Nvidia Kyber NVL144 AI Rack Delayed to 2028 — PCB Midplane Bottleneck, Canceled Rubin Ultra, and What Still Ships This Year

On July 6, 2026, semiconductor research firm SemiAnalysis published a series of posts on X claiming that Nvidia's next-generation Kyber NVL144 rack-scale architecture — designed for the 2027 Rubin Ultra GPUs — has been delayed by more than 12 months, pushing its release to 2028 . The product had been demonstrated by CEO Jensen Huang at GTC 2026 only three months prior .

Manufacturing Root Cause: PCB Midplane

SemiAnalysis attributed the delay directly to the PCB midplane (also called the "orthogonal backplane") — a complex, multi-layer printed circuit board at the center of the rack system. The firm stated: "Kyber NVL144 rack architecture has been delayed to 2028 because the PCB midplane remains challenging from a manufacturing feasibility standpoint" . Multiple outlets, including CNBC, confirmed this as the cited bottleneck .

Jefferies issued a research note on June 22 maintaining a Buy rating on Nvidia (target $300) but acknowledged that supply chain signals from May onward made a Kyber delay or cancellation a "highly probable event" .

Nvidia's Official Response: No Confirmation or Denial

As of the latest reporting, Nvidia has not issued a formal denial or confirmation of the delay. CNBC and other outlets explicitly note that Nvidia has not officially confirmed the reports and did not respond to requests for comment . No public statement from the company denying the SemiAnalysis report was found in available sources.

Additional Roadmap Changes Reported by SemiAnalysis

Beyond the Kyber delay, SemiAnalysis outlined several other roadmap disruptions:

NVL72x2 Back-to-Back Architecture Canceled. SemiAnalysis reported that Nvidia's proposed alternative to Kyber — a back-to-back rack architecture connecting two Oberon racks — has been canceled . The firm stated that cloud service providers and hyperscalers "strongly opposed" the design due to its operational burden .

Four-Die Rubin Ultra GPU Scrapped. On June 29-30, SemiAnalysis reported that the original quad-die Rubin Ultra GPU design (4 compute chiplets + 16 HBM4E modules in one package) had been canceled due to manufacturing execution concerns . The replacement is a dual-die design roughly half the size and half the performance of the original . The root cause is reportedly TSMC's CoWoS-L packaging hitting a physical ceiling at the four-die scale, with its CoPoS successor not expected until late 2028 .

NVL576 Likely Delayed or Limited. The larger NVL576 configuration, which connects eight Oberon racks via co-packaged optics (CPO) on NVSwitches, is also seen as likely delayed or restricted to low-volume shipments due to ongoing CPO challenges .

800VDC Power Architecture Delayed. Earlier in June, SemiAnalysis flagged that 800VDC power architecture shipments are now expected later than previously forecast .

Impact on Asian Supply Chain Stocks

Asian PCB supply chain stocks fell broadly on July 6 following the SemiAnalysis report. Taiwan's Economic Daily News reported that shares across the Asian PCB supply chain declined, with Korea's NH Investment & Securities stating the region's tech weakness was "apparently affected by the SemiAnalysis report" . The report noted that the combination of the Kyber delay, NVL72x2 cancellation, and Rubin Ultra downsizing "increases uncertainty in Nvidia's next-generation scaling roadmap" .

Earlier (June 23), PCB manufacturers had already pushed back against rumors of Nvidia demanding 10% price cuts, calling them "exaggerated and misinterpreted" .

Current-Generation Vera Rubin: On Track and Ahead of Schedule

Nvidia's Vera Rubin NVL144 platform (the current-generation Rubin system, using standard Rubin GPUs rather than Rubin Ultra) is on track or ahead of schedule:

• Vera Rubin NVL144 is still expected to ship in volume production in H2 2026, delivering 8 exaflops of AI compute and 100 TB of fast memory per rack .
• Nvidia reportedly pulled the Vera Rubin program forward by approximately six months from its original 2027 target . Foxconn has completed engineering validation .
• An Nvidia developer blog from March 2026 confirms the Vera Rubin POD architecture with seven chips and five rack-scale systems .

In short: The current-generation Vera Rubin NVL144 is unaffected and shipping this year. The delays and cancellations are concentrated on the next generation: Kyber (rack), Rubin Ultra (GPU die configuration), and the NVL72x2/NVL576 scale-out paths.