Original Title: "SemiAnalysis Again Leaks Before Market Opening: NVIDIA Kyber NVL144 Rack Delayed Over 12 Months Due to 'Difficulties in PCB Midplane Manufacturing'"

Original Author: Long Yue

Original Source: Wall Street Journal

On the morning of July 6, semiconductor industry research institution SemiAnalysis published six consecutive tweets on the X platform (formerly Twitter), revealing significant delays and multiple cancellation decisions concerning NVIDIA's Kyber NVL144 rack architecture. The news attracted market attention before the market opened.

SemiAnalysis stated directly: "Significant delay: Just three months after Jen-Hsun Huang showcased Kyber NVL144 at GTC, the product has encountered a major setback, delayed by more than 12 months, postponed to 2028."

PCB Midplane: The Board That Constraints Kyber

SemiAnalysis analyzed that the direct reason for the delay of Kyber NVL144 points to a key hardware component - the PCB midplane, which NVIDIA officially refers to as the "Orthogonal Backplane."

The institution stated: "The Kyber NVL144 rack architecture has been delayed until 2028 because the PCB midplane still faces significant challenges in the manufacturing process. NVL576, which connects 8x Oberon racks through CPO, is also very likely to be delayed or limited to small-scale production due to current CPO challenges."

The gray board showcased by Jen-Hsun Huang at the GTC event in March is, in fact, the orthogonal backplane of the Rubin Ultra (Kyber architecture) cabinet. Its role is to realize a 90° vertical interconnection between calculation trays and switching trays—calculation trays are inserted vertically, achieving direct board-to-board connections through the midplane with the rear switching tray, completely eliminating the traditional cable jungle.

The manufacturing difficulty of this board is extremely high. According to the aforementioned technical analysis, the backplane uses M9-level copper-clad laminate + quartz cloth (Q cloth) + PTFE composite materials, with a total of 78 layers (formed by laminating 3 pieces of 26-layer boards), with line width and spacing ≤ 25μm, to meet the ultra-high-speed signal integrity requirements at 448G+ SerDes rates.

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

Alternative NVL72x2 Scheme Has Also Been Canceled

Faced with manufacturing difficulties for Kyber, NVIDIA attempted to develop a transitional scheme—NVL72x2 back-to-back rack architecture.

According to SemiAnalysis, the design concept of this scheme was to place two Oberon racks back to back, using pure copper NVLink to expand the scale domain, thus circumventing the manufacturing problems of the Kyber midplane.

However, this scheme ultimately did not come to fruition. SemiAnalysis stated that NVL72x2 "was canceled due to strong opposition from cloud service providers and hyperscale data center operators regarding its peculiar design and heavy operational burden."

With both routes blocked, NVIDIA is facing a phase of blankness in scaling up Rubin Ultra.

NVL576 Also Under Pressure, CPO Challenges Cannot Be Ignored

It is not just Kyber NVL144 that is delayed. SemiAnalysis also pointed out that NVL576— a larger scale system connecting 8 Oberon racks through CPO (Co-Packaged Optics)—"may also be delayed or limited to small-scale shipments due to challenges currently faced by CPO."

CPO is the optical interconnect technology NVIDIA introduced for scaling network expansion in the Rubin Ultra phase. According to a report previously released by SemiAnalysis in March 2026, the design plan for NVL576 is: keeping copper cable expansion inside the rack while connecting NVSwitch between racks via CPO, forming a two-layer fully interconnected network.

However, the maturity of CPO's mass production remains a variable. SemiAnalysis explicitly states in the report that the CPO NVSwitch will not be officially ready until the Feynman generation.

Rubin Ultra's Core Features Also Shrinking: 4-Chip Version Canceled

In conjunction with the aforementioned delay news, an important product-level change was also disclosed.

SemiAnalysis stated that the 4-chip version of Rubin Ultra has been canceled, "only retaining the smaller 2-chip version of Rubin Ultra, with actual performance about half that of the 4-chip version."

This means that even if the Kyber rack is ultimately delivered on schedule, the computing power ceiling for a single rack has been significantly lowered.

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

Competitive Window: AMD and Google May Benefit

The void in scaling domain directly impacts NVIDIA's competitive position in large-scale training scenarios.

SemiAnalysis pointed out: "NVIDIA currently has no verified solutions to expand Rubin Ultra's scaling domain, which leaves room for competitors like AMD MI500X or TPUv8i Broadfly to surpass Rubin Ultra in scaling capabilities."

According to NVIDIA's existing roadmap, CPO NVSwitch will not appear until the next-generation Feynman platform. Until then, the scaling upper limit of Rubin Ultra is constrained.

SemiAnalysis concludes the tweet by noting that the aforementioned delays and cancellations impact the memory, PCB, and ODM supply chains.

The manufacturing challenges of the Kyber midplane directly point to the technical bottlenecks of high-end PCB suppliers. The 78-layer ultra-high-density PCB, M9-level copper-clad laminate, and PTFE composite materials required for this midplane represent the current limits of PCB manufacturing processes.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。