AI sovereignty does not mean that a single country must build every component of the ecosystem, from electricity and chips to models and applications. This is the central conclusion presented by an ITmedia AI Plus analysis, based on the views of Kunihiro Tanaka, president of Sakura Internet, and Junya Ishii, an international political analyst at Sakana AI. The more realistic standard is to determine which layers must be controlled and reduce dependence on a single external source, while retaining the ability to continue operating if a supplier fails or its terms of use change.
Three Aspects of Sovereignty
Tanaka divides AI infrastructure sovereignty into three aspects. The first is data sovereignty: keeping data inside the country and managing it within a local domain. The second is operational sovereignty, which includes having decision-making authority over infrastructure operations, server locations, and capacity. Technological sovereignty concerns the ability to develop the models and software themselves.
Tanaka argues that the risks associated with these aspects are no longer hypothetical. Problems have emerged involving data leaving external servers, foreign providers declining to invest in servers inside Japan, and the risks of software export bans. The article cites Anthropic’s temporary suspension in June of access to Claude Fable 5 and Mythos 5, based on an order from the U.S. government, as an example showing that reliance on a political alliance does not guarantee continued access to software or servers.
These risks are more difficult because they often do not appear under normal circumstances. A service may seem stable and inexpensive, then become unusable when political or commercial conditions change. Tanaka compares this to the “naphtha” problem, in which reliance on cheap imports during peacetime disrupts supply when circumstances change.
Why Is “Complete Localization” Not Enough?
Tanaka and Ishii agree that it is impossible and irrational for Japan to cover 100% of all AI layers. Ishii describes the ecosystem in four layers: infrastructure materials such as electricity and semiconductors; the data-management layer through the cloud or local infrastructure; the foundation-model layer; and, finally, applications.
Under this view, no single country or company can efficiently own all of these components. Ishii therefore proposes reducing foreign dependence at each layer and diversifying products and suppliers when using external technologies. He notes that Japan, South Korea, India, Australia, and Canada, as middle powers, are moving toward the same way of thinking.
Sakura Internet does not consider abandoning NVIDIA graphics processing units a realistic option; this part of the industry depends on a global supplier even for American companies. However, Tanaka distinguishes between the chip itself and the higher layers that Japan can develop, such as cloud infrastructure, inference services, and operational software.
What Changes in Practice for Companies?
Sakura Internet points out that asking, “Is the data stored inside Japan?” is not enough when evaluating an AI service. The more important question is: Is the input used to train the model? Data may physically be located on a Japanese server, but it loses part of its sovereignty if it enters training processes that the company does not control.
Companies should also verify the physical isolation of servers and whether they are dedicated to them, a point receiving particular attention from manufacturing-sector customers. This is not necessarily tied to the model’s nationality; Tanaka argues that an open model from China, such as Kimi or Qwen, can be operated inside local servers so that data does not leave the country, while the other risks associated with the source and software must still be assessed.
The issue of actual usability is also important. Possessing a powerful GPU on paper does not guarantee that it will be available when needed, because of resource shortages, electricity constraints, or a service provider imposing a performance ceiling. One option Tanaka proposes is using smaller, specialized models—even if they are inferior to leading models in general performance—and running them on local inference infrastructure. This may be more cost-effective when agent applications require the consumption of large quantities of tokens.
This approach does not mean isolating Japan from foreign technologies. Sakura Internet is collaborating with Microsoft to make its infrastructure and services available through the Microsoft Azure control plane. However, Tanaka emphasizes the need to use foreign technologies while building local capacity to provide comparable services later, rather than merely consuming them.
Adapting Models Instead of Building Them from Scratch
Sakana AI follows a different path at the model level. Instead of building the model architecture and weights entirely from scratch, it relies on advanced open models and uses post-training techniques to modify their behavior and adapt them to Japanese requirements.
Ishii says that American and Chinese models were developed with enormous resources, making it difficult to catch up by repeating the same approach. At the same time, foreign models may reflect particular values or political biases. Sakana AI therefore developed the Namazu model and, in May, launched the Sakana Chat chatbot powered by it.
The company does not aim to make the outputs heavily biased toward the Japanese viewpoint. Rather, according to Ishii, it seeks to provide neutral answers on historical and territorial issues while clearly presenting the Japanese position, because a one-sided answer could weaken trust among users overseas. The benefit of post-training is that it can be applied to multiple models, reducing the impact of any one model being discontinued.
This need is particularly evident in the government, finance, and defense sectors. Sakana AI is working with a research institution affiliated with the Acquisition, Technology & Logistics Agency of Japan’s Ministry of Defense under a multiyear contract to develop core technologies, with a trial lasting two years. The work includes lightweight AI that operates on drones in environments with weak connectivity, as well as systems that use information collected by aircraft in command and control.
Conclusion: Control Matters More Than Nationality
The experiences of Sakura Internet and Sakana AI offer a practical definition of sovereign AI: not a closed system bearing a “made locally” label, but the ability to control data, operations, and models, with alternatives that allow business operations to continue when a provider or model is lost.
Accordingly, Japanese companies first need to inventory the layers of their existing infrastructure: Is the data used for training? Who manages the infrastructure? What happens if the model stops working tomorrow? Answering these questions is more useful than relying on the model’s nationality alone. Projects that build everything from scratch, such as the approach taken by Preferred Networks, remain an important option, but the article places them among a range of choices rather than presenting them as a single solution. Sakana AI’s use of Sakura Internet’s “Takabi” cloud infrastructure in the GENIAC project also illustrates how local components can be integrated across different layers.