To The Moon

Semiconductor research firm SemiAnalysis published a series of six consecutive posts on the X platform, disclosing significant delays and multiple cancellations for NVIDIA's Kyber NVL144 rack architecture, drawing market attention ahead of the trading session.

SemiAnalysis stated bluntly: "Major delay: Just three months after Jensen Huang showcased Kyber NVL144 at GTC, the product has encountered a major setback, with a delay exceeding 12 months, pushing it out to 2028."

The Midplane PCB: The Component Holding Up Kyber

The direct cause of the Kyber NVL144 delay, according to SemiAnalysis, points to a critical piece of hardware: the PCB midplane, which NVIDIA officially refers to as the "orthogonal backplane."

The firm stated: "The Kyber NVL144 rack architecture has been delayed until 2028 because the PCB midplane still faces major manufacturing process challenges. The NVL576, which connects 8x Oberon racks between NVSwitches via CPO, is also likely delayed or limited to small production volumes due to current CPO challenges."

The grey board Jensen Huang showcased at the GTC conference in March is precisely the orthogonal backplane for the Rubin Ultra (Kyber architecture) cabinet. Its function is to enable a 90-degree vertical interconnect between the compute tray and the switch tray—the compute tray inserts vertically, connecting directly board-to-board with the rear switch tray via this midplane, completely eliminating the traditional cable jungle.

Manufacturing this board is extremely difficult. According to the technical analysis, this backplane uses an M9-grade copper-clad laminate + quartz fabric (Q-cloth) + PTFE hybrid material, with a layer count reaching 78 (formed by laminating three 26-layer boards), and a line width/spacing of ≤25μm to meet ultra-high-speed signal integrity requirements for 448G+ SerDes rates.

Why is this board essential? Technical analysis indicates that the Rubin Ultra NVL144 rack needs to connect 144 GPUs within a single domain. If a copper cable solution were used, it would require over 20,000 cables, increasing weight by more than 30% and causing severe signal degradation. The orthogonal backplane is one of the few feasible solutions under current technological constraints.

Alternative Plan NVL72x2 Also Cancelled

Faced with Kyber's manufacturing hurdles, NVIDIA had attempted to develop a transitional solution—the NVL72x2 back-to-back rack architecture.

According to SemiAnalysis, this design involved placing two Oberon racks back-to-back and scaling the domain via pure-copper NVLink, thereby circumventing the manufacturing difficulties of the Kyber midplane.

However, this plan ultimately failed to materialize. SemiAnalysis stated that the NVL72x2 "was canceled due to strong opposition from cloud service providers and hyperscale data center operators over its odd design and heavy operational burden."

With both paths blocked, NVIDIA faces a temporary gap in scaling capabilities for Rubin Ultra.

NVL576 Also Under Pressure; CPO Challenges Cannot Be Ignored

It's not just Kyber NVL144 facing delays. SemiAnalysis also noted that the NVL576—a larger-scale system connecting eight Oberon racks via CPO (Co-Packaged Optics)—"may also be delayed or limited to small shipment volumes, given the current challenges with CPO."

CPO is an optical interconnect technology NVIDIA first introduced for scale-out networking in the Rubin Ultra phase. According to a SemiAnalysis research report from March 2026, the NVL576 design maintains copper expansion within racks and connects NVSwitches between racks via CPO, forming a two-layer all-to-all network.

However, the production maturity of CPO itself remains a variable. SemiAnalysis explicitly stated in its report that CPO NVSwitch will not be formally ready until the Feynman generation.

Rubin Ultra Itself Scaled Back: 4-Chip Version Canceled

Disclosed alongside the delay news was a significant product-level change.

SemiAnalysis reported that the 4-compute-die version of Rubin Ultra has been canceled, "leaving only the smaller 2-compute-die version of Rubin Ultra, whose actual performance is roughly half that of the 4-die version."

This means that even if the Kyber rack is eventually delivered on schedule, the computational ceiling for a single rack has been substantially lowered.

Regarding this, SemiAnalysis indicated that NVIDIA will compensate for this gap by "significantly increasing sales of Oberon Rubin racks and Oberon Rubin Ultra racks."

Competitive Window: AMD and Google May Benefit

The gap in scale-out capabilities directly impacts NVIDIA's competitive position in large-scale training scenarios.

SemiAnalysis pointed out: "NVIDIA currently has no proven solution for scaling out the Rubin Ultra scale-out domain, leaving room for competitors like AMD MI500X or TPUv8i Broadfly to surpass Rubin Ultra in scale-out capability."

According to NVIDIA's existing roadmap, CPO NVSwitch will not appear until the next-generation Feynman platform. Until then, the scale-out ceiling for Rubin Ultra remains constrained.

SemiAnalysis concluded its posts by noting that the aforementioned delays and cancellations will impact the memory, PCB, and ODM supply chains.

The manufacturing difficulties of the Kyber midplane directly point to the technological bottlenecks of high-end PCB suppliers. The 78-layer ultra-high-density PCB, M9-grade copper-clad laminate, and PTFE hybrid material required for this midplane represent the current limits of PCB manufacturing technology.