Japan's AI sovereignty play is orchestration, not building a new frontier model

The message landed quietly but carried implications for every country that relies on imported AI models. Anthropic restricted export of its frontier model Mythos this year, citing security risks. For Japan, which does not produce a frontier-class large language model, the block was a reminder of a vulnerability built into its technology supply chain. Japan is not alone in this concern; France has also laid out a €2.5 billion blueprint for national AI sovereignty.

Sakana AI, the Tokyo-based unicorn, argues that the answer is not to try to outspend OpenAI or DeepSeek on training runs. Instead, the company is betting on something it calls orchestration. In June 2026, it released Sakana Fugu, a model that acts as a command layer over a pool of existing AI models. A conductor picks the right model for each task, combines outputs, and lets the whole system keep working even if a particular model disappears.

We can replace any model in the pool at any time, said Motoki Sato, head of Sakana AI's defense division. If a high-quality domestic base model appears, we can swap out foreign ones. That is the core idea behind AI sovereignty, and orchestration is the mechanism that makes it practical.

The approach sidesteps the pre-training race. Sakana AI does not build frontier models from scratch. It takes existing open-source models and adapts them through post-training. The company demonstrated this with Namazu, a prototype model that adjusted a frontier model for Japanese language and values. Sato argued that Japan should establish sovereignty in post-training, not in pre-training, because adapting models to local needs is more realistic than outspending U.S. and Chinese tech giants.

Earlier this year, Sakana AI won a contract from Japan's Acquisition, Technology and Logistics Agency to research how multiple AI systems can accelerate observation, reporting, information fusion, and resource allocation on the battlefield. The project involves building a command and control system where AI assists human decision making by automatically converting drone footage and audio into text and plotting it on a map. The system is being developed through agile iterations with actual troops testing the hardware and giving feedback.

Sato was emphatic about one point: the system does not make decisions. It only supports faster situational awareness. Final decisions must be made by humans, he said, especially in defense where handing over command authority to foreign systems raises constitutional issues and risks connectivity failure in a crisis.

The company also works on detecting disinformation online. With Yomiuri Shimbun, it analyzed Chinese state-linked accounts spreading anti-Japan narratives on social media. For Japan's Ministry of Internal Affairs and Communications, Sakana AI built a system that visualizes the information space, flags deepfakes, and proposes countermeasures. Sato cited examples where real images were edited with AI to show Donald Trump next to a person who had been swapped in, or foreign disaster footage was altered to look like an earthquake in Tokyo.

Financing has come from unconventional sources. In its Series B round, Sakana AI accepted investment from In-Q-Tel, the venture capital arm founded by the CIA. Sato said the move reflected growing interest in defense applications, not wariness. The company's website publicly states that its products and services will not take autonomous lethal action, in line with Japan's constitutional framework.

The orchestration system itself is a domestic technology, developed in-house. Sato claimed that Sakana Fugu can match Mythos in certain domains because the orchestration layer compensates for individual model weaknesses. The specific benchmark evaluations were not disclosed in the interview, and Sato acknowledged that U.S. companies still have better single-model performance. The gap is narrowing, he argued, by combining multiple models smartly rather than pouring resources into a single giant one.

For Japan, which has long depended on foreign hardware and software for critical defense infrastructure, the idea of an orchestrated AI layer that can be updated piece by piece has obvious appeal. It does not solve the problem of having no homegrown frontier model, but it reduces the risk of a single point of failure. In a world where AI models can be cut off overnight, building a switchboard may be more valuable than digging one more well.