According to a 2016 blog post by Backblaze, a (particular) 1.5TB HDD uses 3.4W when idle and 5.9W when operating. This gives a lower bound of 20 kWh / TB / year and an upper bound of 35 kWh / TB / year. These numbers are much closer to the low figure than the high one given in the question. We can also see that larger drives use proportionally less power and large-scale cloud storage providers will tend to move towards larger drives (for this and other reasons). Using a 6TB HDD instead results in power usage of 7.2W when idle and 8.8W when operating. This gives a lower bound of 10 kWh / TB / year and an upper bound of 13 kWh / TB / year.
The referenced paper arrives at its number by assuming the drive portion of the power cost is 4.9 kW / 604 TB (unclear what the reference for this is) or around 8 W / TB. This is right in the ballpark of the numbers given by the Backblaze blog post.
The paper then goes on to multiply this power usage by various factors - eg by 1.5 for account for cooling, by 2 for assumed redundancy, and by 2 again for assumed waste due to under-utilization. It then adds in other power factors, routing hardware, all intermediate servers between client and server, etc.
These seem like real contributors to power requirements but the total seems overstated to me. For example, the average 9 servers between client and server necessary to route traffic between them would probably mostly be up and running regardless of that TB of cloud storage. Their cost should be spread across the wide range of services they are providing (mostly Netflix streams at this point, probably).
It is also necessary to compare apples to apples. The paper talks about storage with some form of redundancy in it - assuming that storing 1TB of data requires 2TB worth of disk. The other number in the question, 41 kWh / TB / year, may be for no redundancy or some other level of redundancy. The figures I've computed above are for exactly 1TB - not 2TB.
Properly allocating the power used by all of these shared services is probably a research project all on its own. The authors of this paper have selected values which make Cubbit (an alternative distributed storage system) look like an improvement but I don't see evidence in the paper to support all of the assumptions they've made.
So, I have found no research that provides a complete answer to this question. My personal guess is that usage for a single non-redundant TB is between 2x and 10x the drive requirement - between 20 kWh and 100 kWh / TB / year. It would be great to see a more thorough analysis (or measurement!) spanning the full range of infrastructure, though.