NVIDIA
2026-07-07
NVIDIA Denies Kyber NVL144 Delay, But 78-Layer PCB Bottleneck Exposes AI Hardware Physics Limit
Summary
Key Takeaways
SemiAnalysis reports that NVIDIA's next-gen flagship AI rack system, Kyber NVL144, faces a potential delay to 2028 due to manufacturing bottlenecks in its 78-layer ultra-high-density PCB midplane. This midplane uses M9-grade copper-clad laminate, quartz fabric, and PTFE hybrid materials, laminated from three 26-layer boards, with line width/spacing ≤25μm to support 448G+ SerDes signal integrity.
The report also reveals cancellations: the NVL72x2 back-to-back rack architecture was scrapped due to hyperscaler opposition over “odd design and high maintenance burden”; the 4-chip Rubin Ultra was canceled, leaving only a 2-chip version with half the performance; and the NVL576 system using CPO (Co-Packaged Optics) to connect 8 racks may be delayed or limited to small batches. SemiAnalysis argues NVIDIA lacks a proven scaling solution for Rubin Ultra, opening a window for AMD MI500X and Google TPUv8i Broadfly.
NVIDIA officially responded, stating “Our roadmap is intact,” but did not refute the process-level claims. The current Vera Rubin NVL72 rack system is on track for mass production and delivery to top cloud providers in H2 2026.
Why It Matters
SemiAnalysis's leak exposes a physical scaling limit in NVIDIA's next-gen AI hardware, not a mere supply chain issue. The 78-layer PCB midplane for Kyber NVL144 is a forced material and process stack-up to meet 448G+ SerDes signal integrity, revealing a fundamental conflict between high-speed signal transmission and physical manufacturing capability.
NVIDIA's vague “roadmap intact” denial, without refuting specific process challenges, is a strategic ambiguity. The real risk: if 78-layer PCB yield isn't resolved by 2027, a 12+ month product gap in the trillion-parameter model training market emerges.
The cancellation of the 4-chip Rubin Ultra means NVIDIA's planned “doubling of per-node compute density” path is blocked, retreating to a 2-chip solution. This opens a substantive window for AMD MI500X (with stronger chiplet scalability) and Google TPUv8i Broadfly (optimized for large-scale distributed training).
Furthermore, the potential delay of the NVL576 CPO solution suggests NVIDIA's Co-Packaged Optics maturity may be lower than expected, potentially impacting its inter-rack bandwidth and latency advantages in ultra-large GPU clusters, where tail latency and congestion control become critical.
PRO Decision
【Vendors (AMD, Google, White-box AI hardware camp)】Immediately exploit the potential Kyber NVL144 delay window. Pitch AMD MI500X's Chiplet architecture scalability and mature Infinity Fabric to hyperscalers, emphasizing no need for extreme 78-layer PCB. Accelerate Google TPUv8i Broadfly external supply, highlighting tail latency optimization and proprietary optical interconnects for large-scale training. White-box vendors should offer OCP-standard modular rack solutions to avoid single-vendor manufacturing bottlenecks.
【Enterprises (CIOs & Architects)】Launch a supplier concentration risk audit immediately. Assess dependency on Kyber NVL144 in GPU procurement plans. Demand independent third-party yield and reliability test reports for the 78-layer PCB midplane from NVIDIA, and a detailed technical whitepaper on the Rubin Ultra cancellation and replacement scaling solution. Include milestone-based delivery clauses in contracts, allowing a switch to AMD or Google alternatives if Kyber NVL144 delays exceed 6 months. Plan a multi-architecture training stack (e.g., PyTorch + JAX) for seamless workload portability.
【Investors】Treat SemiAnalysis's leak as a material risk signal, not “pure noise.” Track NVIDIA's progress on physical manufacturing bottlenecks (ultra-high-density PCB, CPO) in next-gen roadmaps, not just PR statements. Reduce NVIDIA positions, increase AMD and Google, and short supply chain firms like Ibiden that may face order delays from process challenges. Long-term, monitor whether Chiplet architecture and optical interconnect adoption accelerates due to NVIDIA's predicament, as this will be the core AI hardware investment theme for the next 3-5 years.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)