What is the reason why NVIDIA's latest GPU uses less Dram? Last night, SemiAnalysis reported that Nvidia has taken the initiative to cut the CPAmeric OCAMM DRAM capacity of its next-generation AI supercomputer platform, the Rubin NVL72 rack, by half. As soon as the news came out, the entire storage sector plummeted. Let's talk in detail: 1. What the hell is this talking about? NVIDIA's main AI rack platform for 2026, equipped with 72 Rubin GPUs and 36 Vera CPUs in one cabinet, fully interconnected through NVLink to form a "giant GPU". HBM4 (approximately 20.7TB per cabinet) is used on the GPU side, and SOCAMM - a pluggable LPDDR5X system memory module - is used on the CPU side for easy maintenance and fault isolation. Original plan: Each Vera CPU uses 8 192GB SOCAMM modules (up to approximately 1.53TB/CPU), and the total DRAM on the CPU side of the entire cabinet is approximately 55TB. Now: Most systems have switched to 96GB SOCAMM modules, and the DRAM on the CPU side of the entire cabinet has decreased to about 28TB (reduced by~50%). Key unchanged: The HBM4 capacity and bandwidth of the GPU remain unchanged, and the core computing power (training/inference FLOPS) is not affected. SemiAnalysis estimates rack hardware cost (BoM): reduced from $7.6M to $6.8M (saving approximately 10%). Simply put, Nvidia proactively downgraded its non core CPU system memory in order to ship the entire machine faster and more smoothly. What is the core reason? In fact, the essence is the face of supply chain reality, with a severe shortage of LPDDR5X high-density SOCAMM. The reason for this downward adjustment is not due to insufficient demand, but rather a supply bottleneck. Due to the extremely difficult packaging yield of 16 layer ultra-high density LPDDR5X, Hynix/Samsung is unable to provide sufficient 192GB of memory to customers. In order to ensure that all Rubin GPUs have corresponding DRAM available, Nvidia actively implemented supply chain risk control and reduced the usage of single cabinet units So it is also a strategic choice for Nvidia. Instead of waiting for high specification memory and causing delays/shortages in the entire cabinet, it is better to proactively reduce to the 96GB version to ensure that all Rubin GPUs can be fully equipped with system memory and quickly enter the production/delivery stage. This actually indirectly verifies that the current storage field is in a state of "supply structural shortage", while AI's demand for memory is exploding (HBM+LPDDR5X). Especially the hbm configuration has not changed at all 3. And it also has its positive impact 1) Accelerate Rubin deployment: lower cost, better TCO, hyperscalers (MSFT, CoreWeave, xAI, etc.) are more willing to purchase on a large scale. MSFT has completed the first cabinet bringing up, and the production pace is faster after the downgrade. 2) Customer profit: Save $800000 per container, save $0.26 per GPU hour, improve long-term token economics, and drive more AI investment. 3) NVIDIA Ecological Enhancement: Continue to control the supply chain (directly buying memory), further consolidating the dominant position of AI factories. Of course, the decrease in CPU side DRAM ratio and memory density (originally 55TB, now 28TB) will have an impact on scenarios such as extremely long context inference, multi-agent parallelism, and ultra large KV cache (which may require more cross node communication or SSD offloading). This should be beneficial for the increase in SSD and optical connectivity usage: The logic lies in the shrinking of system memory used to carry large context (KV Cache) on the CPU side, and the GPU computing bottleneck will inevitably shift to the SSD side and interconnect side. CSPs need to purchase more high-performance SSDs or adopt higher performance in cabinet connection (Scale Up) connection solutions, which is beneficial for the needs of NAND enterprises and optical link enterprises. This article is sponsored by @ bitget_zh, titled 'Bitget Buying US Stocks: Instant Entry, Smooth Trading'