Nvidia Kyber NVL144 Delayed to 2028 as PCB Bottleneck Hits Rubin Ultra

Nvidia's Kyber NVL144 rack architecture faces a significant delivery delay that pushes mass production to 2028. This shift matters because it removes a key piece of infrastructure from the near-term supply chain for large-scale AI training. Buyers relying on this specific rack configuration must now adjust their deployment timelines by more than a year.

Manufacturing challenges push mass production to 2028 and cancel the four-chip Rubin Ultra variant

The delay stems from manufacturing challenges with the orthogonal backplane, also known as the midplane PCB. This central board connects the compute modules and requires extreme precision to function correctly. Nvidia originally planned to launch a four-chip version of the Rubin Ultra GPU on this platform. That larger version has now been cancelled entirely.

• PCB Layers: 78 layers
• PCB Material: M9 grade CCL and quartz cloth
• Architecture Type: Orthogonal Backplane (Midplane PCB)
• Rubin Ultra Version: 2 compute chip version (4-chip version cancelled)

Only the two-chip version of the Rubin Ultra remains in production. The midplane for this product uses a 78-layer structure to handle signal integrity. Engineers use M9 grade copper clad laminate and quartz cloth to build these boards. These materials increase the cost and difficulty of processing the printed circuit boards significantly.

The cancellation of the four-chip model reduces the performance ceiling for this architecture. According to SemiAnalysis, the remaining two-chip version delivers approximately half the performance of the cancelled four-chip design. This performance gap leaves room for competitors like the AMD MI500X or TPUv8i Broadfly to gain ground in scale-up capabilities.

We looked at earlier reports on Nvidia's GPU roadmap while tracking these supply chain updates. The project encountered setbacks just three months after Jensen Huang publicly demonstrated the product. Mass production is now postponed to 2028, marking a delay of over 12 months from the original schedule. The two-chip Rubin Ultra version will be the only variant available on this platform.