If you look into Digits and Vanta, Digits is better for FP4 inference, while a 4090 setup is more cost effective per FP32 FLOP, ideal for training large models and general compute

8x GPUs can handle parallel tasks better than Digits' single chip and scale easier. And 192GB of memory also exceeds its 128 GB unified memory for massive workloads

We’ve got 4090s on deck, stock’s solid, can get a rig built in 1 day and shipped fast