Analyst qinbafrank said on Aug. 2 that market chatter about Nvidia cutting HBM specifications for Rubin has been misread.
According to the post, the issue does not involve the standard Vera Rubin NVL72 systems that have already been delivered. The uncertainty is tied to Rubin Ultra, the upgraded version originally planned for the second half of 2027, whose final specifications have not yet been settled.
Vera Rubin shipments are already underway
qinbafrank said the standard Vera Rubin program is progressing smoothly. Dell delivered the first NVL72 systems to CoreWeave in early June and completed what was described as the industry’s first full boot validation.
By July, dozens of customers had received test racks or initial shipments. The list included Microsoft, OpenAI, Anthropic, Google Cloud, Oracle, Nebius and SpaceX AI. Some of those systems are already running inside customer data centers.
Larger-scale deliveries are set to begin in the fall. The post said the overall rollout is ahead of Blackwell, and rack assembly time without a frame has been cut sharply to roughly the five-minute range.
Questions center on Rubin Ultra’s original design
The more aggressive Rubin Ultra configuration unveiled at GTC 2026 called for four compute chips close to reticle-size limits and 16 HBM4E stacks, targeting about 1 TB of memory in a single package.
That design came under scrutiny in late June after reports from SemiAnalysis and other firms raised concerns about TSMC CoWoS-L substrate warping, reticle-size limits and yield difficulties.
Those reports said the four-chip approach had been dropped and could be replaced by a two-chip design matching the standard version, paired with eight HBM4E stacks. Under that scenario, memory capacity per package would come in at about 384 GB. That would be a sharp reduction from the original target, though some of the system-level performance loss could be offset through rack-scale expansion.
HBM specifications remain unsettled
At the end of July, TrendForce said Rubin Ultra’s HBM configuration was still not locked in. The firm said Nvidia was prioritizing shipment volume and I/O speed as supply stays tight and prices keep rising.
That leaves the company weighing capacity against supply certainty. One possible lower-spec option under consideration, according to the report, includes HBM4E 8-high stacks.
The final Rubin Ultra configuration is expected to be decided after validation in the second half of 2026. Nvidia is still targeting shipments in 2027, but the pace of the program has shifted from aggressive to more pragmatic.