SemiAnalysis: NVIDIA's Multiple AI Rackscale Architecture Delayed or Adjusted, Rubin Ultra Expansion Path Constrained

BlockBeats News, July 6, SemiAnalysis reported that NVIDIA's latest data center interconnect architecture, Kyber NVL144, underwent a major adjustment just 3 months after its release. The launch, originally planned to be delayed by over 12 months until 2028, was primarily due to manufacturing feasibility challenges in the PCB layout design.

At the same time, the NVL72x2 back-to-back rack architecture has been canceled. This scheme was initially intended to enhance the pure copper NVLink extension capability by deploying two Oberon racks back-to-back. However, due to its complexity and the significant operational burden it placed on hyperscale cloud service providers (CSPs), the market strongly questioned it, leading to its ultimate abandonment.

Since Co-Packaged Optics (CPO) technology is not yet mature, NVIDIA's larger-scale expansion plans based on CPO NVSwitch, such as NVL576, may also face further delays or be limited to small-scale pilot production. This means that NVIDIA lacks a stable large-scale scale-up solution until CPO matures.

Furthermore, there have been changes to the Rubin Ultra product roadmap: the previously planned "Quad Compute Chip" version has been scrapped, leaving only the "Dual Compute Chip" version, with the overall system-level performance scale expected to be reduced to about half of the original plan.

This series of adjustments signifies that NVIDIA's scale-up expansion capability in the Rubin Ultra generation is now limited. In the absence of the implementation of CPO NVSwitch in the Feynman architecture, competitors such as AMD MI500X or Google TPU v8i may gain a relative window of opportunity in the scalability of large-scale training clusters. Meanwhile, NVIDIA expects to bridge the market demand during the product transition period by heavily shipping Oberon Rubin racks and their "Ultra" versions to maintain the overall supply chain pace.