Where Domestic AI Stands in Japan: NEDO and GENIAC Support, Sakura Internet, and Winning With Domain-Specific AI Agents

Hello, this is Ryuta Hamamoto from TIMEWELL.

Over the past year or so, the phrase "domestic AI" has gone from rare to hard to avoid. The government is funding generative AI development, domestic cloud operators are installing GPUs at scale, and telecoms and startups have announced their own large language models one after another. And yet you hear, just as often, that none of it can catch the leading foreign models.

Both of those things are true. At the general-purpose frontier, Japan is well behind. But there is still ground where Japan can win. This article maps out where things actually stand. How the NEDO and GENIAC support programmes work, how far each company's domestic large language model has come, and why Sakura Internet keeps coming up as the infrastructure underneath it all. Then, rather than fighting head-on over scale, I want to talk through the path of winning with domain specialisation and AI agents, using TRAFEED, the export-control AI we have built, as a concrete example. I open up the jargon as it appears.

What a domestic AI stack is, and why it is being asked about now

Let me start with the vocabulary. By "AI stack" I mean the whole foundation needed to run generative AI. It helps to picture three layers. At the bottom sits the infrastructure layer: the GPUs and data centres that do the computing. Above that sits the layer of foundation models, the large models trained on enormous amounts of data that form the core of generative AI itself. And at the top sits the application layer, where those foundation models get built into actual work. "Domestic AI stack" refers to the idea of owning that infrastructure and those foundation models with Japanese technology and Japanese capital.

Why is this being asked now? Because generative AI has become a strategic technology on both economic and security fronts. A technology that can produce text, images and code carries huge weight for industrial productivity. At the same time, which country's, and which company's, model your work depends on ties directly to questions of data sovereignty and supply stability. Once a company hands the core of its operations to a single foreign vendor, the risk that pricing changes, that terms of use change, or that access disappears altogether becomes something neither companies nor governments can ignore.

Here I want to be honest about the reality: at the general-purpose frontier, Japan lags badly in capital and compute scale. This is not something you can pin to a specific ranking or market share; it is a qualitative assessment drawn from various analyses and reporting. The frontier general-purpose models are led by a handful of foreign firms able to pour in compute and development budgets of a different order of magnitude. It is not realistic for a domestic model to take them on, at scale, head-on. Which is exactly why the question of where to win matters. The second half of this article takes up those winnable openings in concrete terms.

If you want to get a feel for how your own operations sit at this intersection of technology and regulation, try our free export control self-check. In a few minutes it gives you a sense of which questions you need to be asking.

GENIAC and NEDO: how the state supports foundation-model development

You cannot talk about domestic AI without GENIAC. Its full name is the Generative AI Accelerator Challenge, a programme run since February 2024 by METI and NEDO, the New Energy and Industrial Technology Development Organization, a national body that supports research and development in industrial technology. The name sounds imposing, but what it does is simple at its core: the state helps companies developing generative AI secure the biggest constraint they face, namely computing resources, which means GPUs. Alongside that, preparing the data used for training and building a community where developers can share what they learn are pillars of the support too.

Why support computing resources? Because in generative AI development, performance turns not only on how the model is designed but on how many GPUs you can run, for how long. High-end GPUs are expensive and take time to procure. Even a well-funded startup finds it hard to lock in a large compute base on its own. By backing that from underneath, the aim is to lower the barrier at the entrance to development. The compute base itself is provided by cloud operators.

GENIAC is not a one-off call for proposals; it continues across successive rounds. According to what METI and NEDO have published, the early selections that began in 2024 were followed by further rounds, and June 2026 saw the kick-off of the fourth round. What I find interesting is how the centre of gravity has moved. Early on, the focus was broad foundation-model development. Round by round, the themes have shifted toward the field: efforts to make manufacturing data usable by AI, foundation models that drive robots directly, and real-world deployment in specific industries such as drug discovery and finance. Not "build one general-purpose model and be done," but grow specialised models rooted in the operational data of an industry. That is where policy attention has drifted, and it lines up with the domain-specialisation opening I describe later.

