[a password-cracking semi-necromancer appears]
Sounds like the OP already understands the differences with cracking workloads - which are, indeed, different from mining workloads. I'm not a hardware expert, and I haven't had to solve 30xx power constraints, but I can describe what has worked for me.
I've been running hashcat (and John the Ripper OpenCL as well) for a few years in a single open-air rig with 6x 1080s and 2 Corsair RM1000 PSUs and a 24-pin splitter. For whatever reason, the power aspects have been stable for me, but that may just be luck. I overbuilt it on purpose, and never get anywhere near maxing the PSUs out. The powered ribbon risers can sometimes be a little flaky every few months, but it's not that bad - reseating them usually resolves it. Some wordlist-heavy workloads are indeed slowed down by x1 PCIe, so if you can swing a minimum of x4, that's a good idea. And unlike most mining, you want to have as much system RAM as GPU memory, because hashcat reasons. I also recommend at least one core per GPU, with a little left over in case you want to be doing other kinds of wordlist processing while you have cracking jobs running. And even if the CPU is idle for your current attack, you can always throw hashcat CPU (or john, or even MDXfind) at those cores while hunting for other patterns to exploit. I've undersized CPUs on cracking rigs in the past - and always come to regret it.
Note also that you can throttle the cards to, say, 80 or 90% of their max power, and still have quite decent hashcat performance. atom (primary hashcat dev) did a study of this a while back and there is a sweet spot you can probably find with some testing. If ambient temperatures are stable, you can also lock the fans to 80% or something like that to combat fan-speed variability.
All that being said, you also have an entirely different option: just build two (or more) rigs, and team them with Hashtopolis. There are a few attacks that are trickier to formulate, but for broad scale-out, it's hard to beat for many of the standard attack stacks. And you get to entirely sidestep the "lots of GPUs in one rig" problem. It also makes it easy to temporarily scale out fast (via Vast.ai, etc.)