That part is tricky, because Proof of Capacity (also called space and time) can be faked (also, it is not the first coin with this consensus, there were other failed attempts before, for example Burstcoin). There were fake plots on Chia. Like someone said on Discord, it is like selling a car and assembling it at the same time, when you are trying to pay for it. If you have 1 TB plotted hard drive, then you can play by the rules and get some coins, equivalent to this 1 TB. But if you can plot things in memory, and then save only every second piece, then you need only 500 GB, you will still have some chance to hit something, and you could recompute things when needed. In general, as in ECDSA: you don't need real signature, you need a signature that can pass verification. The same with Chia: you don't need to plot 1 TB, you need to produce a proof that you did it, even if you plotted only a part of that and splitted your work between plotting and on-the-fly computation. So, I guess it could be possible to mount an attack that would plot nothing and focus on mining. For well-designed ASIC, that could be more profitable than buying large disks and plotting everything, also because your read/write to disk also takes time, when RAM operations or some assembly code based on registers could do some parts much faster, so there is no point in storing them permanently on your disk (also because you only need to verify things, not to plot exabytes by yourself to check that it was plotted correctly).

I remember some discussions, where people reached nice results for yespower on FPGAs. And also I remember that some people complained about yespower in early days, even before the coin was released. They were saying that this algorithm can be optimized for non-CPUs, and judging by historical blocks and network difficulty, I guess at least some of them were right (or they had a lot of CPUs, definitely warehouse-level, not home-miner-level).