Explain the Nvidia Rubin reduction rumor in detail: the standard version of Vera Rubin has been delivered to dozens of customers and released in the fall, and the Rubin Ultra specs have not yet been locked

Comparing news, analyst Qinbafrank wrote an article clarifying market discussions on the Nvidia Rubin HBM distribution reduction, pointing out that the actual situation is not a reduction in the distribution of the standard Vera Rubin NVL72 that has already been delivered, but rather that the configuration specifications of the upgraded Rubin Ultra, which was originally planned to be launched in the second half of 2027, have not yet been finalized. The standard version of Vera Rubin is progressing smoothly. Dell took the lead in delivering the first batch of NVL72 systems to CoreWeave in early June and completed the industry's first complete startup verification. By July, dozens of customers had received test racks or initial shipments, including Microsoft, OpenAI, Anthropic, Google Cloud, Oracle, Nebius, and SpaceX AI, some of which are already operating in customer data centers, and will launch a larger scale in the fall Delivery, the overall schedule was superior to Blackwell, and the frameless assembly time was drastically reduced to about 5 minutes.

The Rubin Ultra aggressive configuration announced by GTC 2026 - 4 computing chips close to the limit size of the mask plus 16 HBM4E stacks to achieve about 1TB of memory in a single package - was questioned by reports from SemiAnalysis and other agencies at the end of June. Due to TSMC's CoWos-L substrate warpage, mask size limitations, and yield problems, the four-chip solution was supposedly cancelled, and may switch to the same dual-chip design as the standard version plus 8 HBM4E stacks, with a single package capacity of about 384GB, a significant reduction from the original target, but system level performance can be partially compensated through rack-level expansion.

At the end of July, TrendForce further pointed out that Rubin Ultra's HBM specifications are still unlocked. Nvidia prioritizes shipping volume and I/O speed in the context of tight supply and continued price increases, and may consider lower specification options, including the HBM4E 8-tier stack. The core is the balance between capacity and supply certainty. The final specifications are expected to be finalized after verification in the second half of 2026, and Nvidia's goal is still to ship in 2027, but the overall pace has changed from aggressive to pragmatic.