30 Years After Sega's $5M Lifeline, Nvidia Signs 22 Japanese Companies into Cosmos: The National-Scale Battle for the "Physical AI Era"
Sources: TechCrunch / 21st Century Business Herald / Tencent News Date: 2026-07-19
---
What Happened
During his July 15-16 visit to Tokyo, Jensen Huang signed three major deals:
1. 22 Japanese companies joined the Cosmos alliance — Fanuc, Yaskawa Electric, Kawasaki Heavy Industries, Fujitsu, Hitachi, NEC, Sony, SoftBank, Kubota, Omron, Komatsu, Mitsubishi Heavy Industries, Mitsubishi Corporation, Mitsui & Co., Honda Research Institute, Tokyo Electron, AIRoA, and more. Most of Japan's industrial sector was brought in. 2. Cosmos 3 Edge launched — a lightweight world model that runs on Jetson Thor chips, performing inference inside the robot body itself without pushing all data to the cloud. 3. Noetra, Japan's national AI infrastructure — led by SoftBank, Sony, NEC, and Honda, with 44 participating companies. Tokyo has committed 1 trillion yen (~$6.2 billion) over 5 years, backed by a "Vera Rubin AI factory" with 13,750 Vera CPUs, 27,500 Rubin GPUs, and 140 MW of capacity, operational in June 2028.
Deep Analysis
Three statements from Huang, denser than any visit in 30 years
On July 16, speaking at METI's "Physical AI launch ceremony" — with Trade Minister Ryosei Akazawa on stage and Prime Minister Sanae Takaichi appearing via video — Huang's core messages were:
> "Japan cannot outsource its national intelligence. Japan must own, improve, secure, and deploy AI that is Japan's own."
> "The next industrial revolution will be made in Japan."
> "Japan's excellence in mechatronics is ready to fuse with physical AI."
These compress three concepts — "sovereign AI," "physical AI," and "made in Japan" — onto Japan in one move. Sovereign AI means countries building their own compute clusters rather than relying on foreign public clouds; physical AI refers to embodied intelligence in robots, cars, and factories; made-in-Japan is the execution foundation.
Noetra's three-phase roadmap
- FY2026 (from this April): reasoning models with strong Japanese-language capability — a successor to SoftBank's "Soratoki" LLM.
- FY2028: full multimodality (text / image / video / audio), launching alongside the Vera Rubin AI factory.
- FY2030: "Real-world Native AI" — foundation models designed for robot bodies, opened to developers in phases.
- Fujitsu + Fanuc: factory physical AI, reflecting production changes and shop-floor conditions in real time for plant-wide production planning optimization.
- Fujitsu + Yaskawa: retail and distribution, integrating sales and inventory data for logistics automation.
- Fujitsu + Kawasaki: healthcare, using Nvidia's Holoscan IGX, Isaac for Healthcare, Isaac GR00T, and Cosmos for surgical assistance, care robots, and in-hospital transport robots.
- Deployment of 10 million AI robots across 18 sectors
- $65 billion in public + private investment (~9.6 trillion yen)
- 30% share of the global AI robot market by 2040
- That market is roughly 20 trillion yen ($133 billion)
- Strong dependency on Nvidia hardware — Noetra trains on the Vera Rubin AI factory and infers on Jetson Thor; vendor lock-in is severe and a single point of failure if export controls escalate.
- Is the "third option" realistic? — Noetra aims to be a choice for Japan and other countries, but Japan's own LLMs (Soratoki, Sarashina) visibly lag the first tier; the 30% market-share goal has a long path.
- Physical AI PMF remains unproven — public data on real deployment costs vs. human workers in factories, logistics, and healthcare is scarce.
- Human-robot relations in the cobot era — how Japan's labor structure changes as 10 million robots enter 18 sectors is a big topic Huang's visit didn't touch.
