“Each Grace Hopper Superchip combines 480 GB LPDDR5 CPU memory and 96 GB HBM3, interconnected with NVLink-C2C providing 7x more bandwidth than PCIe Gen5 at one-fifth the power.”
Nvidia’s 144 terabytes of “GPU memory” is mostly CPU memory. Each of the 256 superchips carries 480 GB of LPDDR5 and only 96 GB of HBM3, so about 24.5 TB of the total is actual GPU memory. The rest is reachable over NVLink and counted anyway. Nvidia rounded six parts CPU RAM and one part HBM into a GPU headline.