Transition Plan Fails and Chip Specs Downgraded? SemiAnalysis Claims Nvidia Architecture Is Delayed, Company Urgently Denies Rumors.

TradingKey - NVIDIA ( NVDA) quickly responded to debunk rumors of a delay in its next-generation AI server architecture, stating that the "product roadmap remains unchanged."

On July 6, Eastern Time, semiconductor research firm SemiAnalysis suddenly posted six consecutive tweets on the X platform, pouring cold water on the booming AI chip market. The firm revealed that Kyber NVL144, the next-generation AI server rack architecture heavily featured by NVIDIA at its GTC conference in March this year, will be forced to delay for at least 12 months due to manufacturing bottlenecks for key components, potentially pushing back its original 2027 launch plan to 2028.

SemiAnalysis pointed out that the core reason for the delay of Kyber NVL144 lies in the insurmountable challenges faced in the manufacturing process of a PCB orthogonal backplane required for its unique design.

This midplane, officially referred to by NVIDIA as the "Midplane," is the key component for achieving a 90° vertical interconnection between compute trays and switch trays, allowing a single rack to integrate 144 GPUs and doubling the computing power density compared to the current architecture.

However, to meet signal integrity requirements under the 448G+ SerDes rate, this backplane adopts extremely complex technical specifications. Technical analysis shows that if traditional copper cable solutions were used to connect 144 GPUs, more than 20,000 cables would be required. This would not only increase the weight by over 30%, but the resulting signal attenuation would also severely impact performance.

Therefore, this seemingly ordinary circuit board is actually the only viable solution to achieve the design objectives of the Kyber architecture under current technological conditions.

Alternative Options Fail One After Another

Faced with Kyber's manufacturing challenges, NVIDIA had attempted to develop a transitional solution—the NVL72x2 back-to-back rack architecture—to bypass midplane manufacturing difficulties by placing two Oberon racks back-to-back.

However, SemiAnalysis revealed that this solution met with strong opposition from cloud service providers due to the excessive operational and maintenance burden, and was ultimately cancelled. To make matters worse, the NVL576—a larger-scale system connecting eight Oberon racks via CPO—may also face delays or limited low-volume shipments due to the insufficient maturity of CPO technology.

Adding to market concerns, SemiAnalysis also disclosed that NVIDIA has cancelled the 4-compute-chip version of Rubin Ultra, retaining only the 2-chip version with halved performance.

This means that even if the Kyber racks are delivered on schedule, the single-rack computing power ceiling has been significantly lowered, and NVIDIA currently has no proven solution to fill this scale-out gap.

Nvidia Issues Urgent Clarification, Stirring Market Sentiment

In response to the rumors, an Nvidia spokesperson quickly replied that day, stating that the "product roadmap remains unchanged," though no details regarding specific project timelines were disclosed. Although the official statement stabilized the confidence of some investors, market sentiment has clearly come under pressure.

On July 6, Eastern Time, Nvidia's stock price closed up 0.37% after experiencing volatile intraday trading. However, the stock prices of PCB manufacturers in the Asian market reacted more violently, with supply chain companies such as Ibiden in Japan, Kingboard Laminates in Hong Kong, China, and Taiwan Union Technology in Taiwan, China, all falling by more than 10% in a single day.

Mizuho Securities analyst Jordan Klein believes that such delay rumors have recurred in recent years and are more like "noise to attract attention."

However, some industry insiders pointed out that this turmoil reveals that AI hardware competition has entered a system-level battle, and the coordination complexity among chip design, packaging, thermal management, and server architecture is rising exponentially.

Nvidia needs to find a balance between technological innovation and engineering implementation to maintain its leading position in the AI data center infrastructure sector.

The BofA Securities Taiwan Technology Team pointed out in a research report that although market anxiety has risen, the nature of this event is a reduction in demand volume under supply constraints, rather than a reversal of the demand trend.

The team believes that basic materials such as high-end CCL and PCB account for only a single-digit percentage of the overall cost of AI servers, and short-term delays will have a limited impact on the long-term development of the industry.

More importantly, the supply gap for these high-end materials is unlikely to ease until at least the end of 2027. Long equipment delivery cycles dictate that the pace of supply-side capacity expansion can hardly catch up with demand growth. Therefore, BofA views this stock price volatility as an "enhanced buying opportunity" and maintains a buy rating on the relevant sectors.