[NVIDIA proposes AI memory full-stack collaborative design solution Vera Rubin platform only increases memory capacity per GPU by about 3% and inference throughput increases by up to 30 times] November 6, NVIDIA published an article on November 5 stating that with the expansion of AI model scale, extension of context windows, and popularization of agent applications, GPU memory pressure continues to rise, HBM supply is tightening and prices are rising. Nvidia proposes to improve memory usage efficiency through collaborative design of GPU, CPU, network, storage and software, rather than simply increasing GPU memory capacity. Specifically, NVIDIA compresses model weights and KV Cache through NVFP4 low-precision format, transfers temporarily inactive contexts to lower-cost storage layers, and uses multi-GPU distributed deployment to release HBM space. NVIDIA disclosed that in the SemiAnalysis AgentX test, the unit cost inference throughput of GB300 NVL72 under specific workloads was up to 10 times that of the H200 platform; Vera Rubin NVL72 compared with GB300 NVL72, the unit cost inference throughput was up to 30 times higher, while the single GPU HBM capacity only increased by about 3%.