For companies developing AI models domestically, securing computing resources has long been a persistent headache. While overseas hyperscalers sit on abundant GPU stockpiles, Japanese players have been forced to wait their turn. Then, on July 16, 2026, a consortium called "Noetra," backed by 44 companies including SoftBank, Sony, NEC, and Honda, unveiled a plan centered on 27,500 units of NVIDIA's next-generation Rubin GPU and a 140-megawatt-class data center. Under "FRONTia," a national project led by the Ministry of Economy, Trade and Industry (METI) and NEDO, the government reportedly plans to invest up to ¥1 trillion. But the gaps that must be filled before operations begin in June 2028 are larger than the GPU count alone suggests.
Who Is Noetra? The Identity Behind the 44-Company Alliance
Noetra Inc. is a new company headquartered in Shibuya, Tokyo, that began operations on July 1, 2026. Its representative director and president is Hironobu Tamba, former representative director and CEO of SB Intuitions, who led development of the domestic LLM "Sarashina." A total of 44 companies have invested in the venture, with Sony Group, SoftBank, NEC, and Honda Motor forming its core (these four core companies are included among the 44). The investors span a wide range of industries, from domestic AI developers to manufacturers.
The list of investors also includes AI startups such as Preferred Networks and Sakana AI, as well as Asahi Kasei, KDDI Research, and all three of Japan's megabanks. This article proceeds based on the figure of 44 companies as stated by Noetra's official materials.
In his announcement, Tamba stated that bringing physical AI into the real world poses challenges no single company can solve alone. The development roadmap is divided into three phases: in fiscal 2026, a reasoning foundation model handling Japanese-language comprehension and logical reasoning; in fiscal 2028, an omni-modal foundation model integrating language, images, video, and audio; and in fiscal 2030, a real-world-native AI equipped with spatial awareness and understanding of physical properties. Model development is overseen by Daisuke Okanohara, representative director of Preferred Networks, who also serves concurrently as the joint R&D general manager for the NEDO project and as technical director for AIST's "Exploratory Program." Part of the PLaMo development team has also been seconded to Noetra, hinting at an aim to consolidate Japan's scattered AI talent onto a single computing platform.
Inside the 27,500-GPU Cluster and the 140-Megawatt Facility
The computing infrastructure combines roughly 27,500 NVIDIA Rubin GPUs with approximately 13,750 Vera CPUs, bundled at a 2:1 GPU-to-CPU ratio and built around Vera Rubin NVL72 rack units. The NVL72 design links 72 GPUs via high-speed NVLink to function as a single massive computing device, enabling model training at a scale no single GPU could handle. This is paired with the NVIDIA DSX platform, Spectrum-X Ethernet, and BlueField DPUs. Spectrum-X is a networking platform optimized to eliminate bottlenecks in inter-rack communication, while BlueField DPUs offload data transfer and security processing—tasks traditionally handled by CPUs—onto dedicated chips, freeing GPUs to focus purely on computation.
The 140-megawatt data center capacity NVIDIA disclosed is, in itself, a remarkably large figure. No matter how concrete the GPU procurement plan may be, without an infrastructure capable of stably supplying this much power, the 27,500-unit figure risks remaining nothing more than a paper plan.
What the Term "National AI Factory" Actually Means
Since 2023, NVIDIA has used the term "AI factory" to describe the arrangement whereby companies operate GPU clusters like their own production facilities. This announcement's title claims to be the "world's first national AI infrastructure." However, NVIDIA also partnered with the South Korean government in 2025 to build "sovereign AI infrastructure" and a "National AI Computing Center," meaning government-scale AI computing infrastructure agreements are not, in fact, unprecedented with Noetra. The precise phrasing of this "world's first" claim is "world's first national AI infrastructure for physical AI"—a qualified assertion made by NVIDIA specifically about physical AI. Whether Noetra or METI independently adopted this English phrasing, or whether NVIDIA itself coined it, remains unclear.
In the announcement, NVIDIA CEO Jensen Huang stated that Japan invented modern manufacturing and is now building the AI factories that will power the next industrial revolution. This phrasing reflects a shift underway at NVIDIA: its customer base is expanding from individual companies to entire nations. The buyers of GPUs at scale are broadening from cloud providers and manufacturers to nations themselves. METI Minister Ryosei Akazawa is also reported to have described FRONTia as poised to become the core of Japan's physical AI ecosystem, indicating that the government shares this positioning.
How domestic AI startups not among the 44 investing companies might engage with this computing resource is not addressed in the published materials. For the Japan subsidiaries of overseas hyperscalers such as AWS, Google Cloud, and Microsoft Azure, this marquee "sovereign AI infrastructure" project is now effectively controlled by the framework of the 44-company alliance and METI. NVIDIA's move to turn nations into customers has the potential to reshape the competitive landscape among cloud providers in the Japanese market.
