Leap seconds are stupid, because they create discontinuities. Here's a nice picture: http://hpiers.obspm.fr/eop-pc/index.php?index=leapsecond&lan...
At home I hacked a OpenNTP based time server to hand out UT1 instead of UTC. Right now it's based on a simple software patch,
I yet have to a program to fetch the Bulletin B reports and parse them (ftp://hpiers.obspm.fr/iers/bul/bulb_new/bulletinb.dat). I don't know what they're smoking over there, but they're giving UT1 deviation relative to UTC instead of TAI. So first you've to get to UTC from TAI (respecting the leap seconds, gah), just so that you can go to UT1. Why not simply publish the value for UT1 - TAI instead?
Anyway everytime a report gets out, the clock rescaler in the UT1 timeserver gets adjusted apropriately (manually).
In the long run I plan to make this a small embedded project that uses GPS time to stabilize a high precision temperature compensated crystal oscillator^1 (TXCO), maybe I'll add a rubidum clock either, just because. And this is then to act as a stratum 1 time server on my local network.
----
[1] from my work I know that those have way less short term jitter than "naked" rubidium clocks; most precision frequency normals that come in a nice box with buttons and a display are actually TXCOs that are frequency stabilized by a rubidium clock.
I think it makes the most sense for computers to use TAI internally, and have UTC as a time zone on top of that. Whenever there's a leap second, just push out a TZ update. Most applications don't have special code for time zone updates, which means a lower likelihood of bugs.
Using TAI internally would have another advantage: Computers with GPS receivers could just add 19 seconds to GPS time to get a very accurate TAI. (GPS time is defined to be 19 seconds behind TAI. That was UTC in 1980.)