SemiAnalysis: Multiple AI rack-scale architecture updates from NVIDIA have been delayed or adjusted, and the expansion path for Rubin Ultra is limited

BlockBeats message, on July 6, SemiAnalysis posted that NVIDIA’s latest rack-level interconnect architecture, Kyber NVL144, has undergone major adjustments just 3 months after its release. The original plan was delayed by more than 12 months to 2028, primarily because PCB planar design still faces challenges in manufacturability.

At the same time, the NVL72x2 back-to-back rack architecture has been canceled. This solution was originally intended to enhance pure-copper NVLink expansion capability by deploying two Oberon racks back-to-back, but due to its structural complexity and the high operational burden it places on hyperscale cloud providers (CSPs), it faced strong market skepticism and was ultimately abandoned.

Because CPO (co-packaged optics) technology has not yet matured, NVIDIA’s larger-scale expansion plans based on CPO NVSwitch (such as NVL576) may continue to be delayed, or may be limited to small-batch trial production. This means that before CPO matures, NVIDIA lacks a stable large-scale scale-up solution.

In addition, the product roadmap for Rubin Ultra has changed: the originally planned “four-compute-chip” version has been canceled, leaving only the “dual-compute-chip” version. The overall system-level performance scale is expected to be reduced to about half of the original plan.

A series of these adjustments means that NVIDIA’s scale-up expansion capability in the Rubin Ultra generation is constrained. With CPO NVSwitch unable to be implemented before the Feynman architecture, competitors such as AMD MI500X or Google TPU v8i may gain a relative window of opportunity in scaling large training clusters. Meanwhile, NVIDIA is expected to fill market demand during the product transition period by shipping large quantities of Oberon Rubin racks and their “Ultra” versions, while maintaining the overall supply-chain cadence.