NVIDIA Vera Rubin Memory Costs Dominate Superchip BOM
On August 10, Wccftech and ChosunBiz reported UBS estimates that memory accounts for $24,297 of a $38,902 NVIDIA Vera Rubin Superchip, or about 62%. The reported configuration includes 288GB of HBM4 per Rubin GPU and up to 1.5TB of SOCAMM2 LPDDR5X per Vera CPU. A July 28 TrendForce report described a reportedly lower SOCAMM capacity amid memory prices and supply shortages.
A UBS bill-of-materials estimate reviewed by Wccftech puts HBM4 and SOCAMM2 memory at $24,297 of the estimated $38,902 cost of an NVIDIA Vera Rubin Superchip. That is about 62% of the unit cost, compared with the 53% memory share Wccftech cites for the preceding Grace Blackwell generation.
ChosunBiz, which also reviewed the UBS analysis, reports that the estimated $24,297 memory bill consists of $4,943 in HBM4 attached to the Rubin GPUs and $19,355 in SOCAMM2 memory attached to the Vera CPU. The CPU-side memory alone represents 49.8% of total Superchip cost in the estimate, while HBM4 accounts for 12.7%.
Memory configuration behind the estimate
Wccftech describes the Vera Rubin NVL72 rack, codenamed Oberon, as containing 72 Rubin GPUs and 36 Vera CPUs, arranged as 36 Superchips. In the configuration it reports, each Rubin GPU carries 288GB of HBM4, with up to 22TB/s of memory bandwidth, while each Vera CPU is paired with 1.5TB of LPDDR5X SOCAMM2.
At rack scale, Wccftech calculates that configuration as 20.7TB of HBM4 and 54TB of LPDDR5X. These figures cover accelerator and CPU-attached memory, rather than the full rack bill, which also includes networking, cooling, power delivery, interconnects, and other components.
The UBS-derived Superchip estimate breaks down as follows, according to Wccftech and ChosunBiz:
- •Rubin GPU-related components, including HBM4, packaging, interposer, and peripherals: $9,247
- •Vera CPU-related components, including SOCAMM2: $20,059
- •Other board components: $350
The accounting matters because CPU-attached LPDDR5X, not HBM4, represents the largest individual memory cost in this reported design. For ML infrastructure teams, it illustrates how rack-scale systems can accumulate a large memory bill across host CPUs as well as accelerators.
Capacity report differs from the cited configuration
A July 28 TrendForce report, citing GF Securities analysis and other reports, described a potentially different NVL72 configuration under consideration. TrendForce reported that NVIDIA was reportedly reducing SOCAMM capacity per Vera CPU from 192GB to 96GB, which would lower total CPU memory in the rack from about 55TB to 28TB while leaving GPU HBM4 capacity at 20.7TB.
TrendForce attributed the potential adjustment to elevated memory prices and supply shortages. It reported an estimated Vera Rubin VR200 system BOM of $2.1 million, with memory approaching 29% of that total before configuration changes, and cited an earlier Bernstein estimate that an NVL72 rack could cost as much as $9.1 million.
The reports therefore should not be read as confirming one final shipping memory configuration. The UBS-based breakdown discussed by Wccftech and ChosunBiz uses a 1.5TB-per-CPU configuration, while TrendForce characterizes the lower-capacity design as reported rather than confirmed. ChosunBiz reports that Vera Rubin is now being shipped in earnest, while public reporting has not established whether all deployments use the same SOCAMM capacity.
For the broader AI infrastructure market, comparable rack-scale designs demonstrate why memory supply, module pricing, and packaging capacity can become first-order constraints alongside GPU availability. A system with tens of terabytes of directly attached memory turns changes in LPDDR5X and HBM pricing into material system-level procurement variables, not merely component-level fluctuations.
Key Points
- 1UBS estimates memory contributes $24,297, or 62%, of a $38,902 Vera Rubin Superchip bill of materials.
- 2The reported NVL72 design combines 20.7TB of HBM4 with 54TB of CPU-attached LPDDR5X across 36 Superchips.
- 3Comparable rack-scale architectures make memory supply and pricing material procurement constraints alongside accelerator availability and compute performance.
Scoring Rationale
The reported bill of materials provides a notable view of memory's growing contribution to rack-scale AI infrastructure cost. It is especially relevant to practitioners and infrastructure buyers tracking HBM4, LPDDR5X, capacity planning, and supply constraints, though the configuration and cost estimates are third-party reporting rather than NVIDIA disclosures.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
