Nvidia Reportedly Makes Memory Cuts to Rubin Ultra and Vera Rubin Amid HBM4 Shortage — Here's What Changed

• 12-Hi HBM4e (the original plan)
• 8-Hi HBM4e
• 12-Hi HBM4
• 8-Hi HBM4

No final specification has been set . The shift is a direct response to an expected industry-wide DRAM shortage in 2027 that could constrain the wafer capacity available for HBM production, along with ongoing uncertainty over HBM4e qualification timelines among the three major memory makers .

Vera Rubin NVL72: SOCAMM Capacity Halved

On the Vera Rubin platform, the reported change is more concrete. Nvidia has reportedly decided to halve the SOCAMM memory capacity in the Vera Rubin NVL72 rack-scale system, moving from 192 GB to 96 GB per module . According to analysis cited by TrendForce and GF Securities, the total Vera CPU memory per rack would drop from roughly 55 TB to 28 TB, while GPU HBM4 capacity remains unchanged at 20.7 TB per rack . The move is intended to reduce costs and ease supply constraints, with memory costs having climbed to an estimated 29% of the total system bill of materials (BOM) for the Vera Rubin platform .

Customer-Specific SKUs or Shortage Response?

Some sources note that what looks like a memory cut may actually be Nvidia testing multiple customer-specific SKUs — different memory tiers for different workloads — rather than cutting a single flagship spec . This interpretation aligns with Nvidia's official position.

Nvidia's Official Response

Nvidia has publicly denied that Vera or Rubin were cut down due to HBM shortages. In a blog post on August 7, 2026, Nvidia pushed back on speculation, stating that the configurations in question are different SKUs and customer-specific tuning, not a downgrade forced by supply constraints . The company also said it continually adjusts compute, networking, and memory to deliver the best performance and efficiency for customers .

CEO Jensen Huang has been more direct. In June 2026, he confirmed that all three major memory makers — Samsung, SK Hynix, and Micron — were certified for HBM4 supply, with Vera Rubin in full production and Rubin, Rubin Ultra, and Kyber servers on track . In July 2026, Huang dismissed reports of delays, stating that Vera Rubin production is proceeding with "giant" volumes .

The core tension is clear: TrendForce and other analysts interpret the expanded HBM evaluation as a supply-driven contingency plan; Nvidia frames it as normal product variant planning for different customer workloads.

Memory Cost Pressures Reshaping the Roadmap

HBM4 Is Sold Out for 2026

High-bandwidth memory 4 (HBM4) is sold out for all of 2026, and the DRAM shortage is expected to continue through 2027 . This creates a structural pricing and availability headwind for every next-generation Nvidia platform, from Vera Rubin through Rubin Ultra and beyond.

Memory's Rising Share of System Cost

Memory costs have become a significantly larger portion of total system cost. At roughly 29% of the total system BOM for Vera Rubin NVL72 systems, memory (HBM + SOCAMM) is no longer a supporting component — it is a primary cost driver forcing design trade-offs .

Record-Breaking Supply Deals

To secure enough supply, Nvidia has moved to lock in production capacity at unprecedented scale. Nvidia and SK Group announced a strategic initiative worth over $500 billion that includes a long-term HBM supply partnership with SK Hynix for HBM4 memory covering the Vera Rubin generation . Days later, Samsung Electronics and Broadcom struck a separate $200 billion collaboration . These deals, totaling roughly $700 billion in committed supply, signal that securing enough memory at acceptable cost is now a multi-hundred-billion-dollar strategic priority for the entire AI chip ecosystem.

Product Mix Flexibility as a Strategy

Rather than betting everything on a single flagship HBM4e configuration, Nvidia appears to be preparing a broader range of SKUs — varying HBM stack heights and memory sizes — to serve different price points and availability scenarios . This allows the company to keep shipping in volume even if the highest-end HBM4e stacks are scarce. If the 12-Hi HBM4e parts are slow to qualify, Nvidia can still ship Rubin Ultra with 8-Hi HBM4e or HBM4 parts, keeping customers supplied even if per-chip performance is slightly lower.

Bottom Line

The reports point to real supply-and-cost pressure driving Nvidia to evaluate lower-capacity HBM variants and already cut SOCAMM in the Vera Rubin system. Nvidia officially denies any forced downgrade, calling the variants customer-specific tuning rather than a shortage reaction. However, the record-breaking supply deals and memory's rising share of system cost make it clear that HBM constraints are a major factor shaping the Rubin-era product lineup. The final Rubin Ultra memory specification has not been announced, but the direction is unmistakable: the next generation of AI hardware will be defined not just by GPU performance, but by the availability and cost of the memory that feeds it.