为什么手机内存进入英伟达机柜后，贵过HBM？

TL;DR - Nvidia’s adoption of modular LPDDR5X SOCAMM memory for Vera Rubin systems promises high bandwidth, large capacity, and lower power, but surging AI-server demand is making it scarce and potentially costlier per rack than HBM. Supply constraints could force Nvidia to halve planned CPU-memory configurations.

• SOCAMM2 delivers over 2.5× the bandwidth of comparable RDIMMs while using roughly one-third the space and power, according to Micron estimates.
• A projected Kyber rack pairs 216TB of LPDDR5X with 124.4TB of HBM4E; despite LPDDR5X’s lower unit price, its larger capacity raises estimated rack cost to $2.8 million versus $2.5 million for HBM.
• TrendForce estimates Nvidia may secure only about 60% of its required LPDRAM supply, potentially reducing Vera CPU memory from 1.5TB to 768GB.
• SOCAMM pricing is projected to rise from about $8/GB in early 2026 to more than $13/GB by year-end, with shortages possibly lasting into 2027.