Jim Cramer has doubled down on Nvidia stock, urging investors to buy the chipmaker as it disputes claims of AI rack delays to 2028. According to CNBC, Cramer told investors that chip stocks are staging a 'revenge trade' after last week's 'misguided selling', with the Philadelphia Semiconductor Index gaining 87.8% in the second quarter - its best quarter since records began in 1994. Nvidia's stock traded near $196.58 at the time of reporting, up almost 2% over the last 24 hours. The market reaction came after Japan's Ibiden, Nvidia's largest client, fell as much as 10% on Monday, while Kingboard Laminates tumbled 18% in Hong Kong and Samsung Electro-Mechanics slid 11% in Seoul following the SemiAnalysis report.
Nvidia has pushed back the release date for its Kyber rack-scale architecture to 2028, representing an over-year delay from its originally planned 2027 release. According to reports from SemiAnalysis, the delay is attributed to manufacturing challenges with the PCB midplane, which remains difficult from a manufacturability standpoint. This key printed circuit board serves as a central component that connects electronic modules inside the system, with SemiAnalysis noting that Nvidia cannot yet build it at scale. The delay creates a significant gap in Nvidia's product roadmap, as the company typically releases new architectures annually. However, Nvidia has rejected the claims, telling media outlets that its roadmap remains intact and fresh off launching a revenue-sharing compute program for AI startups.
The Kyber rack represents a significant advancement in AI computing infrastructure, aggregating 144 of Nvidia's more proficient AI chips into a single powerful computer system. As reported by SemiAnalysis, the rack is designed with graphics processing units placed in a vertical manner rather than horizontally to improve density and reduce latency. This configuration provides sufficient computing power for AI companies to train and operate their cutting-edge models effectively, with the density being critical for training and running the largest AI models. The orthogonal backplane that connects Kyber's vertically mounted compute trays presents significant technical challenges, with trade analyses describing three 26-layer sections laminated into one 78-layer stack close to a square meter in area. The board features trace spacing at or below 25μm and impedance held within a tolerance of 5% to maintain 448 Gb/s-class signaling integrity.
The delay affects not only the Kyber NVL144 rack architecture but also the NVL576 system, which couples eight modules together through optical connections. According to SemiAnalysis reports, Nvidia is facing challenges in matching its rapid release schedules with manufacturing capabilities. The company had previously considered merging two current-generation racks together to achieve comparable computing power, but cloud customers pushed back hard on the 'odd design and heavy operational burden,' and Nvidia scrapped the idea. Additionally, Nvidia has killed NVL72x2, a stopgap rack designed to tide customers over, and no proven alternative is now available to widen Rubin Ultra's scale-up in 2027. This leaves Nvidia with 'no proven solution' to scale up its most powerful Rubin Ultra systems, creating a rare technical opening for competitors.
The orthogonal backplane that connects Kyber's vertically mounted compute trays presents significant technical challenges, with trade analyses describing three 26-layer sections laminated into one 78-layer stack close to a square meter in area. The board features trace spacing at or below 25μm and impedance held within a tolerance of 5% to maintain 448 Gb/s-class signaling integrity. SemiAnalysis notes that NVL72x2 would have bolted two Oberon racks back-to-back to reach Kyber-class density, but Nvidia abandoned this solution after its largest customers balked at running two linked cabinets as a single unit. The delay provides a significant opportunity for competitors, as SemiAnalysis notes that Google and AMD already win work from top AI labs with their own chips, with the stumble at the very high end handing competitors a rare technical opening. Nvidia's next earnings report will show whether rack-level friction reaches data center revenue.
