NVIDIA CEO Jensen announced that engineering samples of the Blackwell GPUs for AI applications will start shipping globally this week. NVIDIA plans to launch products featuring the Blackwell architecture in the fourth quarter of this year.

This announcement was made at Siggraph. NVIDIA’s hardware partners, including Foxconn, Quanta, Wistron, Pegatron, and Asus, have already showcased Blackwell-based servers at this year’s Computex in Taipei, indicating they may have had access to Blackwell GPUs for some time. However, not all software developers have access to NVIDIA's latest Blackwell GPUs yet.

NVIDIA and its partners will charge a premium for each AI server cabinet equipped with Blackwell GPUs. NVIDIA has introduced two reference server cabinets: the NVL36, equipped with 36 B200 GPUs, priced at approximately USD 2 million, and the NVL72, equipped with 72 B200 GPUs, priced at USD 3 million.

These cabinets will be offered by traditional partners such as NVIDIA, Foxconn, Quanta, and Wistron, along with new entrants like Asus. NVIDIA projects the shipment of 60,000 to 70,000 B200 server cabinets next year, potentially generating at least USD 210 billion in revenue. It's widely anticipated that major firms like AWS, Dell, Google, Meta, and Microsoft will adopt NVIDIA's latest Blackwell GPUs.