Renowned analyst Ming-Chi Kuo reports that NVIDIA (NVDA, Financial) will no longer offer the dual-cabinet version of its GB200 (NVL36×2) AI server without custom orders. Instead, the company will provide only the single-cabinet versions: GB200 NVL72 and NVL36. This adjustment is attributed to limited resources and customer preferences.
Kuo notes that while this decision will not impact the long-term trends for artificial intelligence or NVIDIA, it has raised concerns over the company's and its supply chain's execution capabilities. He explains that NVIDIA's frequent changes to its AI server product roadmap reflect efforts to balance supply chain execution, competitive advantage, and customer demand, citing the discontinuation of NVL36×2 as an example. This approach showcases a more pragmatic attitude towards product planning but may lead to confusion among some market participants regarding 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: NVL36, NVL72, and NVL36×2. However, managing all three projects, especially the development of two different 72 GPU versions (NVL72 and NVL36×2), proved challenging. Consequently, NVIDIA now focuses solely on NVL72 and NVL36.
NVIDIA's GB200 NVL72 rack consists of 18 compute trays and 9 NVSwitch trays (18 NVSwitch ASICs), with each tray containing two Bianca boards, each equipped with a Grace CPU and two Blackwell GPUs. This is the company's most powerful product, consuming approximately 120kW of power.
Previously, Morgan Stanley analysts indicated that NVIDIA was more inclined to sell AI server cabinets featuring Blackwell GPUs. The NVL36 comes with 36 B200 GPUs, priced between $1.8 million and $2 million, while the NVL72 with 72 B200 GPUs starts at $3 million. Compared to GPUs, GPU modules, and even DGX and HGX servers, cabinets offer significantly higher profit margins. In this scenario, NVIDIA doesn't need to supply 450,000 GPUs to generate $10 billion in revenue.
