LOGO

The “Stargate Project,” a large-scale AI infrastructure development initiative spearheaded by SoftBank Group and OpenAI, is focusing on Sakai City as the central location for its expansion in Japan. Specifically, SoftBank plans to repurpose a former liquid crystal display panel factory owned by Sharp in Sakai. The company has acquired a portion of this facility for approximately 100 billion yen with the goal of transforming it into a cutting-edge AI data center.

This facility will be the third major site for the project, following an existing base in Tokyo and another under construction in Hokkaido. It boasts an impressive power capacity of 150 megawatts, making it one of the largest in Japan. Operations are slated to begin in 2026, with plans to expand capacity to 250 megawatts in the future. Sakai’s favorable location and infrastructure conditions are expected to ensure the long-term stability of the data center’s operations.

SB OpenAI Japan to Drive Domestic AI Development and Adoption

At the heart of this project is “SB OpenAI Japan,” a joint venture established in February 2025 by SoftBank and OpenAI. This company aims to develop large language models (LLMs) specifically tailored for the Japanese language and provide “Crystal Intelligence,” a generative AI service for businesses.

The Sakai data center is planned to host the operation of AI agents powered by GPUs, utilizing the foundational models provided by OpenAI. These agents will be specialized for various corporate functions, such as human resources and marketing, with the aim of delivering customized AI solutions that meet specific business needs.

These efforts have the potential to significantly accelerate the digital transformation of Japanese companies.

Creating the Future Through Massive Investment and Industrial Fusion

SoftBank is planning a large-scale development that will require 100,000 GPUs for this AI infrastructure build-out, potentially amounting to a massive investment approaching one trillion yen based on simple calculations. The GPUs are expected to be supplied by U.S.-based NVIDIA and the Stargate Project itself.

SoftBank President Miyakawa stated, “We aim to make Sakai a hub for the fusion of AI and existing industries, serving as an experimental ground for new business models and solutions to challenges.” This highlights the expectation that the facility will not just be a data center, but a key driver in the evolution of the AI industry both domestically and internationally.

Furthermore, this initiative is poised to be a crucial step in enhancing productivity across various industries and addressing labor shortages.

In March 2025, U.S.-based NVIDIA held its annual developer conference, “GTC,” and announced its new software “Dynamo,” specifically designed for inference processing. This announcement comes against the backdrop of a significant shift in AI’s evolution, moving from a primary focus on “learning” to “inference.”

NVIDIA, a company that has historically excelled in technologies for training AI models, emphasized that its hardware and software are now essential for inference as well. CEO Jensen Huang stressed that accelerating inference processing is key to determining the quality of AI services.

Key Features of the New “Dynamo” Software

Dynamo will be available as open-source software and is designed to accelerate inference processing by efficiently coordinating multiple GPUs. When combined with the latest “Blackwell” GPU architecture, it can reportedly increase the processing speed of the “R1” AI model from the Chinese AI company DeepSeek by up to 30 times compared to previous methods.

A core feature is a technique called “fine-grained serving,” which significantly improves processing efficiency by separating the inference process into two phases: “prefill” and “decode,” and assigning them to different GPUs.

Furthermore, by leveraging a technology called “KV cache” to store and reuse past token information, Dynamo reduces computational load. The “KV Cache Manager” integrated into Dynamo enables efficient cache management to avoid exceeding GPU memory limits.

The Trade-off Problem and Hardware Evolution

In his keynote speech, CEO Huang highlighted the trade-off between “total tokens per second (throughput)” and “tokens per user (latency)” in inference. This illustrates the dilemma where faster response times can limit the number of concurrent users, while supporting more users can lead to increased response delays.

To address this, NVIDIA has adopted a strategy of overcoming this trade-off through hardware enhancements. The newly announced “Blackwell” architecture boasts up to 25 times the processing power of its predecessor, “Hopper,” enabling a balance between quality and scale.

Continued Strong Investment in AI-Related Data Centers

As the primary use case of AI shifts towards inference, the demand for computational processing is experiencing exponential growth. Following “Blackwell,” NVIDIA has unveiled development plans for even higher-performance GPUs, such as “Rubin” and “Feynman,” with Dynamo evolving as the corresponding software foundation.

To support such high-density and high-performance AI processing, distributed and large-scale computing environments are essential. Consequently, with the expansion of AI agents and generative AI, investment in data centers as the underlying infrastructure is expected to remain robust in the future.