SemiAnalysis again leaks pre-market: NVIDIA Kyber NVL144 rack delayed by over 12 months due to “PCB interposer manufacturing challenges”
Original author: Long Yue
Source: Wall Street Journal
On the morning of July 6, the semiconductor industry research firm SemiAnalysis posted six consecutive tweets on X (formerly Twitter), revealing significant delays and multiple cancellations regarding NVIDIA's Kyber NVL144 rack architecture. The news drew market attention ahead of trading.
SemiAnalysis states: “Major delay: Just three months after Jensen Huang unveiled the Kyber NVL144 at GTC, the product has suffered a major setback, delayed by over 12 months to 2028.”
PCB mid-board: the board that stuck Kyber
SemiAnalysis's analysis points to a key hardware component—the PCB midplane, which NVIDIA also refers to as the "orthogonal backplane"—as the direct cause of the Kyber NVL144 delay.
The institution stated: "The Kyber NVL144 rack architecture has been delayed until 2028 due to significant manufacturing challenges with the PCB mid-plane. The NVL576, which connects 8x Oberon racks via CPO between NVSwitches, is also likely to be delayed or limited to low-volume production due to current CPO challenges."
The gray board displayed by Jensen Huang at the GTC conference in March is the orthogonal backplane of the Rubin Ultra (Kyber architecture) cabinet. It enables a 90-degree vertical interconnection between compute trays and switch trays—compute trays are inserted vertically and connect directly to the rear switch trays via this midplane, eliminating the traditional cable jungle entirely.
The manufacturing difficulty of this board is extremely high. According to the above technical analysis, the backplane uses a Class M9 copper-clad laminate combined with quartz fabric (Q-fabric) and PTFE materials, comprising 78 layers (formed by laminating three 26-layer boards), with trace width and spacing ≤25μm to meet ultra-high-speed signal integrity requirements for 448G+ SerDes rates.
Why must this board be used? According to technical analysis, the Rubin Ultra NVL144 rack requires connecting 144 GPUs within a single domain. If the traditional copper cable solution is retained, over 20,000 cables would be needed, increasing weight by more than 30% and causing severe signal attenuation. The orthogonal backplane is one of the few viable solutions under current technological conditions.
The alternative NVL72x2 has also been canceled.
In response to Kyber's manufacturing challenges, NVIDIA attempted to develop a transitional solution—the NVL72x2 back-to-back rack architecture.
According to SemiAnalysis, the design concept involves placing two Oberon racks back-to-back and expanding the scale domain using pure copper NVLink, thereby circumventing the manufacturing challenges of the Kyber interposer.
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 unusual design and heavy operational burden.
Neither path works; NVIDIA is experiencing a temporary gap in scaling up Rubin Ultra.
NVL576 is also under pressure; the challenges for CPO cannot be ignored.
The delay is not limited to Kyber NVL144. SemiAnalysis also noted that the larger system, NVL576—which connects eight Oberon racks via CPO (Co-Packaged Optics)—“may also be delayed or limited to low-volume shipments given the current challenges with CPO.”
CPO is the optical interconnect technology introduced by NVIDIA for the first time at the Rubin Ultra stage to scale network capacity. According to a report released by SemiAnalysis in March 2026, the NVL576 design employs copper cabling for intra-rack expansion and uses CPO to connect NVSwitches between racks, forming a two-layer fully interconnected network.
However, the mass production readiness of CPO itself remains a variable. SemiAnalysis explicitly stated in its research report that CPO NVSwitch will not be fully ready until the Feynman generation.
The Rubin Ultra core has also been reduced: the 4-chip version has been canceled.
Along with the aforementioned delayed message, there is also an important product-level change being disclosed.
SemiAnalysis stated that the 4-chip version of Rubin Ultra has been canceled, “retaining only the smaller 2-chip version of Rubin Ultra, whose actual performance is approximately half that of the 4-chip version.”
This means that even if the Kyber racks are delivered on schedule, the maximum hash rate per rack has been significantly reduced.
SemiAnalysis stated that NVIDIA will make up 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 the scale-out domain directly impacts NVIDIA's competitive position in large-scale training scenarios.
SemiAnalysis points out: "NVIDIA currently has no proven solution to scale the scale domain of Rubin Ultra, leaving room for competitors such as AMD MI500X or TPUv8i Broadfly to surpass Rubin Ultra in scalability."
According to NVIDIA’s current roadmap, CPO NVSwitch will not be available until the next-generation Feynman platform. Until then, the scalability of Rubin Ultra is limited.
SemiAnalysis noted at the end of the tweet that the aforementioned delays and cancellation decisions impact the memory, PCB, and ODM supply chains.
The manufacturing challenges of Kyber's mid-board directly point to technological bottlenecks among high-end PCB suppliers. The 78-layer ultra-high-density PCB, M9-grade copper-clad laminate, and PTFE hybrid materials required for this mid-board represent the current limits of PCB manufacturing technology.