Q&A at
ESG Briefing Session

Date	Friday, March 7, 2025 4:00 pm - 5:30 pm
Speakers	Hironobu Tamba (Vice President, Technology Unit, Data Platform Strategy Division Head & Digital Infrastructure Architect Office Head, SoftBank Corp./President & CEO, SB Intuitions Corp.) Yasuyuki Genda (Vice President, HR Division Head, General Affairs Division Head, Well-being Promotion Office Head, SoftBank Corp.) Masato Ikeda (Vice President, CSR Division Head, ESG Planning & Management Senior Director, SoftBank Corp.) Mariko Fujiwara (Strategic Finance Division, IR Office Head, SoftBank Corp.)

Date

Friday, March 7, 2025
4:00 pm - 5:30 pm

Speakers

Hironobu Tamba (Vice President, Technology Unit, Data Platform Strategy Division Head & Digital Infrastructure Architect Office Head, SoftBank Corp./President & CEO, SB Intuitions Corp.)
Yasuyuki Genda (Vice President, HR Division Head, General Affairs Division Head, Well-being Promotion Office Head, SoftBank Corp.)
Masato Ikeda (Vice President, CSR Division Head, ESG Planning & Management Senior Director, SoftBank Corp.)
Mariko Fujiwara (Strategic Finance Division, IR Office Head, SoftBank Corp.)

What is SoftBank Corp.'s approach to securing power sources for AI data centers?

We will consider locations for future AI data centers based on the availability of renewable energy. In Japan, areas where large amounts of renewable energy can be secured are limited, and securing sufficient land is necessary. We will address the challenge of building distributed AI data centers through discussions with the government and relevant ministries.
With the emergence of AI agents replacing many human tasks, how are business strategies and HR strategies aligned, and how do you plan to restructure the HR portfolio?

We leverage new technologies such as generative AI to enhance efficiency and productivity in our existing businesses, followed by reallocating personnel. In practice, aligning with our medium-to long-term business strategy, we have been reallocating personnel to Next-generation Social Infrastructure and new businesses, resulting in a slight decrease in our standalone employee headcount. For example, to strengthen our cloud business, we implemented a “development-based job posting” system and relocated 140 employees. This system involves reskilling employees over a three- to six-month period, followed by role assignments based on their proficiency. Moving forward, while continuing to streamline operations through AI agents and other technologies, we will aim to reallocate personnel to areas such as the data center business and global operations through reskilling, rather than simply reducing employees.
How does SoftBank Corp.'s approach to human capital management, including its systems and policies, impact the industry?

First, we will fully leverage “Cristal intelligence”—a cutting-edge enterprise AI that securely integrates all systems and data of an individual company and is customized exclusively for that company—within our organization to enhance operational efficiency. We will then assess the extent to which specific tasks can be optimized and identify new roles that should be handled by humans. By accumulating knowledge through this practice, we aim to provide large enterprises with expertise in business reform and contribute to improving efficiency and productivity across the entire industry.
What is the differentiation between “Cristal intelligence” and the homegrown large language model (LLM) that SB Intuitions Inc. is developing?

We adopt a “multi-model strategy,” offering multiple LLMs based on customer needs. The homegrown LLM by SB Intuitions Inc. will be treated similarly to LLMs developed by other companies and provided according to customer requirements.
“Cristal intelligence” is a product built around OpenAI's LLM and is planned to be sold under SB OpenAI Japan, a joint venture with OpenAI. We believe there is much to learn from OpenAI, including what kind of AI the SoftBank Corp. Group should aim for. We will continue striving to become “the company that utilizes AI most effectively.”
What are your initial impressions after using “Cristal intelligence”?

I have not used “Crystal Intelligence” yet, but based on my experience with OpenAI's newly announced services, “Deep Research” and “Operator,” I believe that reasoning (logical inference) will be one of the key directions for the future evolution of AI. For example, running reasoning over a certain period allows the AI to perform deep thinking and inference, similar to “Deep Research”. However, I also recognize that further advancements are required to create a true AI agent. Additionally, the “Browser-use” feature in “Operator” is highly commendable, demonstrating the ability to operate numerous applications through browser-based learning.
Will the AI computing infrastructure required for the introduction of “Cristal intelligence” be provided by OpenAI at data centers in Tomakomai, Hokkaido, and Sakai, Osaka? Or will the AI computing infrastructure include data centers procured under the “Stargate Project”?

