Japan just claimed the world’s first state-owned AI factory

Japan’s Noetra consortium and Nvidia are building a 140MW AI factory with 27,500 Rubin GPUs as the world’s first state‑tendered national AI infrastructure, slated to power the FRONTia physical‑AI programme. ¥387.3 billion in first‑year subsidies has been committed, with up to ¥1 trillion over five years subject to stage‑gate reviews.

The facility will train open multimodal foundation models for robotics and industrial automation, a sovereign‑compute bet that directly challenges the dominance of US hyperscalers and closed cloud platforms. Japan aims to capture more than 30% of the global AI robotics market by 2040.

Japan is building the world’s first state‑tendered national AI factory — a 140‑megawatt data center powered by Nvidia’s latest Rubin GPUs and Vera CPUs, designed to train open models for physical AI. The Noetra consortium, backed by SoftBank, Sony, NEC and Honda, won the public tender on June 30, 2026 and is now moving to lock in hardware and a five‑year funding horizon.

The project, called FRONTia, marks a deliberate bid for sovereign compute. Rather than depend on US hyperscalers or Chinese state clouds, Tokyo is creating a domestic AI stack with the explicit requirement that pre‑trained model weights be shared openly with Japanese developers. That architecture — a hybrid of public money, industry consortium, and open licensing — has no direct parallel among major economies.

The factory built on Rubin‑Vera racks

The facility will house 382 Vera Rubin NVL72 racks, each packing 72 Rubin GPUs and 36 Vera CPUs, for 27,500 graphics processors and 13,750 central processors. Nvidia’s Spectrum‑X Ethernet and DGX SuperPOD reference design tie the system together. The raw compute is meant to handle trillion‑parameter models as the factory scales.

Behind those numbers, the Rubin GPU performs the billions of parallel calculations that foundation models require. Vera CPUs orchestrate data flows and manage non‑GPU tasks. In an NVL72 rack, the close‑coupled design lets the GPUs and CPUs share memory and communication paths, making large‑batch training far more efficient than stringing together generic servers.

Morgan Stanley estimates each NVL72 system costs $5 million to $7 million, putting the rack hardware outlay at $1.9 billion to $2.7 billion before infrastructure. That sum is a financial analyst’s calculation — neither Nvidia nor Noetra has confirmed the figures — and excludes the cost of the data centre shell, power supply and cooling.

Jensen Huang, Nvidia’s chief executive, framed the project as a return to Japan’s manufacturing roots. “Japan invented modern manufacturing,” he said in a statement. “Now, it is building the AI factories that will power the next industrial revolution.”

The consortium’s roadmap runs through fiscal 2030: a reasoning foundation model by fiscal 2026, an omni‑modal model handling text, images, video and audio by fiscal 2028, and “real‑world native AI” with spatial awareness by fiscal 2030. Pre‑trained weights will be shared broadly — a condition of the NEDO tender that aims to seed a national ecosystem rather than lock models inside a few corporate silos.

The sovereign compute race is heating up

Japan’s Noetra project represents a distinctive model for sovereign AI: a state‑tendered facility with a legal mandate to share model weights, operated by an industry consortium rather than a single government body or hyperscaler. This approach differs markedly from the strategies pursued by other major economies.

The United States has built frontier AI capacity through private hyperscalers — Microsoft, Google, Amazon and Meta — assembling clusters that often exceed 100,000 advanced GPUs. These deployments dwarf Japan’s 140 MW facility in raw scale, yet they lack a single, nationally branded “AI factory” designation or a requirement to open‑source model weights for domestic industry.

South Korea is advancing sovereign AI through the K‑AI initiative and large Nvidia‑backed clusters deployed for domestic firms, but has not yet created a state‑tendered facility with the explicit specification and official labelling that characterizes Noetra’s project. The European Union funds distributed compute through EuroHPC and is exploring AI mega‑clusters, but its model remains fragmented across member states; no single installation is designed explicitly as an open‑weight, physical‑AI factory under a unified government tender.

Japan’s hybrid structure — combining state procurement, industrial consortium governance, and mandatory open‑source licensing — allows a mid‑sized power to build frontier compute without ceding control to a single hyperscaler. The catch is the stage‑gate funding model. METI and NEDO will review technical milestones before releasing the bulk of the ¥1 trillion, and a missed checkpoint could slow the rollout just as other nations accelerate.

The real lock‑in, however, may be hardware timing. Nvidia’s Rubin generation moves to volume production in the second half of 2026. By committing 27,500 GPUs now, Japan secures a substantial allocation before global supply tightens under the weight of hyperscaler demand. If the facility expands to trillion‑parameter training, the early reservation could prove decisive. If the funding stalls, the racks still arrive — but the software ambitions would need to be trimmed to match the hardware on hand.

