 500,000 randomly generated lines (18MB) to fit the pattern of data you described, this seems to run about 400% quicker. 40 seconds rather 160.

no warnings;
$^W=0;

open my $in, $ARGV[0] or die "Couldn't open $ARGV[0]:$!";

my ($buffer, %h) = ''; keys %h = 1024*500;

while (sysread($in, $buffer, 16384, length $buffer)) {
    $h{$1}++ while $buffer =~ m[^(?:.+?\|){9}([^|]+)\|]mg;
    $buffer = substr($buffer, rindex($buffer, "\n"));
}

print scalar keys %h;
[download]

You may need to adjust the regex, and you might squeeze some more out by playing with the read size and/or the hash preallocation. In my tests, I had variable results from both, but the differences were of within the bounds of run-to-run error. Especially the latter which I am not really sure how the number of buckets relates to keys.