Rising HBM4 costs are forcing Nvidia to pare back the memory configuration on its next-generation Vera Rubin AI platform, a sign that the memory supply crunch is starting to bend hardware roadmaps. The company is expected to ship Vera Rubin racks with half the originally planned SOCAMM capacity per CPU.
Each Vera CPU will get 96GB modules instead of 192GB, dropping total system memory from 55TB to roughly 28TB per NVL72 rack while keeping the 20.7TB of HBM4 attached to GPUs unchanged. GF Securities estimates the cut saves about $614,000 per VR200 system in memory costs, with further reductions possible if Nvidia drops to one-quarter of the original plan.
The constraint traces back to tight LPDRAM supply. TrendForce data shows Samsung, SK Hynix and Micron can meet only 60% of Nvidia’s estimated demand under current allocation plans. HBM4 pricing is forecast to hit $53 per gigabyte by 2027 according to Bernstein, putting memory at nearly 30% of total system cost versus Nvidia’s preferred 20% threshold.
Cutting SOCAMM lets Nvidia stretch limited LPDRAM supply across more Vera CPU shipments while leaving the GPU memory subsystem intact, where HBM4 delivers the highest performance impact for AI inference and training workloads.
