Source: NVIDIA Blog — 2026-07-15
Summary
NVIDIA detailed a partnership with Noetra Corp., backed by Japan's Ministry of Economy, Trade and Industry, to build a Vera Rubin AI factory pairing 13,750 NVIDIA Vera CPUs with 27,500 Rubin GPUs, delivering 140 megawatts of data center capacity. The facility underpins Japan's FRONTia Project to train open, trillion-parameter-scale multimodal foundation models for physical AI — robotics, digital twins, and intelligent manufacturing — with the resulting model weights shared broadly with domestic developers alongside NVIDIA's Nemotron, Cosmos, Isaac GR00T, and NeMo software.
Key Takeaways
- The facility's hardware footprint: 13,750 Vera CPUs paired with 27,500 Rubin GPUs, running on NVIDIA's DSX platform, drawing 140 megawatts of data center capacity — giving a concrete sense of scale for what a national-level "AI factory" actually means in hardware terms.
- It's explicitly positioned to train trillion-parameter-scale models, not just serve inference — meaning the capacity is aimed at foundation-model training runs for physical AI, not primarily at hosting existing models for API access.
- Output isn't kept proprietary: pretrained model weights get shared broadly with domestic Japanese developers, alongside NVIDIA's own open software stack — Nemotron (language models), Cosmos (world foundation models), Isaac GR00T (robot foundation models), and NeMo (training framework) — lowering the barrier for smaller Japanese firms to build on top of the trained models rather than starting from scratch.
- The project is framed as supporting Japan's AI robotics strategy through 2040, tying a single hardware buildout to a multi-decade national industrial policy rather than a one-off compute deal.
- Funding and backing come from a mix of Japan's METI and domestic industrial leaders, plus Noetra as the operating partner — a public-private structure rather than NVIDIA selling hardware into a purely private data center.
Reel Script
Hook (~18s, 40 words): Japan just committed to a data center with 27,500 of NVIDIA's next-gen GPUs — and the plan isn't to rent that compute out. It's to train open AI models the entire country's robotics industry gets to use for free.
Core Concept (~75s, 165 words): This is what the industry calls an "AI factory" — not a data center that hosts a bunch of unrelated customer workloads, but one built and funded around a specific national goal: training foundation models for physical AI, meaning robots, digital twins, and factory automation systems that need to understand and act in the physical world. The mechanism worth understanding is the pairing: Vera is NVIDIA's CPU, Rubin is the next-generation GPU, and "Vera Rubin" as a combined platform is the successor to the current Grace Blackwell generation — this is Japan committing to next-gen silicon before it's even the current mainstream generation elsewhere. The scale matters because trillion-parameter multimodal models — models that handle language, vision, and robot action all at once — need training runs that only a handful of facilities on Earth can currently do. 140 megawatts is roughly the power draw of a small city, dedicated to one training operation.
Hands-On (~95s, 210 words): Break down the actual numbers, because they're the diagrammable part. 13,750 Vera CPUs plus 27,500 Rubin GPUs — a 1:2 CPU-to-GPU ratio, which is a fairly standard topology for GPU-dense training clusters where the CPUs mostly handle data loading and orchestration while GPUs do the actual matrix math. 140 megawatts of capacity is the number that translates most concretely to viewers: for comparison, that's in the range of power draw for a mid-size town, not a typical office data center. The output side is the more unusual part of this deal structurally: instead of NVIDIA or Noetra keeping the trained model weights proprietary, they're explicitly committing to share them broadly with domestic developers — paired with NVIDIA's existing open stack: Nemotron for language, Cosmos for world-simulation models used to train robots in simulation before deployment, Isaac GR00T as a robot foundation model, and NeMo as the training framework tying it together. That's the flow worth sketching: national compute investment → trillion-parameter training run → open weights → distributed to domestic startups and manufacturers who couldn't otherwise afford to train anything at this scale themselves.
Takeaway (~22s, 48 words): This is less "NVIDIA sells more GPUs" and more a test case for whether national AI infrastructure investment pays off by lowering the floor for an entire domestic industry. Worth watching whether other countries copy this open-weights-out structure instead of keeping trained models locked to the funding government.