Nvidia Shifts Gears: Focus on Sleek AI Servers Amid Resource Constraints

Nvidia has recently decided to halt the development of its GB200 NVL36×2 dual-rack AI server, as reported by renowned analyst Ming-Chi Kuo. This shift allows the company to concentrate its resources on the NVL72 and NVL36 single-rack models. The decision is reportedly due to limited resources and customer preferences, despite earlier indications that the NVL36×2 was highly favored by clients.

The original development plan included three Blackwell GPU-based models for AI and high-performance computing workloads: the NVL36, NVL72, and NVL36×2. However, managing all these projects proved challenging, particularly with the dual 72 GPU versions. Hence, Nvidia is now focusing solely on the NVL72 and NVL36.

The GB200 NVL72 rack is built with 18 compute trays and 9 NVSwitch trays, featuring two Bianca boards per tray, each with a Grace CPU and two Blackwell GPUs. This configuration, while incredibly powerful, comes with an energy demand of approximately 120kW, making it the most power-intensive solution Nvidia offers.

Despite most data centers being unable to support the NVL72's extreme power and cooling needs, prominent customers like Microsoft reportedly favor this option over the NVL36×2. The latter was originally designed as a dual-rack system with 18 Grace CPUs and 36 Blackwell GPUs per rack, allowing for full interconnectivity between 72 GPUs. However, its 132kW power consumption and larger size pose practical challenges.

The shifting focus to the NVL72 underscores Nvidia's intention to maximize performance for its most demanding clients. Simultaneously, continued adjustments in their AI server product roadmap highlight a pragmatic approach in navigating resource limitations while balancing supply chain execution, competitive advantage, and customer needs.

This strategic pivot, while raising questions about execution capabilities, does not impact Nvidia's long-term AI trajectory. Adjusted product launches reflect an evolving strategy aimed at optimizing performance and meeting the intricate demands of major tech players.