Japan's 27,500-GPU Physical AI Plan Starts Before Its Data Center

• Noetra's plan uses existing Japan-based computing providers before its proposed 27,500-GPU Rubin system is ready.
• Construction is scheduled to start in April 2027, with operation expected in June 2028; the announced models and infrastructure are targets, not deployed capabilities.
• A report puts public funding at ¥387.3 billion through March 2027, while the government decision contracts only the fiscal 2026 and 2027 portions of the five-year model program.

Japan has started a nationally coordinated physical AI research program, not a finished national AI facility. Noetra's first model milestone comes before construction of its headline computing system, leaving model delivery, funding reviews and power-intensive deployment as separate tests.

The models come before the Rubin system

Noetra said in its July 16 announcement that it had begun full-scale research and development of a Japan-developed multimodal foundation model. The company will first use AI computing infrastructure operated by providers in Japan. Construction of its dedicated Nvidia system is scheduled to begin in April 2027, and operations are expected to start in June 2028.

Nvidia described the future installation in its announcement as a 140 MW “AI factory” containing 27,500 Rubin GPUs and 13,750 Vera CPUs. The design also specifies Vera Rubin NVL72 racks, the DSX platform, Spectrum-X Ethernet and BlueField data-processing units. A separate infrastructure account repeats those figures but makes clear that the capacity number came from Nvidia.

The distinction between program launch and infrastructure deployment is central to the schedule. Noetra's stated roadmap is to build a reasoning foundation model in phases during fiscal 2026, develop an omni-modal model capable of processing text, images, video and audio by fiscal 2028, and pursue what it calls “Real-world Native AI” with an understanding of spatial and other physical properties in fiscal 2030. A contemporaneous account reports the same order: domestic providers first, Rubin construction from April 2027 and operation from June 2028.

Those dates are development goals. Noetra says external releases will be staged according to research progress and real-world implementation. Nvidia separately says the facility could support trillion-parameter models as it expands. Neither statement demonstrates that the models, the full computing system or the claimed performance exists today.

Domestic model control, one foreign computing stack

Noetra is an attempt to pool industrial data and expertise that are dispersed across Japanese companies. Sony Group, SoftBank, NEC and Honda are its core members, and Noetra says in its release that it has investment from 44 companies and organizations. Its research group draws seconded engineers from investors and core members as well as the National Institute of Advanced Industrial Science and Technology, Preferred Networks and other participants.

The company is not entering an empty domestic model market. A report on the project identifies SoftBank's Sarashina, Preferred Networks' PLaMo and NEC's cotomi as existing Japanese efforts. It says Noetra President Hironobu Tamba, who led development of SoftBank's large language model, described the new company as a way to combine fragmented work. Tamba said Japan could create an alternative to US and Chinese AI systems; that remains an ambition rather than a demonstrated competitive result.

The domestic-control case is clearest for models and data. SoftBank's chief executive said the project should create an environment in which Japanese industrial data can be used securely inside the country. NEC's chief executive presented collaboration as a way to expand model choices and improve economic security. Both are participant claims about the program's intended benefits.

At the infrastructure layer, the plan concentrates rather than diversifies dependence. The announced facility uses Nvidia CPUs, GPUs, networking, data-processing units, reference architecture and software. Japan may gain more control over where industrial data is processed and how its models are distributed while relying on one US supplier for the largest planned compute system.

The distribution terms are also unresolved. Nvidia calls the planned models open and says pretrained weights will be broadly available to domestic developers and companies alongside its own Nemotron, Cosmos, Isaac GR00T and NeMo software. Noetra promises staged external releases but does not identify a license or define what users may inspect, modify or deploy. The eventual terms will determine how much practical control moves beyond the 44 participants.

The public commitment has gates

The project has substantial state support, but the retained evidence does not tie all of that money to the data center. The report says Noetra has been allocated ¥387.3 billion, about $2.4 billion, in public funds through March 2027. The company and Nvidia announcements do not state the facility's procurement price, total capital cost or what share of the public allocation would pay for it.

The official program decision gives the funding process a narrower horizon than the five-year research roadmap. The New Energy and Industrial Technology Development Organization says it reviewed 15 applications for the multimodal foundation-model program and selected an implementation team for a fiscal 2026-through-2030 program. But the current decision contracts only fiscal 2026 and fiscal 2027. From fiscal 2027 onward, revisions and continuation will be decided through annual stage-gate reviews.

That history also limits Nvidia's description of the project as the world's first national infrastructure for physical AI. It is a company superlative without a definition or independent comparison in the retained sources. Japan was already supporting access to foundation-model computing and an AI development community through the GENIAC program from February 2024. What is new here is the physical AI focus, the broad industrial coalition and the scale of the planned Nvidia system.

Scale comparisons need care as well. The comparison calls Noetra's Rubin order substantial but modest beside Microsoft's eventual plans for data centers containing hundreds of thousands of Vera Rubin systems. That establishes the direction of the gap, not a clean ratio: the account does not define the “system” unit or provide comparable power and cost figures.

The first decision arrives before the GPUs

Noetra's fiscal 2026 reasoning model is the earliest evidence available to the annual funding process. The relevant questions are what it can do, how it is licensed and whether companies outside the core group use it—evidence of whether the consortium can turn seconded engineers and existing domestic compute into a shared asset before the dedicated facility exists.

The infrastructure test comes later. None of the retained sources identifies a site, power supplier or construction budget for the 140 MW installation. Nvidia Chief Executive Jensen Huang said in the report that Japan would need much more data-center and power-grid infrastructure to meet AI demand, underscoring a constraint without explaining how this project will secure its own capacity.

By April 2027, the relevant evidence will be whether construction begins and whether the next program stage passes review. By June 2028, it will be whether the Rubin system operates at the announced scale. The central question is not whether Japan can announce 27,500 GPUs, but whether Noetra can deliver models that justify continued public support and give domestic users meaningful control despite the project's dependence on Nvidia infrastructure.