Direct answer – What is Noetra’s 27,500-GPU AI factory?
Noetra and NVIDIA plan a 140-megawatt Japanese AI computing facility with 27,500 Rubin GPUs and 13,750 Vera CPUs. It is infrastructure for training physical-AI models, not a factory automation system manufacturers can deploy today. Construction is scheduled for April 2027, operations for June 2028, and model access will arrive in stages.
NVIDIA said on July 16, 2026 that it will work with Noetra Corp. on a 140-megawatt AI factory in Japan, built with 13,750 Vera CPUs and 27,500 Rubin GPUs.
Noetra says construction of the computing infrastructure is scheduled to begin in April 2027 and operations are expected in June 2028. The company has already begun model development on computing infrastructure operated by Japan-based providers.
The manufacturing distinction matters: this is a national computing and model-development program, not a packaged robot brain ready to install on a line. The hardware may train models for robots, digital twins and AI agents, but factories will still need task data, safety controls, integration and production proof.
Key Takeaways
- Noetra plans 27,500 NVIDIA Rubin GPUs, 13,750 Vera CPUs and 140 megawatts of data-center capacity.
- Construction is scheduled for April 2027, with operations expected in June 2028.
- Noetra targets a reasoning model in fiscal 2026, an omni-modal model in fiscal 2028 and Real-world Native AI in fiscal 2030.
- The project has investment from 44 companies and organizations, with manufacturing heavily represented.
- Noetra has not published factory benchmarks, access terms or a deployment schedule for individual manufacturers.
What Noetra and NVIDIA actually announced
NVIDIA calls the project the world’s first national AI infrastructure for physical AI. The planned system will use Vera Rubin NVL72 racks, the NVIDIA DSX reference architecture and Spectrum-X Ethernet networking. Japan’s Ministry of Economy, Trade and Industry supports the initiative through the FRONTia Project.
The intended outputs are open multimodal foundation models for AI agents, digital twins, robotics and other physical-AI applications. NVIDIA says pretrained weights will be made broadly available to domestic developers and enterprises alongside its Nemotron, Cosmos, Isaac GR00T and NeMo software. Neither announcement provides detailed licensing, eligibility or support terms.
Noetra’s joint announcement with Sony, SoftBank, NEC and Honda says 44 companies and organizations have invested. The research group includes engineers from the core members, the National Institute of Advanced Industrial Science and Technology and Preferred Networks, with industrial participants ranging from FANUC and OMRON to DMG MORI and Yaskawa.
Why 27,500 GPUs do not mean factory-ready AI
A large compute cluster can train and serve larger models. It does not supply the clean, labeled plant data that teaches a model how a particular weld, fixture, route or recovery procedure behaves. That work remains inside each deployment, which is why our industrial automation stack starts with stable controls and measurable constraints rather than an AI layer by itself.
The phrase AI factory can also confuse two different factories. Noetra’s facility is a data center that produces model capability. A manufacturer’s factory produces parts, assemblies or process output. Moving a model into collaborative robot applications requires edge or on-premise inference, permissions, integration with MES and control systems, validation and a safe fallback when the model is uncertain.
Our read: the size of the cluster is evidence of national commitment, not evidence of plant-floor readiness. The strongest factory AI use cases still succeed on narrow tasks with a visible baseline. Work on natural-language robot training faces the same test: a model becomes useful only when it improves a defined metric without increasing intervention, scrap or safety risk.
The 2027 to 2030 timeline manufacturers should watch
Noetra’s schedule begins before the new facility opens. In fiscal 2026 it plans a reasoning foundation model with Japanese-language understanding, logical reasoning and instruction-following abilities for AI agents and language processing. That phase will use existing domestic computing providers.
Construction of the Rubin-based infrastructure is scheduled for April 2027, and operations for June 2028. By fiscal 2028, Noetra aims to develop an omni-modal model that processes text, images, video and audio. By fiscal 2030, it wants Real-world Native AI that understands physical properties such as spatial relationships and is designed for real-world deployment.
Noetra says external provision and release will happen in stages based on research progress and real-world implementation. That wording leaves an important procurement gap: manufacturers do not yet know which models will be downloadable, which will be hosted, who qualifies for access, or what production support accompanies them. The same gap appears in current Physical AI funding claims, where model ambition arrives before useful-hour and intervention data.
What manufacturers should verify before deployment
Start with one task and one operating number. A robot model should be tested on intervention rate, recovery time, useful-hour rate and cost per completed operation. A scheduling or digital-twin model needs a baseline for lateness, changeovers, throughput or planning time. A model that cannot be scored against the current process is still research.
Ask how plant data moves into the system, whether video and production traces leave the site, who owns improvements trained on that data and whether the workload can run locally when connectivity is limited. Plants should also require version control, rollback, audit logs and a defined human approval point for consequential actions.
Finally, separate model access from deployment support. An open weight is not an integrated application. Manufacturers still need tooling, safety review, system integration and someone accountable for failures. Noetra can lower the national compute barrier; it cannot remove those local engineering obligations.
Frequently Asked Questions
It is a planned 140-megawatt computing facility in Japan using 27,500 NVIDIA Rubin GPUs and 13,750 Vera CPUs. Its job is to train and operate multimodal foundation models for AI agents, robotics, digital twins and physical AI. It is not a conventional manufacturing plant or an off-the-shelf automation system.
Noetra says construction is scheduled to begin in April 2027 and operations are expected in June 2028. Model research has already started using computing infrastructure from Japan-based providers, so the research program and the new Rubin facility have different start dates.
The roadmap starts with reasoning and Japanese-language capabilities, then adds combined text, image, video and audio processing. By fiscal 2030, Noetra aims for models that understand spatial and other physical properties for real-world use. Specific manufacturing tasks and performance benchmarks have not been published.
Not as a finished production product described in these announcements. Noetra says models will be provided and released externally in stages as research and implementation progress. The announcements do not yet define general availability, licensing, geographic eligibility, integration packages or factory-support terms.