Nvidia's Kyber NVL144 reportedly pushed back more than a year, Asian suppliers drop
Nvidia's Kyber NVL144 AI server system is delayed beyond 2028 due to critical manufacturing defects in the PCB midplane. Nvidia has scrapped the NVL72x2 rack design and the four-die Rubin Ultra chip, retaining only the smaller two-die version. The advanced CPO-NVSwitch interconnect technology is postponed until the subsequent Feynman generation, limiting near-term scalability. Asian supplier stocks (e.g., Ibiden, Kingboard Laminates) plummeted on the news, though analysts attribute this largely
Analysis
TL;DR
- Nvidia's Kyber NVL144 AI server system is delayed beyond 2028 due to critical manufacturing defects in the PCB midplane.
- Nvidia has scrapped the NVL72x2 rack design and the four-die Rubin Ultra chip, retaining only the smaller two-die version.
- The advanced CPO-NVSwitch interconnect technology is postponed until the subsequent Feynman generation, limiting near-term scalability.
- Asian supplier stocks (e.g., Ibiden, Kingboard Laminates) plummeted on the news, though analysts attribute this largely to profit-taking rather than reduced AI demand.
Why It Matters
This development highlights the physical manufacturing bottlenecks that are beginning to constrain the theoretical scaling of AI hardware, specifically in high-density interconnects. For investors and industry observers, it signals that while AI capital expenditure remains robust, Nvidia's monopoly on cutting-edge form factors is facing temporary erosion, creating a strategic window for competitors.
Technical Details
- PCB Manufacturing Challenges: The primary cause of the delay is the inability to produce defect-free PCB midplanes for the Kyber NVL144, which serve as the central interconnect for all components in the rack.
- Design Cancellations: The NVL72x2 configuration (two Oberon racks back-to-back) was abandoned due to operational overhead and cloud provider resistance. Additionally, the quad-die Rubin Ultra chip was canceled, leaving only the dual-die variant with approximately half the compute performance.
- Interconnect Delays: CPO-NVSwitch, a technology essential for linking multiple chips into large-scale systems, is deferred to the Feynman generation, meaning Nvidia currently lacks a proven method to scale the Rubin Ultra architecture effectively.
Industry Insight
- Competitive Opportunity: The delay and reduction in chip specs provide AMD (MI500X) and Google (TPUv8i Broadfly) with a critical timing advantage to capture market share in large-scale AI deployments.
- Supply Chain Volatility: Investors should anticipate continued volatility in Asian semiconductor supply chain stocks, as they remain highly sensitive to Nvidia's production roadmaps despite long-term demand fundamentals remaining strong.
- Focus on Existing Form Factors: Nvidia is likely to pivot revenue growth toward selling more of the existing Oberon-Rubin rack configurations in the interim, suggesting that near-term infrastructure upgrades may rely on volume rather than architectural leaps.
Disclaimer: The above content is generated by AI and is for reference only.