The memory is too expensive, can't stand it! NVIDIA Rubin drastically cuts configuration: CPU memory directly halved.

due to high storage chip prices and supply shortages, NVIDIA will significantly reduce the memory configuration of its Vera Rubin NVL72 rack-scale AI system.

Specifically, the SOCAMM memory module capacity in the rack has been halved from 192GB to 96GB, the total memory for the Vera CPU has been cut from 54-55TB to 28TB, and the CPU rack section will also use 96GB SOCAMM modules.

The only unchanged component is the GPU-side HBM4 memory, which remains at 20.7TB per rack.

The direct reason for NVIDIA's move is the tight supply in the LPDDR5X market. Affected by memory price increases, the cost of a Rubin NVL72 rack may reach as high as $9.1 million, making the earlier estimate of $7.8 million outdated. Meanwhile, HBM4 prices could rise to $53 per GB by 2027.

Without adjustments, memory costs would account for 29% of the total VR200 system material cost, approximately $2.1 million, far exceeding the ideal level of 20%.

After halving the configuration, the effect is immediate: the LPDDR5X cost for the VR200 dropped from the earlier estimate of $1.2 million to $586,000. If further compressed to a quarter of the original, it could even be reduced to $293,000.

It's worth noting that NVIDIA, having previously secured long-term storage agreements, has been the company most able to withstand this wave of price hikes. But now even it has to make cuts, highlighting just how much pressure the storage market is under.

For context, the Vera Rubin NVL72 is already one of the most powerful AI systems available, having been released just this January. Based on test data, when processing MoE large model workloads, the Blackwell platform achieves a throughput of 80,000 tokens/s per megawatt of power consumption, while the Vera Rubin NVL72 reaches 800,000 tokens/s—a full 10x improvement.