DGX Spark is widely considered to be a worthless crap.
Purely from GPU / VRAM pespective, it's mostly these things that are widely accepted as go-to choices:
- RTX 3090 (with 2 of them, you can run some strong local models at 48GB VRAM)
- 5090, 4090, 4090 48GB mod (obviously expensive af to get more than 1 card)
- 7900 XTX (using llama.cpp Vulkan instead of CUDA)
- RTX Pro 6000 Blackwell (one 96GB card is enough, super expensive)
- stacking a few RTX 5060 Ti 16GB cards (slower than RTX 3090 by 1.5 to 2 times?)
While all the gamers whine about "AI enthusiasts" stealing their GPUs, those "AI enthusiasts" largely don't give a shit about anything other than a bunch of used RTX 3090 cards (at least when they aren't into image/video generation; for that stuff, 5090 would be quite significantly better).