SemiAnalysis releases another pre-market scoop: NVIDIA Kyber NVL144 rack delayed by over 12 months due to "difficulties in PCB mid-board manufacturing."

July 6th morning, semiconductor industry research institution SemiAnalysis released six consecutive posts on the X platform (formerly Twitter), disclosing that Nvidia's Kyber NVL144 rack architecture has encountered major delays and several cancellations. The news drew market attention before trading. SemiAnalysis stated bluntly: "Major delay: Just three months after Jensen Huang showcased Kyber NVL144 at GTC, the product has suffered a significant setback, delayed by more than 12 months and pushed back to 2028." PCB Midplane: The Board Stalling Kyber SemiAnalysis analyzes that the direct cause of Kyber NVL144’s delay points to a key piece of hardware—the PCB midplane, which Nvidia officially calls the “Orthogonal Backplane.” The institution said: "Kyber NVL144 rack architecture is now delayed to 2028 because the PCB midplane faces significant manufacturing challenges." NVL576, which connects 8x Oberon racks via CPO on NVSwitches, is also likely to be delayed or limited to small batch production due to current CPO challenges. The gray board Jensen Huang displayed at the GTC conference this March is precisely the orthogonal backplane of the Rubin Ultra (Kyber architecture) cabinet. Its function is to achieve a 90° vertical interconnection between the compute trays and the switch trays—the compute trays are vertically inserted and directly connected to the rear switch trays via this midplane, completely eliminating the traditional cable jungle. Manufacturing this board is extremely difficult. According to technical analysis, this backplane uses M9-grade copper-clad laminate + quartz cloth (Q cloth) + PTFE mixed materials, has 78 layers (laminated from three 26-layer boards), and line width/spacing of ≤25μm, to meet the ultra-high-speed signal integrity requirements at 448G+ SerDes speeds. Why can't other solutions replace it? According to technical analysis, the Rubin Ultra NVL144 rack needs to connect 144 GPUs in a single domain. If using copper cable solutions, more than 20,000 cables are required, increasing weight by over 30% and causing severe signal attenuation. The orthogonal backplane is one of the rare feasible solutions under current technological conditions. Alternative Solution NVL72x2 Also Cancelled Facing Kyber’s manufacturing challenge, Nvidia once tried to develop a transitional scheme—NVL72x2 back-to-back rack architecture. According to SemiAnalysis, the design concept was to place two Oberon racks back-to-back, expand the scale domain using pure copper NVLink, thus bypassing the manufacturing difficulty of the Kyber midplane. However, this solution ultimately did not materialize. SemiAnalysis said NVL72x2 "was cancelled due to strong opposition from cloud service and hyperscale data center operators regarding its peculiar design and heavy operational burden." With both paths blocked, Nvidia faces a temporary vacuum in scale expansion of Rubin Ultra. NVL576 Also Under Pressure, CPO Challenges Can't Be Ignored It’s not just Kyber NVL144 that’s delayed. SemiAnalysis pointed out that NVL576—a larger system connecting eight Oberon racks using CPO (Co-Packaged Optics)—“could also face delays or be limited to small batch shipments due to current CPO challenges.” CPO is an optical interconnect technology introduced by Nvidia for network scale expansion at the Rubin Ultra stage. According to SemiAnalysis’s research report from March 2026, the NVL576 design keeps copper expansion inside racks, but connects racks via CPO to the NVSwitch, forming a two-layer fully interconnected network. But the mass-production maturity of CPO itself remains variable. SemiAnalysis made clear in the report that CPO NVSwitch will not be fully ready until the Feynman generation. Rubin Ultra Shrinks: 4-Chip Version Cancelled Alongside the above delay news, there’s also an important product-level change. SemiAnalysis claims the 4-compute-chip version of Rubin Ultra has been cancelled, “with only the smaller 2-compute-chip version of Rubin Ultra retained, whose actual performance is roughly half that of the 4-chip version.” This means that even if the Kyber rack is finally delivered as scheduled, the ceiling for single rack computational power is now much lower. To fill this gap, SemiAnalysis states Nvidia will “substantially increase sales of Oberon Rubin racks and Oberon Rubin Ultra racks.” Competitive Window: AMD and Google May Benefit The gap in scale expansion directly affects Nvidia’s competitive position in large-scale training scenarios. SemiAnalysis points out: “Nvidia currently lacks a validated solution to scale Rubin Ultra, leaving room for competitors like AMD MI500X or TPUv8i Broadfly to surpass Rubin Ultra in scale expansion capability.” According to Nvidia's current roadmap, CPO NVSwitch will only appear with the next-generation Feynman platform. Until then, Rubin Ultra’s scale expansion ceiling remains constrained. At the end of the thread, SemiAnalysis reminds that the above delays and cancellations impact memory, PCB, and ODM supply chains. Kyber midplane’s manufacturing challenges point directly to the technological bottlenecks of high-end PCB suppliers. The required 78-layer ultra-high-density PCB, M9-grade copper-clad laminate, and PTFE mixed materials represent the limits of current PCB manufacturing processes. Risk disclosure and disclaimer The market has risks; investment must be cautious. This article does not constitute individual investment advice and does not consider the specific investment objectives, financial situations, or requirements of any particular users. Users should consider whether any opinions, views, or conclusions in this article suit their circumstances. Investing based on this is at your own risk.