- TechCrunch: https://techcrunch.com/2026/07/19/what-to-watch-for-after-jensen-huangs-japan-visit
- 21st Century Business Herald: https://new.qq.com/rain/a/20260716A0ALVK00
- Tencent News (commentary): https://news.qq.com/rain/a/20260717A0B9L200
- Nvidia official: https://nvidianews.nvidia.com/news/japans-robotics-and-manufacturing-leaders-build-on-nvidia-cosmos-to-advance-physical-ai-frontier
Key point: Noetra will train foundation models that "robots can run on," not chatbots. Hiroshige Tanpo, who led SoftBank's "Soratoki" LLM, is Noetra's president — his words: "a genuine third option that Japan and other countries can choose."
Cosmos 3 Edge + Jetson Thor: the AI brain inside the robot
Jetson Thor is Nvidia's robot-specific chip, announced in 2024 and mass-produced from 2025. Cosmos 3 Edge is a lightweight world model tailored for Thor — meaning robots can perform most inference on their own bodies, with latency and bandwidth no longer bottlenecks of cloud-based approaches.
The Fujitsu × Fanuc × Yaskawa × Kawasaki collaboration chain:
"Getting most of Japan's industrial sector to sign the same robotics underlying protocol" — the industrial significance of this may be greater than the Cosmos models themselves.
Toyota: Drive platform + factory simulation + conservative ADAS
Japan's largest automaker works with Nvidia on three layers:
| Layer | Content | |---|---| | Vehicles | Next-generation models will use the Nvidia Drive platform (promised at CES 2025) | | Manufacturing | Production line design via Nvidia simulation | | Software | Drive OS running in-vehicle software and road-traffic interpretation systems |
However, Toyota's ADAS strategy is "advanced driver assistance, driver still required" — not the driverless approach of Waymo / Tesla. This detail is underreported in Chinese media but signals clearly in English-language circles that Japan's automotive AI path is more conservative than China's or America's.
National 2040 target: 10 million AI robots
Japan's AI Robot Strategy (published March 2026) targets:
The Takaichi cabinet has made AI and semiconductors the core of its growth strategy, with total public-private investment of 370 trillion yen ($2.3 trillion) across 17 strategic fields through 2040. Noetra's Vera Rubin AI factory is the first substantive piece of this national AI strategy to land.
The 30-year-old story as a mirror
In 1995-96, Nvidia bet wrong on 3D graphics, the NV2 chip collaboration with Sega failed, and the company was near bankruptcy. Sega CEO Akio Ishikawa provided $5 million in emergency funding, keeping Nvidia alive until the RIVA 128 turnaround in 1997.
On July 15, Huang returned to Akihabara — standing on the site of the former Sega arcade, reuniting with current Sega CEO Haruki Satomi and Ishikawa. Thirty years ago, $5 million saved an American GPU company; today, that company is giving back to Japan's entire industrial robotics chain. This isn't sentiment marketing — it's a textbook case of supply-chain diplomacy.
Why It Matters
1. A national AI strategy landing through a single American company for the first time — sovereign AI has been a slogan for years; Noetra is the first plan with specific compute, specific companies, and a specific timeline. 2. A de facto standard for the robot's on-board AI brain — Cosmos 3 Edge + Jetson Thor + 22 Japanese companies; Nvidia's "Android moment" in physical AI may have already happened. 3. Physical AI is the next battleground after LLMs — as LLM dividends converge, embodied intelligence is starting collectively in 2026 H2 (including Nvidia Cosmos and various domestic Chinese players). 4. The Japan-vs-China embodied AI window — Japan has the industrial robot base (Fanuc, Yaskawa, 40 years of accumulation); China has software and data (world models, VLA, data governance). Nvidia is a supplier betting on both sides. 5. $65 billion + $2.3 trillion investment pools — physical AI funding in 2026 H2 vastly exceeds language models. For Chinese embodied/physical AI startups, the direct customer base is growing.
Risks and Open Questions
---
One-Sentence Summary
22 Japanese companies + Cosmos 3 Edge + 1 trillion yen Noetra + a 2028 Vera Rubin AI factory — this is Nvidia's key campaign pushing "physical AI" from concept to national infrastructure, and Japan's biggest bet in the sovereign AI era. The window for a Japan-vs-China embodied AI showdown is officially open.
Source links: