SoftBank has deployed what it claims to be the world’s largest Nvidia DGX B200 SuperPod cluster.

Comprising more than 4,000 Nvidia Blackwell GPUs, SoftBank said the system has been scaled with Nvidia Quantum-2 InfiniBand networking.

In total, the AI compute platform built and deployed by SoftBank for its SB Intuitions Corp. subsidiary now consists of more than 10,000 GPUs, delivering 13.7 exaflops of combined AI compute performance.

The computing platform, which began operations in September 2023, was initially deployed with 2,000 Nvidia Ampere GPUs, offering 700 petaflops of AI compute power. Just over a year later, SoftBank announced it had installed an additional 4,000 Nvidia H100 GPUs into the cluster, bringing the total number of GPUs to 6,000 and its performance to approximately 4.7 exaflops.

Launched in August 2023, SB Intuitions is dedicated to developing Japanese large-language models (LLMs). SB Intuitions plans to offer its 70 billion parameter commercial model, Sarashina mini, by the end of March 2026, while its AI compute platform will be made available via an infrastructure-as-a-service model to companies and research institutions in Japan.

First launched in March 2024, an Nvidia DGX BB200 SuperPod is a cluster of eight or more DGX B200s, the sixth generation of the chipmaker’s air-cooled, traditional rack-mounted DGX design that includes eight Nvidia B200 Tensor Core GPUs and two 5th Gen Intel Xeon CPUs.

SoftBank was the first company to receive Nvidia DGX B200 systems, which it used to build its DGX B200 SuperPod. That announcement, from November 2024, formed part of a wider partnership unveiled by the two companies, with other Japan-based initiatives including the development of an AI supercomputer based on the chip designer’s Grace Blackwell platform, and a 5G AI-RAN network.