I can think of 2 reasons to use higher sampling rates. Neither is mentioned by that article, which focuses on irrelevant points. Indeed, his argument applies to analog recording too, since it also is bandwidth limited. In many cases, the microphones are the bandwidth limit, not the recording medium. However we record, it's always bandwidth limited. What is the point of recording at 192 kHz to get Nyquist to 96 k, if the mic itself doesn't capture anything above 30 k? Another important point he fails to mention is the energy or amplitude of those frequences around 20 kHz is very small (with most sounds anyway) and often masked by the much louder sounds at lower frequencies. So if they get attenuated, perceptually, most of the time, it doesn't make a big difference.
So why use higher sampling rates?
First, because they allow a gentler anti-aliasing filter. At 44.1 k sampling, the transition band is only 20,000 to 22,050. That's a narrow range for a filter to go from 0 to -infinity, hard to do in real-time with limited hardware without passband artifacts. Higher sampling rates give a wider transition band, gentler slope, which is easier to implement. Evidence of this is how common it is for DACs measured here to have filters that extend the transition band to 24 k, which is above Nyquist at 44.1 k sampling. They wouldn't cheat like this unless they had a reason... If we used a sampling rate even just a little higher, like 48 k, they wouldn't have to cheat.
Second because some people, with some audio signals, under ideal conditions, can detect the difference between 44.1k and higher sampling rates. A while back, Amir posted one such ABX test that he did. I believe the test signal was jangling keys in front of a mic? It was either that or some other signal that had a lot of extreme HF content. Another study has been referenced here too, showing a small but not insignificant % of people detected signals in the range of 20k and slightly higher.
In suggesting that higher sampling rates have some benefit, I'm not saying we need to go crazy. 48 k would likely be sufficient. 88.2 k would be more than enough.