I would humbly submit that MTBF is not the best figure of merit one can cite or consider as a "measure" of reliability. It has already been noted repeatedly that this is a deceptive figure. Who, for example, measured the 2.6 million years that these power supplies lasted (actually, if 2.6 million is the mean, then the measurements would have more likely lasted 5.2 million years!?! Really!?)? It's obviously erroneous.

In the IPC Designers Council standards, there are three broad levels of reliability. Consumer electronics are at the lowest level of reliability. They need to work well and outlast their obsolescense. But, if they do fail, there is no risk of life or limb.

The intermediate level of reliability is one where the products must perform in order to protect life and limb. Medical devices, automotive, military, these are typical examples.

But the highest level of reliability is an application wherein the device absolutely must work, first time every time, when it is activated. Examples of this might include the parachute deployment system on the old Apollo command modules. Or the telemetry systems on a Mars lander.

I would suggest that a far more informative, and less deceptive, measure of reliability for a given design is the infant mortality rate. As someone has already pointed out, it's pretty difficult to get any electronic device to fail if it works to begin with (and for convenience sake I will naively ignore the possibility of such things as memory leaks and lost pointers in the firm/software).

There's my two cents worth.