OpenAI receives early access to Nvidia Blackwell DGX B200 systems ahead of general availability.

OpenAI has obtained one of the first NVIDIA Blackwell DGX B200 systems, marking a significant leap for artificial intelligence. The company shared a photo of its staff with an early DGX B200 engineering sample via its official X handle, signaling readiness to train its advanced AI models with this cutting-edge hardware. NVIDIA's B200 GPUs, which use the Blackwell architecture, are the firm's fastest data center GPUs to date, and orders have begun rolling in from multinational corporations seeking to boost their AI computing capabilities.

The DGX B200 is an all-in-one AI platform designed for training, fine-tuning, and inference with eight B200 GPUs per unit. Each DGX B200 can provide GPU memory of up to 1.4 TB, backed by a maximum HBM3E memory bandwidth of 64 TB/s. According to NVIDIA, the DGX B200 is capable of delivering training speeds of up to 72 petaFLOPS and inference speeds of up to 144 petaFLOPS, making it well-suited for training LLMs, chatbots, and recommender systems.

Blackwell GPUs have long captured the interest of industry leaders. OpenAI's CEO Sam Altman even hinted at the possibility of employing them to train their AI models. Major global tech giants including Amazon, Google, Meta, Microsoft, Tesla, xAI, and Dell Technologies have already committed to using Blackwell GPUs. xAI intends to use 50,000 B200 GPUs in addition to its existing 100,000 H100 GPUs, while Foxconn has announced plans to construct the fastest supercomputer in Taiwan using B200 GPUs.

Compared to NVIDIA's Hopper GPUs, Blackwell is both more powerful and more power efficient, making it an ideal choice for OpenAI's AI model training needs. The DGX B200 boasts three times the training performance and fifteen times the inference performance of earlier generations, positioning it as a transformative technology for advancing AI capabilities across the industry.