Thông tin đồn đại về cấu hình NVIDIA Rubin HBM đã được làm rõ: Phiên bản tiêu chuẩn đã được giao, phiên bản Ultra vẫn chưa được quyết định

The Huoxing Finance report states that on August 2, analyst qinbafrank clarified market discussions regarding alleged downgrades to NVIDIA’s Rubin HBM, noting that the actual situation is not a downgrade of the standard Vera Rubin NVL72 already delivered, but rather that the configuration specifications for the upgraded Rubin Ultra, originally planned for release in the second half of 2027, have not yet been finalized. Progress on the standard Vera Rubin is proceeding smoothly: Dell delivered the first NVL72 systems to CoreWeave in early June and completed the industry’s first full system validation; by July, dozens of customers—including Microsoft, OpenAI, Anthropic, Google Cloud, Oracle, Nebius, and SpaceX AI—had received test racks or initial shipments, with some already operational in customer data centers. Large-scale deliveries are set to begin in autumn, with overall progress outpacing Blackwell and rack assembly time significantly reduced to approximately five minutes. The aggressive configuration for Rubin Ultra announced at GTC 2026—featuring four compute dies approaching mask size limits and 16 HBM4E stacks to achieve ~1TB of memory per package—was questioned by institutions such as SemiAnalysis at the end of June due to issues with TSMC’s CoWoS-L substrate warping, mask size constraints, and yield challenges. The four-die design is reportedly canceled and may shift to a dual-die design identical to the standard version, paired with eight HBM4E stacks, resulting in ~384GB per package—a significant reduction from the original target but partially offset by rack-level scaling to compensate for system-level performance. At the end of July, TrendForce further noted that the HBM specification for Rubin Ultra remains unconfirmed. Amid supply constraints and continuously rising prices, NVIDIA is prioritizing shipment volume and I/O speed, potentially considering lower-spec options such as eight-layer HBM4E stacks, with the core focus being a balance between capacity and supply certainty. The final specification is expected to be confirmed after validation in the second half of 2026; NVIDIA still aims for 2027 shipments, but the overall timeline has shifted from aggressive to pragmatic.