Sovereign AI as a growth driver:
How the development
of homegrown LLMs
will drive
SoftBank's next leap forward

I see the development of this foundation model as a crucial initiative that lays the cornerstone for the future of AI development in Japan. Today, AI development starts by developing a very high-performance, large-scale AI that acts as a “teacher.” This “teacher” AI possesses a vast amount of knowledge, but it requires significant computing resources and electricity to operate. This presents challenges in terms of cost and response speed for everyday business use. Therefore, from the knowledge held by this “teacher” AI, we are developing a lighter, faster, and more power-efficient “student” AI. It is this optimized “student” AI that will actually be used by many of our customers.
In other words, to continuously create high-performing “students,” a superior “teacher” that serves as their source is essential. If we were to depend on “teachers” from other countries, we would face the risk of suddenly losing our freedom of development due to changes in licenses or regulations. This is the background to why SB Intuitions took on the challenge of developing a large-scale foundation model with 460 billion parameters, and I believe it was an essential step for Japan to build its own AI ecosystem for the future.

The strategic intention behind developing this model is to accelerate the practical deployment of AI across society and to enable further advancement of future LLMs.
First, a model that is well-balanced in terms of response speed, answer accuracy, and cost is essential for many client companies to implement AI. “Sarashina mini” has been optimized using techniques such as “model distillation*” to meet diverse needs while maintaining, as much as possible, the high-level performance of the large-scale foundation model developed earlier.
However, the role of this model does not end there. “Sarashina mini,” while being an outstanding “student,” also becomes an important component for creating the next generation of AI. Specifically, by combining multiple 70 billion-parameter models with different areas of expertise, we will evolve them into an even more powerful AI, like a “team of specialists.” By continuously training this “team of specialists,” we will efficiently build the next-generation “teacher” AI with even higher performance in a short period, aiming to accelerate the development of a one trillion-parameter scale LLM.

The response has been highly encouraging, and we are steadily preparing for commercialization in the fall of 2025, as originally planned. SoftBank employees, who regularly use cutting-edge AI like ChatGPT Enterprise and Gemini, have high standards, so to be honest, the initial internal feedback was tough. However, this feedback was extremely valuable to SB Intuitions, and we have made updates almost weekly, incorporating the high-level requests from employees. As a result, employee satisfaction has steadily improved, and I am confident that the product is heading in the right direction.

The strengths of “Sarashina” lie in being a “high-quality LLM rooted in Japanese language and culture” and in its ability to realize “sovereign AI.”
As a homegrown LLM, being a “high-quality LLM rooted in Japanese language and culture” allows it to accurately understand the complex nuances and contexts of the Japanese language and respond appropriately to specific wording and professional tasks. For example, the concept of “guardrails,” which control inappropriate expressions that AI should not generate or information it should not output, differs greatly depending on the country and culture. Even interpretations of copyright vary across nations. This level of meticulous attention to legal regulations, social norms, and even the sentiments of native Japanese speakers is a strength that only a homegrown LLM can achieve. With this homegrown LLM as a foundation, we want to evolve it into a truly customized AI that aligns with the specific operations of each client company, thereby contributing deeply to their businesses. Such initiatives are already underway. For example, we are jointly developing an LLM and AI agents specialized for the pharmaceutical industry with Chugai Pharmaceutical Co., Ltd., aiming to speed up new drug development. We have also begun joint research and development of an LLM specialized for finance with Mizuho Financial Group, Inc.
Being a “sovereign AI” becomes even more important when AI plays a core role in business. Dependence on overseas LLMs carries not only the concern of a company's confidential information being unintentionally transferred abroad, but also serious risks related to business continuity itself. For example, if an AI is controlling a manufacturer's production line, and its use is suddenly restricted due to changes in foreign laws or international affairs, the factory's lines could be shut down. To solve this challenge, SoftBank emphasizes three types of sovereignty—“technology,” “data,” and “operations”—as well as the nation's own “laws and regulations.” “Technological sovereignty” means developing our own LLM and controlling what data it learns; “data sovereignty” means safely managing customer data in domestic AI data centers; “operational sovereignty” means having domestic operators manage that infrastructure; and “laws and regulations” means establishing rules for AI use and development based on Japanese laws and ethics. By building a sovereign AI that fulfills these elements, we can meet the high-level needs of government, universities, research institutions, and corporations that handle confidential information.

