Nvidia Kyber rack systems delayed to 2028 amid manufacturing challenges
Nvidia is facing reports of operational delays regarding its next-generation Kyber rack-scale architecture, with production now pushed to 2028. Despite this setback, the firm remains a central focus for large-scale AI infrastructure investments by major hyperscalers.
Key Takeaways
- Kyber architecture is designed to house 144 Rubin Ultra chips in a single unified computing system.
- The delay stems from technical challenges in manufacturing the 78-layer PCB midplane (orthogonal backplane).
- Hyperscalers reportedly rejected a stopgap 'NVL72x2' dual-rack design due to operational complexity and costs.
- SemiAnalysis notes the delay may create a high-end market opening for competitors like AMD and Google.
- Nvidia shares lagged the broader chip sector's 4.3% rally on Monday, gaining only 0.5%.
Why It Matters
The delay of Kyber suggests that Nvidia's aggressive annual product cadence is colliding with the physical limits of rack-level systems engineering. While silicon roadmaps for individual chips like Rubin Ultra may remain intact, the failure to deliver the accompanying rack-scale interconnect at yield means customers may face a performance ceiling in 2027. This technical stumble is particularly significant for the streaming video industry, as hyperscale providers rely on these high-density systems for the massive inference workloads required by generative AI and agentic video applications. Watch for whether Nvidia officially updates its public roadmap or if competitors AMD and Google secure incremental design wins at major cloud providers in the 2027 window.
Additional Context
The reporting of the Kyber delay comes at a friction point in the semiconductor market where investor focus is shifting from pure GPU demand toward a broader infrastructure mix. Per Goldman Sachs (May 2026), while Nvidia's data center revenue continues to exceed consensus—reaching over $75 billion in a single quarter—the market is increasingly pricing in growth from memory, custom AI silicon, and standalone CPUs. Nvidia CEO Jensen Huang has identified a secondary $200 billion total addressable market for the company’s new 'Vera' CPUs, projecting nearly $20 billion in standalone CPU revenue for 2026 as agentic AI shifts the workload balance from model training to inference (per Investing.com, May 2026). Despite the reported Kyber setback, Nvidia officially maintains that its broader product roadmap remains 'intact' (per Tom's Hardware, July 2026). This follows a string of upward revisions to the company's long-term guidance, with Huang pegging cumulative AI compute demand at $1 trillion between 2025 and 2027. However, external analysis by SemiAnalysis suggests the PCB bottleneck is among the most complex engineering hurdles in commercial computing. The 78-layer board requires maintaining signal integrity and thermal stability across 144 vertically mounted GPUs—a density that is reportedly straining the current capability of global PCB suppliers like Japan's Ibiden (per 24/7 Wall St, July 2026). Competitors are positioning themselves to capitalize on any perceived execution risks. AMD has notably seen a shift in valuation, trading at a premium multiple relative to Nvidia by mid-2026 as it gains traction in sixth-generation EPYC server CPU deployments with Meta and Oracle (per 24/7 Wall St, July 2026). As hyperscalers like Alphabet and Microsoft forecasted combined 2026 capital expenditures exceeding $700 billion, any delay in Nvidia's high-end rack availability could accelerate the adoption of custom-designed silicon or alternative platforms better aligned with 2027 capacity schedules.
Read full article at tr.tradingview.com
Get this in your inbox → Subscribe
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source