“A single 10-kilowatt SambaNova SN40 rack outperforms a 130-140 kilowatt rack of NVIDIA GPUs for inference tasks.”