SoftBank's “multi-LLM strategy” is based on the idea of selecting the optimal AI according to the sensitivity of information handled by the client company.
For summarizing general information or generating text, using excellent overseas-made LLMs is a valid option. On the other hand, the more a task involves confidential or proprietary information and the greater the need for domestic specificity and information security, the more important a homegrown LLM becomes. It is in such operations that “Sarashina,” with its deep understanding of Japanese legal language and business customs, demonstrates its true value. Furthermore, when industry-specific expertise is required, we will provide customized models and AI agents for each industry based on “Sarashina.” And for handling the most sensitive data related to national interests, we envision its use as a true “sovereign AI” that can also comply with government certifications.
Thus, the “multi-LLM strategy” and the sovereign AI strategy are inextricably linked. With SoftBank managing the entire scope of AI utilization as a partner to client companies, they can enjoy the convenience of overseas AI while protecting their important data and operations with an optimal homegrown LLM centered on “Sarashina.” Providing diverse options while having a secure, domestic option at the core—this is the essence of SoftBank's approach to sovereign AI.

SoftBank's strategy is not to wait until the entire infrastructure is fully built out before monetization. Rather, we aim to consistently generate profits along the way from individual businesses such as AI solutions and cloud services. It might be easier to understand if you imagine the business model of Japan's private railway companies. They did not wait until their entire railway lines were completed before monetization. Instead, in the process of extending their lines, they generated revenue by developing commercial complexes at stations to offer various services and by developing residential areas along the railway lines. Similarly, SoftBank believes that this combination of a long-term development perspective and short-term business viability is the key to achieving sustainable growth.

At SoftBank, we see movements in the industry like DeepSeek not as a risk, but rather as an opportunity that validates the soundness of our strategy.
The case of DeepSeek is a prime example of the latest trends in AI development, and SB Intuitions has been closely monitoring and analyzing it in detail. Traditionally, in the “tuning phase” after a foundation model has completed its basic learning, it was common to have the AI learn from a set of questions and answers created by humans. However, the brilliance of DeepSeek's approach is that they have the AI, not humans, generate this question-and-answer set itself, making the process more efficient. Through this, they have succeeded in enhancing logical reasoning for tasks with definitive answers, like mathematics. This “training AI with AI” method is a clear validation of the development strategy that SB Intuitions has been promoting, where a “teacher AI nurtures a student AI.”

However, the important point here is that this tuning method is only possible with a powerful “teacher” AI at its base—in other words, a high-performance foundation model. To build this foundation model from scratch and continuously evolve it, large-scale investment in AI computing infrastructure is still essential. In the competition to develop the core part that determines a model's performance, the scale and quality of the AI computing infrastructure remain the key to success.
Therefore, these industry trends actually highlight the importance of SoftBank's investment in large-scale AI computing infrastructure and our development of a high-performance foundation model. The technological evolution shown by DeepSeek is not a risk to SoftBank's strategy; on the contrary, it proves its validity.

I believe the homegrown LLMs developed by SB Intuitions will become a high-value-added proprietary technology that drives SoftBank's future growth. SoftBank is making large-scale investments to build one of Japan's largest AI computing infrastructures and is also planning to open massive AI data centers in Hokkaido and Osaka Prefecture. By integrating SB Intuitions' LLMs with SoftBank's state-of-the-art AI infrastructure, we will realize a “sovereign AI” and “sovereign cloud” with competitive advantages that no other company can match.
Until now, SoftBank has been an operator that excels at quickly introducing superior products and services from around the world and popularizing them among consumers and businesses. From now on, in addition to this strength, we will evolve into a company with product-creation capabilities—what in Japan is called “monozukuri”—to develop the foundational technologies for society and pioneer new markets. I am confident that the challenge undertaken by SB Intuitions will lead to the sustainable growth of the entire Group and the enhancement of its corporate value.