ChainThink report: On August 2, analyst qinbafrank posted that the market's discussion regarding NVIDIA's Rubin HBM "downgrade" refers not to the standard Vera Rubin NVL72 already delivered, but to the yet-to-be-finalized specifications of the upgraded Rubin Ultra, originally planned for release in the second half of 2027.
The standard version of Vera Rubin is progressing normally. Dell delivered the first NVL72 systems to CoreWeave in early June and completed the industry’s first full boot validation;
As of July, dozens of customers, including Microsoft, OpenAI, Anthropic, Google Cloud, Oracle, Nebius, and SpaceX AI, have received test racks or initial shipments, with some already operational in customer data centers. Larger-scale deliveries are expected to begin in the fall.
Previously announced at GTC 2026, the Rubin Ultra solution features four compute chips approaching mask size limits, paired with 16 HBM4E stacks, totaling approximately 1 TB of memory per package.
At the end of June, institutions such as SemiAnalysis reported that the four-chip design was reportedly canceled due to issues including TSMC’s CoWoS-L substrate warping, mask size limitations, and yield concerns, potentially shifting to a dual-chip design with eight HBM4E stacks, resulting in a single package capacity of approximately 384GB.
TrendForce further noted at the end of July that Rubin Ultra's HBM specifications have not yet been finalized.
Amid supply constraints and rising prices, NVIDIA may prioritize shipment volumes and I/O speed, considering lower-spec options such as HBM4E with 8-layer stacking; the final specifications are expected to be finalized after validation in the second half of 2026, with a target shipment date of 2027.