Somehow I wish it weren't true because I hate monopolies in any markets. Nvidia has been enjoying their absurd premium on consumer and production GPUs and intentionally removed the NVLink capability in the GeForce 4090 series (3090 had them) to avoid cannibalizing their A100 and H100 business. With NVLink, people could chain multiple GPUs and increase the overall VRAM without much reduction in bandwidth. But without it, if you chain two 4090s, your effective bandwidth gets shrunk, so higher VRAM but much slower for LLM inference.
I'm with Linus Torvalds on this—f*** NVidia.