Japan’s Robot AI Replays a 1980s National Computer Bet

Japan’s industry ministry picked Noetra Corp. and the national lab AIST on June 30, 2026 to build a homegrown AI model for robots. The first-year budget is ¥387.3 billion (about $2.4 billion), and the state will only lock in fiscal 2026 and 2027 unless the work clears a yearly review.

This is the same ministry that spent 11 years and ¥54 billion on a national AI computer in the 1980s, then shut the project in 1992 after the machines found no market.

Fifteen Bids, Two Winners, a Two-Year Contract

The Ministry of Economy, Trade and Industry, working with the New Energy and Industrial Technology Development Organization, said Noetra and the National Institute of Advanced Industrial Science and Technology will lead research on a domestic multimodal foundation model. That model is meant to read language, images, video, audio, and sensor data together so a machine can interpret a room and then move.

METI’s English notice puts the case in industrial language. For Japan, physical AI that uses on-site data from a wide range of industries is “a winning path,” and a domestic model is required “while safeguarding on-site data.” The same notice flags Japan’s low energy self-sufficiency and the need to cut the power that AI burns.

NEDO ran the call from March 24 through April 22, 2026 and reviewed 15 applications for the work. The named implementers are Noetra and AIST. NEDO will sign contracts only for fiscal 2026 and 2027. From fiscal 2027, it will run a stage-gate review each year and can stop the project. The stated window is fiscal 2026 through fiscal 2030. Noetra’s own notice dates the research from June 2026 to March 2031, which is the same span in calendar form.

The government has given the effort a public name, FRONTia. NEDO also ties it to GENIAC, the compute-and-community program it has run since February 2024 to seed domestic foundation-model work. Noetra is a new company sitting on top of that stack, not a lab that has been training robot brains for a decade.

HOW THE TENDER UNFOLDED

1. January 7, 2026: Noetra is founded in Shibuya as Japan AI Foundation Model Development Co., with Hironobu Tamba as president.
2. March 24, 2026: NEDO opens the multimodal-model call, which closes on April 22.
3. June 1, 2026: The company takes the Noetra name.
4. June 30, 2026: METI and NEDO name Noetra and AIST as the implementers.
5. July 1, 2026: Noetra says research and development begins.
6. July 16, 2026: METI stages a Tokyo event with Tamba, AIST, NEDO, Economy Minister Ryosei Akazawa, and NVIDIA chief Jensen Huang, plus a video from Prime Minister Sanae Takaichi.

The 13 other bids were not named. That silence is part of the design. Once the state chooses the integration layer for pilots, data formats, and safety checks, suppliers tend to build around the winner because that is where the contracts sit.

MITI Tried This Once Already

From 1982 to 1992 the Ministry of International Trade and Industry, METI’s predecessor, funded the Fifth Generation Computer Systems project through the Institute for New Generation Computer Technology. The Information Processing Society of Japan’s computer museum records that the state put about 54 billion yen over 11 years into parallel inference machines meant to handle knowledge, not just arithmetic.

The hardware was homemade. The software bet was logic programming. Five parallel-inference models were built, including machines with 512 and 256 processors, and some of that kit now sits in the National Science Museum in Tokyo. The project advanced parallel processing and trained a generation of researchers. It did not produce a computer the market wanted. After it ended, MITI stopped funding large-scale computer research of that kind, and the momentum drained away.

Year one of FRONTia already spends about seven times that entire 11-year budget. If the five-year envelope of up to 1 trillion yen is paid in full, the new project would spend about 18 times as much. The institutional rhyme is still loud. A ministry vehicle, a national lab, a cluster of electronics and auto names, and a promise that Japan will own the next thinking machine.

THE 1982 PROJECT AND THIS ONE

The lesson METI appears to have kept is narrow. Do not invent the chips. Do not keep the weights inside a closed lab. Do not write a 10-year blank check. The rest of the playbook, a national champion fed by the usual industrial names, is intact.

27,500 GPUs, Online in June 2028

Noetra will not train the big model on Japanese silicon. Honda, Sony Group, SoftBank Corp., and NEC said the company will use domestic compute at first, then, with NVIDIA, put 27,500 NVIDIA Rubin GPUs and 13,750 Vera CPUs into a factory built on NVIDIA’s DSX platform and Spectrum-X Ethernet. NVIDIA puts the site at 140 megawatts. Honda’s joint notice says construction starts in April 2027 and that the system should run from June 2028.

That lag is the sharpest fact in the file. The tender is already awarded. The first models are supposed to appear during fiscal 2026. The dedicated racks are two years behind. Early training will ride on machines other Japanese operators already run, which is a practical start and a reminder that “national infrastructure” is, for now, a purchase order.

NVIDIA’s newsroom spelled out the rack count on July 16, 2026, the same day Huang was in Tokyo with Akazawa.

Japan is laying the foundation for the future of physical AI.

Partnering with Noetra and supported by METI, the NVIDIA Vera Rubin AI factory built on the NVIDIA DSX platform will bring world-class AI infrastructure to manufacturing, robotics, healthcare, and beyond.

✅ 27,500…

— NVIDIA Newsroom (@nvidianewsroom) July 16, 2026

Huang told that audience that Japan’s plant knowledge, the vocabulary of takumi, kaizen, kanban, and genba, is national property that should not be handed outward, and that physical AI’s next industrial turn should be built in Japan. The hardware that will actually train the model is still his. Takaichi’s video at the same event called overseas dependence in domestic AI development a problem, then welcomed a “domestic model” development base. Both things can be true at once: the weights may be Japanese, the boards will not be.

