Do you know whether it' hardware limitation or limitation of IPFS software which isn't designed to host hundred TB data? If it's hardware limitation, using SSD would solve the problem, just like how you run ETH full node.
With the amount of data we're talking I believed he used hard disk RAID (5/6/10?) so not only are SSDs economically infeasible here but with the networking lag I described ("crosstalk", as in someone else's IPFS site traffic being passed through you), I get the impression it's not designed for that volume of data.
The IPFS service in question was even sparely used.