Japan METI GENIAC Phase 2 Compute Awards: 12,000 H100 GPU-Months for Sovereign Foundation Model Startups
Japan’s national push for generative AI independence accelerated today as METI distributed massive sovereign compute resources across premier domestic research labs.
Securing Japan’s Technological Sovereignty in Generative AI
Recognizing that dependency on foreign foundation model providers poses critical long-term risks to national industrial competitiveness, corporate data security, and cultural heritage, the Government of Japan has undertaken aggressive state-backed interventions. Today, Japan’s Ministry of Economy, Trade and Industry (METI) officially announced the recipient organizations and compute allocations under Phase 2 of the GENIAC program (Generative AI Accelerator Challenge).
Backed by a state supplementary budget allocation exceeding ¥80 Billion ($530 Million), GENIAC Phase 2 distributes over 12,000 GPU-months of high-end NVIDIA H100 and B200 computing power to domestic startups, premier universities, and corporate research consortia.
1. Phase 2 Consortium Recipients & Strategic Focus Areas
Unlike general-purpose academic grants, GENIAC compute grants are tied to rigorous, measurable foundation model development milestones:
- Sakana AI: Awarded 2,500 GPU-months to scale its nature-inspired evolutionary model merging algorithms, training dense multimodal reasoning models optimized for Japanese language and cultural nuance.
- Preferred Networks (PFN): Allocated 3,000 GPU-months to develop sovereign scientific foundation models targeting multi-modal materials discovery, battery chemistry, and industrial robotics control.
- NTT Corporation & Tokyo Tech: Dedicated compute to train next-generation lightweight Japanese foundation models capable of running at ultra-low power within optical IOWN telecommunications networks.
- Kotoba Technologies: Developing voice-first spoken Japanese models tailored specifically for call center automation and eldercare humanoid robotics.
2. The ABCI 3.0 Sovereign Supercomputer Backbone
The physical compute for GENIAC Phase 2 is anchored by the newly commissioned ABCI 3.0 (AI Bridging Cloud Infrastructure) supercomputer, operated by the National Institute of Advanced Industrial Science and Technology (AIST). Featuring thousands of liquid-cooled NVIDIA H200 and B200 tensor core GPUs connected over high-speed InfiniBand fabrics, ABCI 3.0 provides domestic researchers with an air-gapped national compute fabric protected against foreign sanctions or supply chain disruptions.
GENIAC Phase 2 Compute Allocation Overview
| Recipient Organization | Allocated Compute | Primary Development Target | Target Industrial Application |
|---|---|---|---|
| Sakana AI | 2,500 GPU-Months | Evolutionary multimodal foundation models | Enterprise reasoning & cultural nuance |
| Preferred Networks (PFN) | 3,000 GPU-Months | Scientific AI (Chemistry & Materials) | Next-gen EV batteries & semiconductor IP |
| NTT / Tokyo Tech | 2,500 GPU-Months | Low-power IOWN optical models | Telecom infrastructure & smart cities |
| Kotoba Technologies | 1,500 GPU-Months | Spoken vernacular Japanese voice AI | Eldercare robotics & public administration |
| Academic Consortia | 2,500 GPU-Months | Open research checkpoints & safety evals | University research & AISI evaluations |
A Coordinated Industrial Policy for the AI Age
Through the GENIAC program, Japan demonstrates how state-directed industrial policy can effectively marshal capital, sovereign compute, and private entrepreneurial talent. By subsidizing the steepest capital hurdles of foundation model training, Japan is ensuring that its manufacturing powerhouses—from Toyota and Sony to Panasonic and Hitachi—are powered by indigenous, culturally aligned machine intelligence.