Jensen Huang in Tokyo: Japan’s $6.2B Bet on Physical AI and Nvidia

Nvidia CEO Jensen Huang spent July 15 and 16 in Tokyo, leaving with agreements that span Japan's entire tech stack: a national AI data center, robotics platform partnerships with companies like Fanuc and Yaskawa, and supply deals with chip-material firms powering Nvidia's next chips. Japan is committing up to 1 trillion yen ($6.2 billion) over five years to build what it calls physical AI, foundation models designed to run robots, vehicles, and factory floors, all built on American silicon despite the push for sovereign infrastructure.

What happened

Detail	Fact
Visit dates	July 15-16, 2026, Tokyo
Japan’s AI investment	Up to 1 trillion yen ($6.2 billion) over five years
Noetra AI factory specs	13,750 Vera CPUs and 27,500 Rubin GPUs, 140 megawatts, launching 2028
Cosmos robotics partners	Fanuc, Yaskawa, Kawasaki Heavy, Fujitsu, Hitachi, NEC, Sony, SoftBank, Kubota, AIRoA
Japan’s 2040 robotics target	10 million AI-equipped robots across 18 sectors
Global robotics market target	30%+ share of a market Tokyo values at roughly 20 trillion yen (~$133 billion)
Takaichi growth plan	370 trillion yen ($2.3 trillion) in public and private investment by 2040

Huang’s Tokyo stop was the latest in a regional tour that also included a keynote in Taiwan and a 50,000-GPU deal in Seoul last fall. In two days he met the CEOs of Toyota, Fanuc, Yaskawa, Fujitsu, and Kawasaki over lunch, and supply-chain leaders over skewers and whisky at a Kanda izakaya. Prime Minister Sanae Takaichi joined the government’s physical-AI launch by video on July 16, alongside trade minister Ryosei Akazawa.

Three deals worth watching

1. Noetra: Japan’s sovereign AI factory

Noetra is a consortium of roughly 44 Japanese firms, with SoftBank, Sony, NEC, and Honda at the center. Its job is to build AI models that run machines, not language apps. Nvidia will supply “a Vera Rubin AI factory,” a data center built around its next-generation Vera CPUs and Rubin GPUs, expected to go live in 2028.

Noetra’s model roadmap runs in three stages:

1. A Japanese-language reasoning model, fiscal 2026.
2. An omni-modal model handling text, images, video, and audio by 2028.
3. “Real-world Native AI” built to control robots, released to outside developers from 2030 onward.

The political framing is sovereignty. Japan wants its own data, its own compute, and less dependence on foreign infrastructure. The tension, noted by TechCrunch, is that this independence still runs entirely on American chips.

2. Cosmos 3 Edge: Nvidia’s robotics platform goes on-device

Nvidia launched Cosmos as an open-model initiative in May with a group of global AI labs. In Tokyo, it unveiled Cosmos 3 Edge, a version designed to run directly on its Jetson Thor chips inside robots and factory machines. Ten major Japanese manufacturers, including Fanuc and Yaskawa, said they plan to build on Cosmos models. Honda R&D and Omron are already using the tools. Some partners are testing a shared robot control system together.

3. Toyota: chips in cars and on the factory floor

Toyota committed its next-generation vehicles to Nvidia’s Drive platform at CES in January 2025. The newer work extends Nvidia further: into manufacturing simulations used to design production lines, into vehicle software, and into traffic-reading systems. Toyota’s driver-assistance approach still requires a human behind the wheel, a more conservative position than Waymo or Tesla.

Why it matters

Japan has a shrinking workforce and a government that has made AI and semiconductors the centerpiece of its national growth strategy. The country’s AI Robotics Strategy, released in March, targets more than 30% of the global AI robotics market by 2040. Noetra’s factory, which Nvidia calls “the world’s first national AI infrastructure,” is where models at the scale of trillions of parameters would actually be trained.

For Nvidia, Japan is not just a customer. It is a validation story. If Japanese manufacturers, who invented the precision factory systems the rest of the world copied, adopt physical AI at scale, that becomes a powerful reference for every other industrial nation watching.

The historical footnote is worth noting. Thirty years ago, a $5 million investment from Sega helped keep a near-bankrupt Nvidia alive. Today the relationship has flipped: Japan’s manufacturers need Nvidia’s chips to stay relevant in the next industrial era.

Our take

Physical AI (AI models designed to perceive and control machines rather than generate text) is a real category, not just rebranded robotics. The Noetra specs are concrete enough to take seriously: 27,500 Rubin GPUs in one facility is a meaningful amount of compute for training large industrial models. The 2028 launch date gives this some grounding in reality rather than pure aspiration.

That said, sovereign AI built on a foreign chip monopoly is a structural contradiction Tokyo has not solved. If Nvidia restricts exports again, or if U.S.-Japan trade dynamics shift, the entire factory depends on a supply chain Japan does not control. Businesses building AI workflows right now should watch whether Cosmos 3 Edge becomes the de facto standard for industrial robot software, because that would matter for anyone selling automation or factory software into this market. Our AI integration work increasingly touches exactly these decisions, specifically which model layer to build on when the ground is still moving.

For context on how other frontier labs are positioning around hardware sovereignty, see our earlier piece on Alibaba’s 2.4 trillion parameter Qwen3.8-Max.

What to do about it

1. If you sell software or automation services to manufacturers, check whether your stack is compatible with Nvidia’s Jetson Thor or Cosmos API, because that is where Japanese and likely European industrial buyers will standardize first.
2. Watch Noetra’s fiscal 2026 reasoning model release. It will signal how capable Japanese-language industrial AI is before the bigger 2028 buildout.
3. Track whether Cosmos 3 Edge adoption spreads beyond Japan. If Fanuc and Yaskawa ship products using it, global competitors will follow quickly.

Physical AI is moving faster than most business buyers realize. Getting familiar with the Nvidia Cosmos ecosystem now costs nothing; scrambling to catch up in 2028 will.

Source: TechCrunch · AI