This is not what he did.
Looks like you don't understand what NVLink did in this case.
In his case he had to use NVLink to link two of his GPUs together to increase the usable amount of vram because the 24 GB of VRAM wasn't enough for this scene.
NVLink made it possible to add the vram of two GPUs together to be used for that scene.
To make it a bit more simple, NVLink made it so that those two GPUs acted like one single big one for rendering purposes instead of two separate ones.
Edit:
To explain it a bit more.
If you have three GPUs with 24 GB VRAM each you might think that you have 72 GB of usable VRAM for rendering purposes but that's wrong.
You only have 24 GB because for each card to be able to work on the same scene the same data has to be inside the VRAM of each GPU meaning that if a scene takes up 12.4 GB of VRAM then all three of your GPUs have that 12.4 GB scene in their VRAM and the only way to circumvent that limitation in this case is NVLink.
Disadvantage is that if you only have three 3090s then only the two GPUs linked via NVLink can work on that scene.
You would need four 3090s and two NVLink bridges to be able to use all four GPUs for scenes needing more than 24 GB.
Edit2:
Here is a link to see how much VRAM is roughly needed for Gen 8 figures
You must be registered to see the links
(tl;dw 14 models 17.8 GB VRAM)
But keep in mind that the scene in the video is very simple so a more complex scene also means higher VRAM usage.