Summary of Key Points
This news report focuses on the delays in NVIDIA's Kyber rack architecture: The Kyber rack, originally designed for the 2027 Rubin Ultra chips (a solution that integrates 144 chips into a single supercomputer cabinet), is reportedly being postponed by more than 12 months to 2028 due to difficulties in manufacturing the critical printed circuit boards (PCBs). NVIDIA's alternative approach of stacking two cabinets back to back was rejected by cloud service providers due to its unusual design and high costs, leaving no mature solution available to expand the Rubin Ultra computing cluster for the time being. This situation presents opportunities for AMD and Google. In the market, A-share companies related to PCB manufacturing saw initial declines followed by gains; NVIDIA's stock price in the US market experienced only slight fluctuations. Both the company and experts suggest that there's no need for excessive concern, as NVIDIA has experience in dealing with supply chain challenges.
1. The Core Issue: Delays in the Kyber Rack Architecture Due to PCB Manufacturing
What is the Kyber rack architecture? In simple terms, it involves combining 144 NVIDIA chips within a single server cabinet to create a supercomputing system, specifically designed for the Rubin Ultra chips expected to be released in 2027. However, there are now indications that the launch will be delayed until 2028, mainly due to challenges in PCB manufacturing. The PCBs act as the bridge that connects the chips to other components, and their production proved more difficult than anticipated, raising doubts about the feasibility of this approach. Another rack design, called NVL576, may also face delays, and if it does get released, it will likely be available in limited quantities only.
2. The Alternative Approach Fails, Leaving Competitors with Opportunities
NVIDIA had prepared a backup plan: stacking two cabinets back to back to increase computing power. However, this option was dismissed by cloud service providers such as Alibaba Cloud and Amazon Web Services due to its awkward design and higher operational costs (e.g., increased energy consumption and cooling requirements). This means NVIDIA currently lacks a mature method to build large-scale Rubin Ultra computing clusters, giving AMD (NVIDIA's chip competitor) and Google (which develops its own AI chips and provides computing services) an opportunity to gain market share.
3. Market Reactions
After the news was released, A-share companies involved in PCB manufacturing saw declines on Monday: Guanghua Technology fell 5.16%, Nuode Co., Ltd. fell 8.63%, Tianhai Electronics fell 6.77%, and Weilao Group fell 8.32% as investors feared that the Kyber rack delays would reduce PCB orders. By Tuesday, these stocks rebounded, with Guanghua Technology rising more than 4% and Nuode Co., Ltd. rising over 6%. The reason for this reversal could be that markets believed the issue was not permanent and that NVIDIA would be able to resolve it. NVIDIA's stock price in the US market fell 1.39% on Monday and then rose 0.37% on Tuesday, indicating that investors' long-term confidence in the company remained relatively stable.
4. Company and Expert Responses: No Need for Excessive Anxiety; NVIDIA Has Experience
In response to the delays, NVIDIA stated that its product roadmap remains intact, suggesting that the overall plan is not disrupted. Experts from consulting firms also advised against overreacting to the delays, noting that NVIDIA has previously overcome similar technical issues through collaboration with partners (such as PCB manufacturers). NVIDIA CEO Jensen Huang has mentioned supply chain bottlenecks, emphasizing that the rapid growth in AI demand has created widespread challenges. They have made advance plans for the supply chain but anticipate that demand may exceed expectations. The underlying message is that the delays are temporary, and NVIDIA is capable of addressing them.
5. The Big Picture: The Competition in Computing Power Enters a New Phase
This incident highlights a new trend in the AI computing power race: having chips alone is no longer enough; more efficient rack architectures are needed to accommodate large-scale AI models. NVIDIA has been focusing on integrated solutions (chips + cabinets + software), but the PCB issues highlight the complexity of hardware integration. For ordinary users, there's no need for excessive worry, as NVIDIA possesses strong technical capabilities and a robust supply chain. Whether competitors can seize this opportunity will depend on their ability to develop comparable products.