A word of caution on numbers. Figures reported around GENIAC, such as total funding or subsidy rates, circulate largely as secondary information in summary articles, and to confirm any amount or ratio precisely you need to go to the primary sources from METI and NEDO. For our purposes here, it is enough to hold the shape: a state programme built around securing computing resources that keeps running to support generative AI development. Relatedly, NEDO also operates GENIAC-PRIZE, a results-based prize programme that encourages the development of AI services addressing social challenges such as labour shortages, using a prize format rather than the usual detailed inspection and verification process ¹. State support, too, is diversifying, from uniform subsidies toward an emphasis on results.

How far each company's domestic LLM has come

With the support framework in hand, let me turn to the actual models. LLM stands for large language model, the core of generative AI, a language model trained on vast amounts of text. Over the past year or two, domestic LLMs have grown steadily in number. I cannot cover all of them, but here are enough to show the differences in character.

PLaMo, developed by Preferred Networks, is one of the leading purely domestic models. It is strong at handling Japanese, and its standing has risen, including winning a domestic product and service award in early 2026. The company has also moved to collaborate with Sakura Internet and the National Institute of Information and Communications Technology on jointly developing a successor model, a good example of pairing domestic infrastructure with a domestic model. ELYZA, part of the KDDI group, is a startup with roots in the Matsuo Lab at the University of Tokyo, and it has released Japanese-specialised models in a form that can be used commercially. Its development leans toward practical usability, such as speed and power efficiency, with models built on new methods for making inference more efficient.

NTT's tsuzumi 2 was announced in October 2025 as a purely domestic model aiming for high performance while staying lightweight ². What stands out is that it can run inference on a single GPU, or even on CPU alone. That design speaks directly to the strong needs of finance, healthcare and government, which want to run AI without sending confidential information to an outside cloud, inside a closed environment or on-premises (operation within a company's own equipment). Its training efficiency is high too, and it is pitched on being able to pick up specialist knowledge from a small number of examples. Rakuten's Rakuten AI 3.0, released in March 2026, is a model of roughly 700 billion parameters (a parameter being a measure of model size, here around 700B scale), offered free of charge and available for commercial use under the permissive Apache 2.0 licence ³. It claims results on Japanese benchmarks that surpass large foreign models, and it was developed with GENIAC support. SoftBank's Sarashina has a confirmed track record in the 460-billion-parameter model released in November 2024, and it emphasises data sovereignty through operation in domestic data centres. A larger model is something management has set out as a target, so the stated goal and the scale actually shipped need to be kept apart.

The place where these models come together is the Digital Agency's government AI, "Gennai." This is a shared generative AI platform for government use, with the domestic models to be trialled chosen through open selection. Several domestic models were picked from many applicants, and validation is planned across numerous ministries and agencies and a workforce numbering in the low hundreds of thousands. Reporting says the specific line-up of selected models includes tsuzumi, ELYZA, PLaMo, Sarashina and a model from a major electronics maker, but for the official list and naming, the surest source is the Digital Agency's published materials. Either way, the government itself trialling domestic models in earnest and moving to formally procure the strongest is a major force creating demand for domestic AI.

Sakura Internet, the base beneath domestic AI infrastructure

I have been talking about models, but without the infrastructure to run them, they are castles in the air. This is where Sakura Internet's presence has been growing. The company is building domestic AI infrastructure around its Koukaryoku GPU cloud. Koukaryoku comes in three forms: a bare-metal type that lets you use a whole physical server, a type usable through containers (lightweight execution environments), and a virtual-machine type, so developers can choose according to their use. On top of that it offers the Sakura ONE managed supercomputer and the Sakura AI Engine, a platform that puts generative AI into a ready-to-use form, presenting a safe and trustworthy AI platform on domestic data centres and a domestic cloud base ⁴.

