NVIDIA
2026-07-06
NVIDIA Kyber NVL144 Delayed to 2028: Midplane PCB Manufacturing Becomes AI Scaling Bottleneck
Summary
Key Takeaways
On July 6, 2026, SemiAnalysis disclosed that NVIDIA's Kyber NVL144 is delayed over 12 months to 2028, just three months after GTC 2026 showcase. The core bottleneck is the Orthogonal Backplane manufacturing—a 78-layer PCB (three 26-layer boards laminated) using M9-grade copper-clad laminate + quartz cloth + PTFE hybrid, with ≤25μm line width/spacing, required for 448G+ SerDes signal integrity. NVL144 connects 144 GPUs in a single domain; copper cables would require 20,000+ cables with 30% weight increase and severe signal loss. The interim NVL72x2 (back-to-back Oberon racks, copper NVLink extension) was cancelled due to hyperscaler objections over design complexity and operational burden. NVL576 (CPO connecting 8 Oberon racks) may be delayed or limited volume; CPO NVSwitch won't be ready until Feynman generation. The 4-die Rubin Ultra is also cancelled, leaving only a 2-die version with half performance. SemiAnalysis notes NVIDIA has no proven solution for Rubin Ultra scaling domain, opening a window for AMD MI500X or Google TPUv8i Broadfly. Current Rubin systems remain on track for H2 2026 deliveries to AWS/Azure/GCP. The event exposes physical manufacturing limits in NVIDIA's annual product cadence—the PCB midplane becomes the 'invisible ceiling' for AI infrastructure scaling.
Why It Matters
This event reveals NVIDIA's manufacturing weakness in defending against AMD and Google's scaling攻势. The Orthogonal Backplane is not a simple PCB upgrade but a lock-in mechanism—once deployed, customers must buy proprietary racks, power, and cooling. However, the 78-layer PCB yield and 448G+ SerDes integrity are nearly unsolvable at scale; NVIDIA downplayed the engineering nightmare of M9 laminate + Q-cloth + PTFE thermal expansion mismatch and impedance consistency. The cancellation of NVL72x2 exposes control plane shift issues: back-to-back rack design requires data center retrofitting, with operational burden severely underestimated. The 4-die Rubin Ultra cancellation indicates tail latency and inter-die bandwidth fell short; the 2-die version halves performance, undermining NVIDIA's TCO advantage in large clusters. For buyers, NVIDIA's annual cadence is unreliable—each architecture transition (Hopper→Blackwell→Rubin) faces delivery delays. The PCB midplane bottleneck is a physical ceiling, forcing customers to consider AMD's MI500X Infinity Fabric or Google's TPUv8i Broadfly in 2027-2028.
PRO Decision
【Vendors (AMD, Google, Intel)】Exploit NVIDIA's 18-month window. AMD should highlight MI500X modular rack design avoiding complex Orthogonal Backplanes, offering lower TCO and faster deployment. Google should promote TPUv8i Broadfly CPO maturity vs NVIDIA's PCB risk. Intel can partner with white-box ODM to deliver open rack solutions based on UEC standards, breaking NVLink physical lock-in. 【Enterprises (CIO/Architects)】Audit NVIDIA roadmap dependency: assess if current Rubin deployment can be delayed to 2028; if not, launch POC for AMD MI500X or Google TPUv8i. Demand independent third-party yield test reports for the Orthogonal Backplane and 448G+ SerDes signal integrity data. Include delay penalty clauses in contracts and reserve 30% of compute budget for non-NVIDIA solutions to mitigate supplier concentration risk. 【Investors】This event exposes unsustainability of NVIDIA's annual cadence. The PCB midplane bottleneck is a physical constraint, not short-term fixable. Watch for NVIDIA lowering 2027 revenue guidance. Short NVIDIA near-term, long AMD and Google AI hardware. Long-term, NVIDIA valuation must shift from 'monopoly shovel seller' to 'competitive hardware vendor facing manufacturing ceilings'.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)