On November 13, at the 2024 AI Summit in Japan, NVIDIA and SoftBank announced a series of cooperation plans aimed at accelerating the development of Japan’s sovereign AI project (Sovereign AI, national-level AI infrastructure) and strengthening NVIDIA’s Global technology leadership while bringing billions of dollars in new revenue to the global telecommunications industry.
NVIDIA founder and CEO Jensen Huang revealed in his keynote speech that SoftBank is using the Blackwell platform to build Japan's most powerful AI supercomputer and plans to upgrade to the Grace Blackwell platform in the future.
At the same time, NVIDIA announced that SoftBank has successfully used the NVIDIA AI Aerial platform to achieve the world's first integration test of AI and 5G telecom networks. This breakthrough will open up billions of dollars in new revenue channels for telecom operators.
In addition, NVIDIA and SoftBank also jointly announced that they will use NVIDIA AI Enterprise software to create a market that meets Japan’s local secure AI computing needs. This new service will support AI training, edge reasoning and other functions, helping SoftBank become the core of Japan's AI grid and create, distribute and use new business opportunities for artificial intelligence services for various industries, consumers and enterprises.
Huang Renxun said that this cooperation with SoftBank will push Japan into the AI industrial revolution and lead a new era of growth in industries such as telecommunications, transportation, robotics, and healthcare.
Junichi Miyagawa, President and CEO of SoftBank, emphasized that the world is accelerating the adoption of AI to promote social development, and the long-term cooperation between SoftBank and Nvidia will assist this transformation. With its powerful AI infrastructure and innovative distributed AI-RAN solution "AITRAS", SoftBank will reshape 5G networks and accelerate the pace of innovation worldwide.
SoftBank is about to receive the world's first NVIDIA DGX B200 system, which will become the core building block of its new NVIDIA DGX SuperPOD supercomputer.
SoftBank intends to use this Blackwell-powered DGX SuperPOD to not only promote its own generative AI research and development and AI-related businesses, but also support the AI needs of Japanese universities, research institutions and enterprises.
It is expected that after completion, SoftBank's DGX SuperPOD will become Japan's most outstanding supercomputing system. It comes with NVIDIA AI Enterprise software and NVIDIA Quantum-2 InfiniBand networking, which is particularly suitable for developing large language models.
In addition to DGX SuperPOD, SoftBank is also planning another supercomputer focused on highly computationally intensive tasks. The computer is initially planned to be built based on the NVIDIA Grace Blackwell platform and integrate the NVIDIA GB200 NVL72 multi-node, liquid-cooled, rack-level system, perfectly combining the NVIDIA Blackwell GPU with the efficient Arm architecture NVIDIA Grace CPU.
SoftBank has worked closely with Nvidia to achieve a technological milestone - the development of a new telecommunications network that can take into account both AI and 5G workloads, which the industry calls AI Radio Access Network (AI-RAN). This innovative infrastructure is widely favored by the telecom industry ecosystem because it helps operators transform base stations from cost burdens into AI revenue-generating resources.
In an outdoor trial in Kanagawa Prefecture, Japan, SoftBank verified that its AI-RAN solution based on NVIDIA acceleration has achieved operator-level 5G performance and can smoothly run AI inference tasks while utilizing the remaining capacity of the network.
Traditional telecommunications networks are designed to handle peak hours, but average utilization is only one-third. The general computing capabilities of AI-RAN are expected to give telecom companies the opportunity to convert the remaining two-thirds of capacity into AI inference services for commercialization.
According to estimates by NVIDIA and SoftBank, telecom operators are expected to obtain approximately US$5 in AI inference revenue for every US$1 of capital expenditure they invest in new AI-RAN infrastructure. Taking into account operating costs and capital expenditures, SoftBank predicts that for each additional AI-RAN server, the return rate on its infrastructure can reach up to 219%.