You can't connect -a- transistor's service life to a
-population's- statistics directly. You can only get
a "confidence interval" (range in which the failure
will probably fall, the confidence level widens that
the more confidence you demand). Without peering
into the HTOL qual statistics (and hopefully this
includes ongoing product surveillance, not just the
first two lots out the pipe and believe nothing ever
changes subsequently) you won't know how much
wearout life varies lot-lot, quarter to quarter. And
it must - how much, is the question.
You also would not want to advertise the 125C life
without couching it in terms of a worst case, and
might like to also put out a "use model based" life
(how many systems truly spend their entire service
life at max rated temp? Usually there are swings in
ambient temp and power dissipation, and you need
a sophisticated analysis to get within a factor of 10
if your temp swing is 30C (/2 per 10C is a rule of
thumb, but who knows where that thumb has been?).
If you were a significant customer, putting this to the
vendor's applications engineering org with a request
for a reliability assessment of you r use case, might
get some attention. But not if you are trying to buy
cheapo commercial and apply outside the datasheet
limits, instead of paying the industrial or automotive
grade price.
I tend to see 90% or 95% in HiRel applications, but
that's for stuff that sees 100% burnin and lot level
HTOL sample qual. If you're buying floor sweepings
by the sackful, might want a 110% confidence. Heh.