On June 30, 2026, Japan’s Ministry of Economy, Trade and Industry (METI) announced the launch of the “Program for Developing Multimodal Foundation Models for AI Robots and Physical AI.” Following a competition conducted by the government agency NEDO, two organizations were selected from fifteen applications: Noetra, established by SoftBank, Sony, NEC, and Honda, and the National Institute of Advanced Industrial Science and Technology (AIST). The approved first-year budget for the program includes 378.3 billion yen, or about $2.4 billion. Over five years, the amount of support could reach 1 trillion yen.
Betting on Manufacturing Data
Japan’s program stands out from other national AI initiatives for its narrow specialization. The money is not going to general-purpose language models, where American and Chinese developers dominate, but to “physical AI” — models for controlling robots, autonomous production lines, and transport. In addition to text, such models process images, video, sound, sensor readings, and physical characteristics of the environment, such as wear and heating of components.
The rationale for the choice is stated explicitly in METI’s documents: Japan’s main asset is the manufacturing data accumulated over decades. Data from factories, construction sites, logistics centers, medical institutions, and social service facilities are barely represented on the internet and are therefore inaccessible to developers of general-purpose models. Minister of Economy, Trade and Industry Ryosei Akazawa described this as “a path to success through data accumulated in the field of elderly care, disaster response, at production facilities, and in the decommissioning of Fukushima Daiichi.”
There is also a defensive rationale. Japanese companies operating on foreign closed models are transferring production data outside the country and depend on decisions made by foreign regulators and vendors. According to the ministry, a domestic foundation model is needed to “protect production data and use it confidently in the future.” The third argument is energy: Japan has low energy self-sufficiency, so model energy efficiency has been built into the program’s targets from the outset.
How the program is structured
The program is designed for five fiscal years, from 2026 to 2030. Noetra is developing a multimodal foundation model based on the needs of Japanese companies, while AIST is conducting forward-looking research together with Japanese and international research centers. The consortium’s proposal was titled “Research and Development of Core Technologies for Physical AI for Real-World-Native Applications.”
Noetra is a project company established by four corporations for this purpose (formerly known as Japan AI Foundation Model Development). It is headed by Hironobu Tamba, a senior executive officer at SoftBank and a long-time supporter of Japanese-developed models. The company plans to expand its circle of participants and investors to more than forty companies (Nikkei puts the figure at 44) from the automotive, electronics, logistics, telecom, IT, and financial sectors. The idea is to create a cross-industry framework from research to deployment, in which corporations share data without disclosing commercial secrets to one another.
The program continues the GENIAC initiative, an NEDO project launched in February 2024 through which the government subsidizes computing resources for Japanese foundation model developers. The bulk of the new funding is directed to a single foundation model, while GENIAC continues to support sector-specific and robotics models in parallel.
Conditional funding
Although the total support is estimated at roughly ¥1 trillion, contracts have been signed only for the first two years. Each subsequent year of funding goes through a stage-gate process: a committee with external experts reviews the results and the case for further investment twice a year — first approving the budget request, then the continuation of the project and the allocation of funds for the following year. Quarterly review meetings are scheduled between audits to adjust the course. The audit results will be published subject to confidentiality requirements.
The target indicators are set out in the program’s logical model.
From 2026 to 2030, at least one model is to be released each year, with trained weights transferred to Japanese developers.
The first version is due to be released in the current fiscal year at the baseline performance level of widely used open models, followed by annual expansion of modalities and reasoning depth. By 2030, the model is expected to be used by at least 80% of Japanese companies implementing physical AI with government support. On the research side, the program targets 30 publications at leading conferences (CVPR, NeurIPS, ACL, ICRA) in the first year and 60 by 2028. The program is also expected to contribute to reducing CO₂ emissions by around 60 million tons by 2035 through optimization of industry, transport, and buildings.
Place in the broader strategy
The program is based on two higher-level documents. The Basic AI Plan, approved by the Cabinet on December 23, 2025, became the first strategic document implementing Japan’s law to promote the development of AI. The AI Robotics Strategy, adopted by the interministerial council on March 26, 2026, sets the goal of capturing more than 30% of the global AI robotics market by 2040: METI estimates this segment at roughly 20 trillion yen, with the total physical AI market at around 60 trillion yen (McKinsey estimate). On June 30, while announcing the program’s implementers, Minister Akazawa presented an updated strategy: to increase Japan’s fleet of autonomous AI robots to 10 million units by 2040, expanding coverage from 16 to 18 industries by adding food service, the food industry, and healthcare. The foundation model program is intended to provide the technological basis for these goals.
The stated figures are planning benchmarks. Both the 1 trillion yen over five years and the 10 million robots by 2040 depend on annual audits, the budget process, and Noetra’s technological results.
What matters here for Kazakhstan
For Kazakhstan, which has declared 2026 the Year of Digitalization and Artificial Intelligence, the Japanese program is of interest primarily for its design: the budgets are incomparable, but the institutional solutions are transferable. The basic elements of the country’s own AI policy are already in place: the first Artificial Intelligence Law in Central Asia has been adopted, a dedicated ministry is operating, the Alem.Cloud and AI-Farabium supercomputer clusters with a combined AI capacity of about 3.6 exaflops have been launched, and the national language models KazLLM and Alem LLM are being developed. The next-stage question is how to turn infrastructure into a sustainable ecosystem. Four takeaways can be drawn from the Japanese experience.
First, selecting a niche based on comparative advantage. Japan is building the program around an asset that gives it a clear edge — industrial data. For Kazakhstan, the same logic points to domains with unique national data: the Kazakh language and a multilingual environment, the mining and metallurgical sector, agribusiness, transit logistics, and public services.
Second, results-based funding. A stage-gate process with annual reassessment, external experts, and publication of the results makes it possible to launch long-term projects without unconditional multi-year commitments. The mechanism should be assessed after the first year of the Japanese program.
Third — the terms of public support. The program’s target indicators include the annual transfer of trained weights to third-party Japanese developers, so the entire ecosystem gains access to the results.
Fourth, energy efficiency as a baseline requirement. Japan built low model energy consumption into the KPIs from day one, recognizing that the growth in AI workloads is constrained by the power system. This requirement should be incorporated into national AI projects at the technical specification stage, taking into account Kazakhstan’s commitments to achieve carbon neutrality and its plans to develop AI infrastructure.
The first results of the Japanese experiment are expected soon: the release of the first model is scheduled for the current fiscal year, the first stage-gate — approval of the budget request for the following year — will take place by autumn 2026, and the funding decision will be made in early 2027.