https://nvidianews.nvidia.com/news/nvidia-puts-grace-blackwell-on-every-desk-and-at-every-ai-developers-fingertips https://nvidianews.nvidia.com/news/nvidia-put

https://nvidianews.nvidia.com/news/nvidia-puts-grace-blackwell-on-every-desk-and-at-every-ai-developers-fingertips https://nvidianews.nvidia.com/news/nvidia-puts-grace-blackwe... GB10, or DIGITS, is $3,000 for 1 PFLOP (@4-bit) and 128GB unified memory. Storage configurable up to 4TB. Can be paired to run 405B (4-bit), probably not very fast though (memory bandwidth is slower than a typical GPU's, and is the main bottleneck for LLM inference).