Beyond the headline

The Bigger Picture

Japan’s AI factory anchors a broader move toward sovereign compute. Tokyo is blending public R&D funds with an industry consortium and mandatory open‑weight licensing, creating a path that sits between the American hyperscale model and China’s state‑directed ecosystem. The result will influence how other mid‑sized economies build frontier AI capacity without surrendering control to foreign platforms.

The Reach

For global industrial automation suppliers, the Noetra factory creates a direct competitive dynamic. As home‑grown multi‑modal models mature, Japanese robotics and control‑system makers will have first access to models trained on domestic data, eroding the edge that Western off‑the‑shelf solutions currently hold. European and US industrial software exporters may soon find that compatibility with Japan’s physical‑AI ecosystem becomes a condition for staying inside its manufacturing supply chain.

The Timing

The announcement comes as Rubin GPUs enter volume production and as Japan’s AI Robotics Strategy moves from plan to execution. With hyperscalers building ever‑larger cloud clusters, Tokyo needed to secure a substantial GPU allotment before supply tightened. Committing to 27,500 Rubin chips and a five‑year funding window gives Japan a seat at the training table during the current hardware cycle — not a waitlist for the next one.

The ripple from Tokyo’s AI factory

As Japan locks in hardware and an open model mandate, Western companies and investors face a changed competitive map for industrial AI.

• Western semiconductor procurement manager

  Monitor Nvidia’s investor communications and public production updates for Rubin and Vera output — especially any mention of backlog or allocation criteria for sovereign buyers. If further governments replicate Japan’s large‑scale tender, lead times on advanced GPUs are likely to stretch, potentially forcing you to adjust procurement schedules and pricing assumptions.

• US-based investor with APAC emerging market exposure

  Watch the valuations of Japanese industrial‑automation stocks and robotics integrators that stand to gain preferential access to Noetra’s open models. Also track the response of regional competitors in South Korea and China; a widening model‑development gap could shift relative positions in the $133 billion global AI robotics market that Japan has directly targeted.

• European industrial automation solutions provider

  Re‑examine your product roadmaps for the Japanese market. The availability of free, high‑quality, Japan‑specific foundation models may reduce the willingness of domestic manufacturers to license external AI tooling. Consider forming technical partnerships with Japanese consortia or developing compatibility layers that plug into FRONTia‑derived model pipelines.

• Western policy professional tracking AI sovereignty

  Study the Noetra model as a policy template: a transparent tender, a multi‑company consortium, mandatory open‑weight release, and stage‑gate funding. Compare its governance to the EU’s AI Act risk classifications and the US NIST AI Risk Management Framework. Japan’s early move may shape allied norms for sovereign compute before multilateral standards crystallise.

Explainer

Noetra

Noetra Corp. is a consortium founded by SoftBank, Sony, NEC and Honda, with investment from 44 companies and organisations according to NEC. It won the Japanese government’s public tender to build and operate the national AI factory for the FRONTia programme. The consortium was formed specifically to pool industrial resources for training Japan‑origin multimodal foundation models at a scale no single company could afford alone.

FRONTia

FRONTia is Japan’s state‑funded physical AI programme, formally titled “Development of Multimodal Foundation Models with a View to AI Robotics and Physical AI.” Overseen by METI and NEDO, it runs from fiscal 2026 through fiscal 2030 and requires that pre‑trained model weights be broadly shared with Japanese developers. The programme is the policy vehicle for Japan’s ambition to become a global leader in AI‑driven robotics and industrial automation.

Vera Rubin NVL72

Vera Rubin NVL72 is Nvidia’s rack‑scale building block that packages 72 Rubin GPUs with 36 Vera CPUs, interconnected via high‑bandwidth NVLink for tightly coupled parallel processing. In Noetra’s factory, 382 such racks form the compute backbone, enabling efficient training of trillion‑parameter models and physical‑AI workloads. The name combines “Vera” (the CPU architecture) and “Rubin” (the GPU generation) into a single integrated unit that Nvidia sells as a pre‑configured system.

Sovereign AI

Sovereign AI refers to a nation’s policy of building, owning and controlling its own artificial intelligence infrastructure and foundation models, reducing dependence on foreign technology providers. Japan’s FRONTia factory is an example: a domestically tendered, publicly funded computing platform with a legal mandate to share model outputs with local industry. This contrasts with relying on US hyperscale clouds or Chinese state‑run models, giving governments greater control over data, algorithmic priorities and supply‑chain security.