There have been some decent results historically with checkerboard and other separated reconstruction techniques. Nvidia was working on some new checkerboard approaches before they killed off SLI.
A decade or two ago most people I knew had dual GPUs, itbwas quite common for gamers and while you were not getting 100x utilisation it was enough to be noticeable and the prices were not mega bucks back then.
On your point of buying 1 card vs many, I get that, but it seems like we are reaching some limitations with monolithic dies. Shrinking chips seems to be getting far harder and costly, so to keep performance moving forward we are now jacking up the power throughput etc.
Anyway the point I'm trying to make is that it's going to become so costly to keep getting these more powerful monolithic gpus and their power requirements will keep going up, so if it's 2 mid range gpus for $500 each or 1 high end gpu for $1.5k with possibly higher power usage im not sure if it will be as much of a shoe in as you say.
Also if multi chiplet designs are already having to solve the problem of multiple gpu cores communicating and acting like one big one, maybe some of that R&D could benefit high level multi gpu setups.
If these were stories I was picking up to implement I would be asking the BA to elaborate some more 😂