NVIDIA Vera Rubin Ultra Kyber Racks Could Shift Memory Cost Pressure From HBM4E to LPDDR5X
NVIDIA’s next-generation Vera Rubin Ultra platform is becoming one of the most closely watched developments in the AI hardware market. While the company has not finalized every detail publicly, industry estimates suggest a major shift may be underway in how memory costs are distributed across future NVIDIA AI racks.
Recent market analysis indicates that NVIDIA’s upcoming Vera Rubin Ultra-based Kyber racks may end up spending more on LPDDR5X memory for the Vera CPU side than on HBM4E memory for the Rubin GPUs. That is a notable change, especially because high-bandwidth memory has often been viewed as one of the most expensive and critical components in modern AI accelerators.
Rumors have also suggested that NVIDIA may consider reducing the amount of HBM4E used per GPU in order to protect margins. Whether that happens or not, the broader trend is clear: as AI servers grow larger and more memory-intensive, the total memory bill is expanding rapidly across both GPU and CPU components.
The base Rubin GPU is currently expected to feature 288GB of HBM4 across eight stacks, with bandwidth reaching up to 22TB/s per GPU. This is a huge leap for AI training and inference workloads that require extremely fast access to massive datasets.
At the same time, NVIDIA’s Vera CPU takes a different route from traditional server memory. Instead of relying on standard RDIMM modules, Vera is expected to use LPDDR5X-based SOCAMM2 memory modules. These modules use a compact compression connector that sits flat against the motherboard. This design helps shorten electrical traces, improving signal integrity and reducing latency, which are both essential for high-performance data center systems.
Each Vera CPU reportedly includes eight SOCAMM2 slots. With 192GB modules, that creates a theoretical capacity of 1.5TB of LPDDR5X memory per CPU. In large rack-scale deployments, that adds up quickly.
According to recent estimates, HBM4E in Rubin may cost around $19.76 per GB. However, in a 144-GPU Kyber rack, LPDDR5X could become the larger memory expense. The same estimates suggest that each Kyber rack may include roughly 216TB of LPDDR5X memory, costing about $2.8 million. By comparison, the HBM4E portion is estimated at about 124.4TB, with a cost of around $2.5 million.
That means LPDDR5X could exceed HBM4E as the larger memory cost in NVIDIA’s Vera Rubin Ultra Kyber rack configuration. The full Kyber rack is estimated to cost approximately $41.6 million, showing just how expensive next-generation AI infrastructure is becoming.
Another estimate breaks down the cost of the Vera Rubin superchip more directly. A Rubin GPU is believed to cost around $9,247 when including HBM4, advanced packaging, the interposer, and related components. The Vera CPU is estimated at around $20,059, including SOCAMM2 LPDDR5X memory, plus roughly $350 for additional board components.
The most striking detail is how much of the system cost is tied to memory. Nearly half of the Rubin GPU cost is believed to come from DRAM-related components. On the CPU side, the impact is even more dramatic. Without memory, the Vera CPU itself is estimated to cost only around $704, meaning most of its total cost comes from the LPDDR5X memory attached to it.
This highlights an important shift in AI server economics. HBM remains essential because GPUs need extremely high bandwidth to handle AI model training, large language models, and accelerated computing workloads. But as NVIDIA builds bigger rack-scale systems with larger shared memory pools, CPU-side memory is becoming just as important from a cost perspective.
Compared with NVIDIA’s Blackwell generation, Vera Rubin appears to take memory scaling to a new level. In Blackwell servers, Grace CPUs featured 480GB of memory, while Blackwell GPUs offered 192GB of HBM3E in GB200 systems and up to 288GB of HBM3E in GB300 configurations.
Vera Rubin dramatically expands the total memory footprint. Rubin racks are estimated to reach a combined memory pool of around 74.7TB when counting both HBM4 and SOCAMM2 memory. Kyber racks go much further, potentially reaching an enormous 340.4TB of total memory.
For AI customers, this massive memory expansion could be critical. Larger memory pools allow data centers to support more complex AI models, larger context windows, faster inference pipelines, and more efficient training clusters. As AI workloads continue to grow, memory capacity and bandwidth are becoming just as important as raw compute performance.
For NVIDIA, however, the challenge is balancing performance with profitability. HBM4E is expensive, advanced packaging remains costly, and LPDDR5X capacity at rack scale is no longer a minor line item. If these estimates prove accurate, memory suppliers could become even more important partners in NVIDIA’s AI hardware roadmap.
The Vera Rubin Ultra generation is shaping up to be more than a simple GPU upgrade. It represents a broader change in AI system architecture, where the cost and performance of every memory tier matter. HBM4E will remain a core part of NVIDIA’s GPU strategy, but LPDDR5X SOCAMM2 memory may become one of the biggest cost drivers in the company’s future AI racks.
As hyperscalers and enterprise AI customers prepare for the next wave of accelerated computing, Vera Rubin Ultra and Kyber racks could define the next phase of AI data center design: bigger systems, larger memory pools, higher costs, and even greater demand for efficient rack-scale performance.
