Renowned analyst Ming-Chi Kuo stated that, in the absence of customer customization requirements, $NVIDIA (NVDA.US)$ the dual-cabinet version GB200 (NVL36×2) AI server will no longer be provided, only the single-cabinet versions GB200 NVL72 and NVL36. The reason for this move by Nvidia is reportedly due to limited resources and customer preferences.

Guo Mingxi added that this move will not affect artificial intelligence or Nvidia's long-term trends, but it does bring into question the company's and its supply chain's execution capabilities.

Guo Mingxi explained, "In my view, Nvidia's recent frequent changes to the AI server product roadmap reflect their attempt to achieve a better balance between supply chain execution, competitive advantage, and customer demand under limited resources (stopping the development of NVL36×2 is an example)." "This is a good thing, indicating that Nvidia is taking a more pragmatic approach to product planning, but this shift may confuse some market participants about supply chain changes."

Initially, Nvidia planned to develop three AI server cabinets based on the Blackwell GPU GB200, for AI and high-performance computing (HPC) workloads, namely NVL36, NVL72, and NVL36×2. However, managing all three projects became challenging, especially when developing two different 72 GPU versions simultaneously (NVL72 and NVL36×2). Therefore, Nvidia is now focusing only on NVL72 and NVL36.

Nvidia's GB200 NVL72 rack includes 18 compute trays and 9 NVSwitch trays (18 NVSwitch ASICs), each tray has two Bianca boards, each board has one Grace CPU and two Blackwell GPUs. This is Nvidia's most powerful product, but also the highest power-consuming solution, consuming about 120kW.

Previously, Morgan Stanley analysts pointed out that Nvidia is more willing to sell AI server cabinets equipped with the Blackwell GPU. NVL36 is equipped with 36 B200 GPUs, expected to be priced between 1.8 million to 2 million dollars, while NVL72 equipped with 72 B200 GPUs is expected to start at 3 million dollars. Selling cabinets is much more profitable compared to GPUs, GPU modules, or even DGX and HGX servers. In this scenario, Nvidia can earn 10 billion dollars without needing to provide 450,000 GPUs.

Editor/Lambor