NVIDIA’s new GB200 AI server line is gearing up for production, with shipments expected to begin in mid-September for the NVL36 model and late October for the higher-end NVL72. These servers are crucial to NVIDIA’s AI strategy, with a projected 80% of the 5 million Blackwell chips produced in 2025 destined for GB200 systems. Each GB200 AI GPU carries a hefty price tag of around $30,000, while the GB200 Superchip (CPU + GPU) costs over $70,000. The NVL72 AI server cabinet, boasting 72 x B200 GPUs and 36 x Grace CPUs, is priced at approximately $3 million per unit. The NVL36 model features 36 x B200 AI GPUs and 16 x Grace CPUs.
NVIDIA’s move towards liquid-cooling for its GB200 servers has increased complexity and costs. Taiwan, home to NVIDIA’s production facilities, is experiencing a surge in demand for AI components, with Taiwanese companies working tirelessly to meet the growing needs. However, recent challenges have arisen due to liquid-cooling leaks, leading to delays in GB200 shipments.
Despite these hurdles, demand for AI servers remains strong, with volume shipments expected to ramp up in January 2025. The GB200’s performance and advanced liquid-cooling system are key to meeting the growing demands of the AI landscape, but the challenges of implementation will be a key factor in determining the success of this crucial product line.