Preferred Networks president Daisuke Okanohara is also Noetra’s chief of development, and engineers from PFN, SoftBank, NEC, and AIST are being folded into the same room. That is a denser technical bench than ICOT had. It is still one room, and the chips still arrive from Santa Clara.

Who Holds the Factory Data?

SoftBank Corp. president Junichi Miyakawa said the data inside Japanese industry is the competitive asset, and that Noetra’s job is to make a place where that data can be used inside the country. Honda is in the core group because the model is supposed to move bodies, cars, and robots, not chat. NEC said it can take a model from a blank training run through to a working system. Sony Group brings cameras, sensors, and entertainment robots. Tamba’s line at the NVIDIA announcement was simpler, and shorter than the budget.

Bringing physical AI into the real world requires enormous computing, data and foundational technologies

Hironobu Tamba, CEO, Noetra, July 16, 2026 announcement

Around those four sit a further 40 investors. Named users in the original brief include Mitsubishi UFJ, Sumitomo Mitsui, Mizuho, Nippon Steel, and Kobe Steel. Banks do not weld robots. They sit on factory-loan books, insurance, and settlement data, and they will care whether a state-backed model becomes the default way those plants log video, torque, and defects. Steel mills care because a model that reads furnaces and rolling lines from Japanese sensor feeds is a product they would rather not import.

WHO PUTS MONEY AND DATA IN

• SoftBank Corp.: Core investor, and the firm that says it will supply AI infrastructure while arguing that industrial data must stay onshore.
• Honda: Core investor, putting plant and mobility practice into a model that is supposed to act in the physical world.
• NEC: Core investor, offering end-to-end model work from a blank training run through to deployed systems.
• Sony Group: Core investor, with cameras, sensors, and machines that already live in homes and on factory inspection lines.
• AIST: The public lab, tasked with longer-horizon research and with ties to institutes in Japan and abroad.
• Preferred Networks: Not a core shareholder in the notices, but its president is Noetra’s development chief and its engineers are inside the build.

Akazawa went further at the July 16 event. He said Japan’s “winning path” is that it is a super-aged society and a disaster-prone country, which means it holds unusual volumes of data from elder care, disaster response, busy plants, and the Fukushima Daiichi decommissioning site. That is a grim inventory, and it is also the one dataset US chat labs do not already own. The hard part is turning plant-floor access into training-ready data that lawyers will let leave the building, a gap that will matter more than the logo on the GPU crate.

Climate Bonds With a Yearly Kill Switch

The fiscal 2026 slice is not a science-grant curiosity. METI listed the multimodal-model project among GX promotion measures for the year, which means the robot brain is being paid for through the green-transition account. METI’s GX pages describe about 20 trillion yen in government support meant to pull through more than 150 trillion yen of public and private GX investment. Putting a foundation model on that list treats energy use, plant efficiency, and robot substitution as climate work. It also means the largest new line in that account is a compute project whose own racks will draw 140 megawatts.

Officials have described public funding of up to 1 trillion yen over five years. NEDO’s public decision does not print that ceiling. It prints the two-year contract and the yearly gate. Investors who treat the full sum as money in the bank will be marking a speech, not a signed schedule. The real calendar is the stage-gate, because a missed milestone can zero the following year.

THE MONEY THAT IS ACTUALLY SIGNED

• Fiscal 2026 slice: ¥387.3 billion is in the industry ministry’s books for the domestic foundation model that is supposed to lead into physical AI.
• Two-year lock: NEDO will contract fiscal 2026 and fiscal 2027 only, even though the project is described through fiscal 2030.
• The gate: From fiscal 2027, continuation is decided each year by a stage-gate review.
• The envelope: Officials have talked about up to ¥1 trillion over five years; that figure is not the same thing as the signed two-year work.

Takaichi used a much larger number in her July 16 video, saying the cabinet’s growth strategy assumes 80 trillion yen of public and private investment by fiscal 2040 in physical AI, including semiconductors, and that FRONTia is the first step. That 80 trillion yen figure is a whole-sector map, not Noetra’s budget. Mixing the two is how a two-year contract becomes a moonshot on a slide.

The First Model Lands Before the Racks Do

Noetra’s public roadmap is staged on purpose. Fiscal 2026 is a reasoning model with a heavy Japanese-language brief. By fiscal 2028 the target is an omni-modal system that handles text, images, video, and audio together. Fiscal 2030 is “real-world native AI,” a model that is supposed to understand space and physical properties and run on machines that move. Trained weights are supposed to go out to domestic model shops and users so they can grow specialist systems on top, which is the mothership theory in plain clothes.

METI’s Satoshi Nohara told the July 16 hall that the wider robot policy still aims to put 10 million AI robots into Japan by 2040, across 18 sectors that include manufacturing, infrastructure, health care, caregiving, disaster response, and defense. The growth-strategy target sitting behind that deployment number is more than 30 percent of the global AI-robot market by the same year. Yutaka Matsuo, who chairs the government’s AI strategy expert panel and teaches at the University of Tokyo, called physical AI a fight Japan cannot afford to lose, because the labor shortage is already here and because robots and plants are the one AI field where Japan still has a shop floor.

Tamba, speaking later in August 2026, said a lineup like this will not come together again, and that if the effort fails there is no next chance. That is the Fifth Generation hangover in one sentence. Stuffing Honda, NEC, Sony Group, and SoftBank Corp. into a single national vehicle is how Japan has always tried to win a computer race. China let many robot and model shops fight, and most of them died. Japan is paying one champion, then reserving the right to cut it off each year after 2027. The early models will be judged on rented computers. Construction of the GPU factory is scheduled to start in April 2027, the racks are due to run in June 2028, and the first stage-gate comes before either date can cover a missed year.