Governance Risk in Public-Private Alliances: Lessons from Two Past National Projects
METI and NEDO plan to provide support of up to ¥1 trillion through fiscal 2030, but it is Noetra and the National Institute of Advanced Industrial Science and Technology (AIST) that will actually run the project—not the government directly owning and operating the infrastructure. The locus of decision-making is not singular. A similar structure characterized the Fifth Generation Computer Project, led by ICOT under the Ministry of International Trade and Industry from 1982 to 1992, which invested roughly ¥54 billion over 11 years in pursuit of artificial intelligence computing. The official final evaluation credited the project with achieving several technical goals—including knowledge base mechanisms and inference/PIM mechanisms (achieving over 150 MLIPS against a 100 MLIPS target)—while noting that some systems, such as the intelligent interface, fell short of their targets.
At the time, a misreading of the technology trend toward PC downsizing is considered to have led domestic manufacturers astray in their decisions. More recently, Rapidus, which aims to mass-produce next-generation semiconductors, is set to receive additional support in fiscal 2026, bringing cumulative government support to an estimated ¥2.9 trillion, with total investment including private funds reportedly exceeding ¥7 trillion. This structure—where the scale keeps growing through annual budgetary allocations—shares with FRONTia a dependence on sustained political commitment.
FRONTia has built in annual stage-gate reviews and manages progress through two pillars—a development track and an exploratory track—which can be read as an effort to avoid repeating the fate of the Fifth Generation Computer Project, which achieved some technical goals but never reached practical implementation. Yet even though Noetra and AIST are the two entities actually running the project, the investment structure behind it is spread across 44 companies spanning diverse industries. Generally speaking, the more investors involved, the more decision-making speed tends to lag behind a single-company-led structure, and such an arrangement is likely to carry coordination costs that a solely corporate-led national project like Rapidus does not face.
The Power Procurement Void: What Comparison with Stargate UAE Reveals About Feasibility
Neither NVIDIA's official announcement, Noetra's official announcement, nor major Japanese media coverage mentions how the 140 megawatts of power will be procured (whether from renewable energy, nuclear power, grid electricity, or on-site generation), nor what cooling method will handle the heat generated during operation. The data center's location is likewise undisclosed. While it is true that SoftBank is proceeding with development of an AI data center on the site of the former Sharp Sakai plant, there is no official announcement confirming this as the construction site for Noetra's computing infrastructure, and it would be premature to link the two.
As a benchmark for scale, G42-led Stargate UAE plans to bring 200 megawatts online in its first phase in 2026, and has announced the adoption of NVIDIA GB300 chips. External estimates have put the deployment scale corresponding to that 200-megawatt capacity at roughly 100,000 units, yet the U.S. Department of Commerce's November 2025 approval of advanced semiconductor sales to G42 caps out at the equivalent of up to 35,000 GB300 units (it is not explicitly stated that this allocation is exclusively for Stargate UAE). Noetra's figure of 140 megawatts and 27,500 units is, if anything, closer in scale to this export-approved unit count—though because GB300 and Rubin belong to different generations, a direct comparison of computing performance is not straightforward.
The difference in governance structure is even more pronounced. Stargate UAE operates under a framework in which G42 handles construction, OpenAI and Oracle handle operations, and Cisco, SoftBank Group, and NVIDIA provide support—whereas Noetra is an indirect model in which a private alliance of 44 companies is mediated through a commissioned project with NEDO, resulting in a different design for both decision-making speed and the locus of responsibility. METI's fiscal 2026 support amount of ¥387.3 billion converts to roughly $2.39 billion at the then-prevailing exchange rate (approximately ¥162/dollar), and the maximum ¥1 trillion over five years equates to roughly $6.17 billion—figures worth keeping in mind when comparing against the scale of investment in Stargate UAE.
The Rubin GPUs NVIDIA plans to deploy are a next-generation product whose mass-production shipment track record has not yet been publicly confirmed as of this writing, and the roughly two-year span from the start of construction in April 2027 to the start of operations in June 2028 carries the character of something closer to a forward contract—announced well before actual operations begin. Whether the 44 companies can align their decision-making within this two-year window is the first question to be answered. Finalizing the source of the 140 megawatts of power and getting Rubin GPU mass production shipments on track must also proceed in parallel within the same timeframe. The figure of roughly 27,500 units is, after all, merely an approximate number as of the time of announcement. Given how the Fifth Generation Computer Project lost momentum due to a misjudgment of technology trends, what will determine Noetra's feasibility is whether the voids of consensus-building and power procurement can be filled before construction begins in April 2027.