Why does the domestic character of the infrastructure matter this much? This is where the economic security context comes in. Sakura Internet's cloud has received national certification as a plan to secure the stable supply of a specified critical material defined under the Economic Security Promotion Act, specifically cloud programs. A specified critical material is one that the state designates as essential to people's lives and the economy and problematic if its supply becomes overly dependent on a particular country, and cloud is positioned as one of them. Under this framework, the state has decided on grants to multiple operators for building AI computing resources, and Sakura Internet is among those covered. To avoid any misreading: the grant figures reported are a ceiling against the combined total for several operators, not an amount received by a single company on its own.

The company also keeps expanding the compute base itself. It is scaling up, including the 2026 start of AI infrastructure fitted with large numbers of the latest-generation GPUs on the grounds of its Ishikari Data Center in Hokkaido ⁵. To be precise on one point: news of these individual facility expansions and the national certification or grants under the Economic Security Promotion Act sit on separate tracks of information. It would be wrong to simplify this into "the state directly subsidised the latest GPUs"; policy support and the company's own investment are more accurately understood as distinct. Even so, taken as a whole, the policy direction is clear: protect the domestic cloud base from foreign dependence and support domestic AI development from the compute side. Sakura Internet has become a symbolic carrier of that direction.

On top of this infrastructure, domestic models grow, and companies and government actually use them. That circulation starting to turn is where the domestic AI stack now stands.

The realistic path: winning with domain specialisation and AI agents

Now for the heart of the matter. At the start I wrote that Japan lags on scale at the general-purpose frontier. So where is the ground where Japan can win? I see two main openings.

The first is domain specialisation. A general-purpose model aims to handle any topic, broadly if shallowly. There, the volume of training data and the size of the model translate almost directly into performance, so it becomes a contest of capital and compute. A model narrowed to a specific task or industry is a different story. On ground such as reading and writing Japanese, Japanese business practice, or a specific industry's expertise, what counts is how carefully you gather that field's data and how faithfully you build for the actual work. Here there is real room to close a gap in scale through engineering. Many of the domestic models I mentioned claim results that hold their own against the big foreign names on Japanese-language or task-specific benchmarks, which is precisely this domain-specialisation thinking showing through. Rather than aiming for the world's largest model that can answer any question, it is about winning on the point of being reliably useful in this field.

The second is AI agents. This one needs some explanation. A large language model is, in effect, a brain that generates text. But a brain alone does not get the actual work done. Fetch the needed information from outside, drive the processing along a defined procedure, and produce a conclusion together with its basis. The arrangement that sees this whole sequence through to the end is an AI agent. Two components are key here. One is a trustworthy knowledge base for the model to consult. The other is a framework that defines what the model should do, in what order, which is called a harness. A harness originally refers to the fittings or wiring bundle that safely holds a machine or a person in place; in the AI context it means the arrangement that gives the model-as-brain hands and tools and controls it so it does not stray from the defined path.

Why does this second opening become a way for Japan to win? Because even if you cannot win a race to build the world's largest standalone model, if you build a full agent that includes the knowledge base and harness, you can compete perfectly well on practical usefulness in a specific task. In Japan too, a startup with strength in evolutionarily merging models of differing architectures has adopted a domestic GPU base as its compute and is working on developing a large model built to operate as an agent ⁶. This is a development target rather than a finished product, but I read it as a symbolic sign of the direction: going after the win as an agent, not as a standalone model.

In this context, TRAFEED, which we have built, is an AI agent designed exactly around that combination of domain specialisation and agent architecture. Let me get concrete about what is inside it in the next section.

What TRAFEED shows: one concrete answer

TRAFEED is an AI agent specialised in the single domain of export control. Export control refers to the whole practice of determining whether your company's products or technology fall under the targets defined in law, and, where they do, going through the required procedures before exporting. At the core of that determination sits the work of classification, checking whether your goods or technology match the control lists by comparing the requirements in the provisions against the product's specifications, one by one, and recording the result with its basis. If you want the fundamentals of export control itself, read what are dual-use items alongside this.