We are currently in discussions with OpenAI regarding the implementation approach. We recognize that in order to use “Cristal intelligence” within our Company and sell it to Japanese companies through “SB OpenAI Japan,” it is necessary to securely manage data within Japan and establish AI computing infrastructure domestically.
How does SoftBank Corp. evaluate open-source AI models such as “DeepSeek”?

We believe the content released by “DeepSeek” is useful for the development of the LLM by SB Intuitions Inc. Traditionally, fine-tuning AI required using human-created data, which was time-consuming and costly. However, according to the paper published by “DeepSeek”, they appear to be using AI-generated data for reinforcement learning. This approach is particularly valuable for enhancing logical reasoning in areas like mathematics, where there is a single correct answer. At SB Intuitions Inc., we have also been advancing efforts to train AI with AI, and we reaffirmed that this approach is on the right track.
How do you perceive the business model of open-source AI?

Utilizing open-source AI is one option. The example from “DeepSeek” demonstrated that once a sufficiently large LLM is built, the costs associated with reinforcement learning and tuning can be reduced. When the inner workings of an AI model is not a concern, or when used for fixed purposes, tuning and using such models cost-effectively is a viable option.
However, building the foundational LLM and adding new features still involves significant costs, necessitating ongoing research. Particularly when adding new features, the workload can be akin to rebuilding the AI model, making in-house development of an LLM a viable option.
How many GPUs are required for AI training and inference, respectively?

We have not disclosed specific numbers. If the number of parameters for the LLM we aim to develop is large, the number of GPUs required for AI training will also be significant. If the GPUs used for training are large-scale and high-performance, the time required for AI training can be reduced.
The amount of GPUs required for inference is expected to vary greatly depending on future AI use cases. To reduce AI latency (response time), we anticipate that data centers with GPUs for inference will need to be deployed in a distributed manner, placing centers closer to users.
Will the GPUs we are currently planning to procure be insufficient, requiring the acquisition of additional GPUs?

The GPUs purchased by our Company are being utilized as an AI computing infrastructure for AI-related research and development within Japan, including use by domestic research institutes and universities, with the support of government subsidies. Since the use of GPUs by “SB OpenAI Japan” and OpenAI is for commercial purposes, we believe that the AI computing infrastructure needed within Japan should be discussed separately from the resources secured by SoftBank Corp.
Will the GPUs placed in the AI data centers in Tomakomai, Hokkaido, and Sakai, Osaka, be used for AI training by SoftBank Corp. Group?

These GPUs are expected to be installed not only for our group but also to meet the GPU needs of Japanese companies. We aim to build an environment that can be used for both AI training and inference. If OpenAI requires AI computing infrastructure in Japan, SoftBank Corp. may provide the necessary infrastructure as an AI data center operator.
NTT has proposed the concept of “IOWN” as an all-photonics network. Will it collaborate with SoftBank Corp.'s “all optical network” in the future?

We understand that detailed discussions on compatibility have not taken place, but both concepts are based on “photon-electron convergence” technology and are progressing in a common direction in a broader sense.
What is SoftBank Corp.'s stance on the government's cloud strategy, such as the “Government Cloud”?

I understand that the “Government Cloud” is focused on how government and municipalities provide application services. Looking ahead, when considering how to handle sensitive information in cloud and AI environments, and how to manage data generated from IoT and other societal sources, these aspects fall outside the scope of the “Government Cloud.” There is a recognition of the need for the concept of a “Sovereign Cloud,” which is currently being discussed with the government. While the roles of the “Sovereign Data Center” supporting the “Sovereign Cloud” and “Sovereign AI” are being discussed individually, we believe it is important to consider them as a unified entity.
Regarding smart buildings, is it possible to ensure data quality and integrate services through the standardization of building OS?

Data integration remains a challenging issue. We recognize that it will be necessary for multiple companies and organizations to collaborate in building a data integration platform. A “connector” that can accurately convert different data formats will be needed, as well as an authentication method to ensure the accuracy of the data. We believe it is important to proceed with discussions in cooperation with the government.