It's an architectural let down. These cards should have a direct bus bridge that makes the cards parallel process like chiplets. The reason why SLI should go away is to optimize parallel processing instead of getting rid of it. This shows a lack of vision by Nvidia. If Nvidia wanted to they could design their architecture to properly linearly scale using a common bus across all cards. Problem is that was what SLI was supposed to be but they never achieved it even in the NVLink iterations. This is passing the buck to the API developers to duct tape parallel processing together in software, instead of hardware. Software solutions are almost never better than hardware solutions because of driver and API level bugs given whatever version being used. Gotta realize we're passing the buck to Microsoft to make parallel processing GPUs work and they screw with the render pipelines every 6 months now with Windows 10 updates (often for no apparent reason other than random graphic user interface tweaks).
Instead of championing features, they're abandoning them. The original dumped was 3D Vision. Also VitualLink was quietly dumped with a total lack of commentary. They left both dead awhile back. I expect G-Sync will be next abandoned in favor of the open Adaptive V-Sync (VESA's FreeSync specification).
Just imagine those people going 3090 to get SLI to conflate their benchmark scores just to find their second $1500 card will be a paper weight next year.