Why does a general-purpose large model struggle with this domain? Three reasons. First, the basis for a decision lies in an extremely concrete comparison of legal text against product specifications, which requires a kind of accuracy entirely separate from the ability to fluently produce plausible-sounding prose. Second, control lists are updated by each country on its own schedule, so unless you are always connected to the latest primary sources, last year's correct answer becomes this year's wrong one. Third, because the work handles highly confidential product and counterparty information, the question of how much to hand to which model, the matter of data sovereignty, comes into play. Vaguely asking a general-purpose model and getting a plausible answer back satisfies none of these three.

So we built TRAFEED not as a standalone model but as an agent that includes the knowledge base and the harness. The knowledge base is a knowledge graph of more than 200 million records, on the order of 90 million papers, 100 million patents and 300,000 researchers. A knowledge graph is a network of knowledge that draws lines connecting elements such as people, organisations and technologies to represent their relationships, and having it means checks can account for the connections behind a name, not merely a match of the name itself. The harness is the arrangement that drives the model through the classification procedure and leaves the basis on the record. In joint validation with Okayama University using roughly 30,000 past screening records, we confirmed AI classification accuracy of 95% or higher (internal study). This export-control-specialised AI agent is, to our knowledge, the world's first initiative in Japan's security export control field (as of March 2026, per our research), holds Patent No. 7862062, and is already in use at more than 20 organisations. That said, the final classification is made by each company's export control officer. What the AI handles is performing the comparison between provisions and specifications without gaps and at speed, and leaving the reasoning on the record. We draw that line clearly.

How domain-specific AI like this fits within the larger theme of economic security is laid out in the economic security and AI ecosystem. And for the research-security practicalities that sit right alongside export control, I put together Japan's research security procedures manual. Within the big picture of a domestic AI stack, I think of TRAFEED as one concrete form of going after the win somewhere other than the general-purpose contest of scale.

Summary

Let me restate where the domestic AI stack stands.

• The AI stack is easiest to grasp as three layers: infrastructure, foundation models and applications. The domestic-production debate is aimed mainly at the lower two
• GENIAC is METI and NEDO's support for generative AI foundation-model development, built around securing computing resources, and its centre of gravity has moved from general-purpose to industry-specific
• Domestic LLMs have grown into a real set of choices, including PLaMo, ELYZA, tsuzumi 2, Rakuten AI and Sarashina, and validation is advancing in the government AI "Gennai"
• Sakura Internet forms a central part of domestic AI infrastructure through Koukaryoku, the Sakura AI Engine and more, and has received national certification under the Economic Security Promotion Act
• While Japan lags on scale at the general-purpose frontier, there is a realistic winnable opening in domain specialisation and in AI agents that include the knowledge base and harness
• TRAFEED shows one form of that, as an AI agent combining export-control specialisation, a knowledge graph and a classification harness

If you set catching the giant foreign models on scale as the only goal, the story of domestic AI is bound to feel bleak. But choose where to win, and build something reliably useful there, knowledge base and harness included. Switch to that frame of mind and domestic AI turns out to have plenty of room to compete. Where in your own operations could a domain-specific AI like that make a difference? If you are considering export control or economic security, talk to our TRAFEED team.

---

Footnotes

1. NEDO, "GENIAC-PRIZE" https://geniac-prize.nedo.go.jp/ ↩

2. NTT Research and Development, "tsuzumi 2" https://www.rd.ntt/research/JN202606_39542.html ↩

3. Rakuten Group press release (17 March 2026), "Rakuten AI 3.0" https://global.rakuten.com/corp/news/press/2026/0317_01.html ↩

4. Sakura Internet, "Sakura AI" https://ai.sakura.ad.jp/ ↩

5. Sakura Internet news release (25 February 2026) https://www.sakura.ad.jp/corporate/information/newsreleases/2026/02/25/1968223641/ ↩

6. Sakura Internet news release (16 July 2026) https://www.sakura.ad.jp/corporate/information/newsreleases/2026/07/16/1968225411/ ↩