The "bends" in the curve they plot are too abrupt. There must be something else going on.
Looking at the original article, they had only about 3500 drives around 2009. That's 4 years ago. So their "4 year" survival rate is not based on the 25000 drives they have now, but only on the 3500 that they had in 2009. With the sharp bends in the curves around 1.5 years and 3 years, I think they significantly changed their buying policy around those moments. Or the manufacturers started shipping them different drives.
How else can the drive "know" that it's been on for 1.5 years? The annual failure rate drops by a factor of four inside a month.
The explanation of the bathtub curve eplains it a bit, the random failures is apparently about 1.4% per year. The initial failure is about 5.1-1.4= 3.7 per year. But instead of the initial failures "tapering off" to "small" values around 1.5 years, they stay constant for 1.5 years, and then suddenly drop to zero. To me this points to something like: "they bought a big batch of drives about 1.5 years ago that has such a high random-failure-rate to pull the average first-1.5-year average up to 5.1%/year".
Do the same analysis 3 months from now, and the "1.5 year bend" moves over to 1.75 years. That's my hypothesis based on the data they publish. Having the underlying data and some time to spare, the current data may debunk or prove my hypothesis already. (e.g. if you run the analysis on the data that is now older than 3 months will, if my hypothesis is correct, show the bend around 1.25 years. If that happens, it makes my hypothesis very likely.....)