I'm using nvidia right now with a 3060. It doesnt use much power, I got it for pretty cheap on ebay, and it encodes/decodes everything except for av1 encoding which I dont have use for. Looking at the charts in the link below, if you need to encode av1 you'ld need a 4000 series.
I've found nvidia to work pretty well for jellyfin, I use docker with the nvidia container toolkit and it just worked with hardware encoding out of the gate. I also have some other docker containers running gen ai and the 3060 handles them well as long as the modle will fit in vram.
I think it entirely depends on your use case and hardware. I have a rack server, I need the extra power relatively frequently, as well as the 16x 2.5" bays and the 4 NICs. A rack server is a fairly power efficient package to get all those features in. However, it means that I am limited to discrete graphics, as Xeons don't have Intel QSV. There's also no monitor connected, and no 3D rendering happening, so the card is gonna idle at >5W and probably only use 20-30W while transcoding. Compared to a system that's idling at ~250W that's nothing.