Japan's Fugaku Supercomputer Trains 500B Parameter LLM

Japan's national research institute RIKEN has trained a 500 billion parameter large language model using the Fugaku supercomputer, one of the world's most powerful computing systems. The initiative represents Japan's most ambitious effort yet to build a sovereign AI foundation model optimized for the Japanese language — and signals a broader global trend of nations investing in homegrown AI capabilities.

The model, developed in collaboration with several Japanese universities and research partners, aims to address the well-documented performance gap that Western-trained LLMs exhibit when processing Japanese text. Unlike English-centric models such as GPT-4 or Meta's Llama 3, this Japanese-native model is designed from the ground up to handle the complexities of kanji, hiragana, katakana, and the nuanced grammar structures unique to the Japanese language.

Key Facts at a Glance

• Model size: 500 billion parameters, making it one of the largest non-English LLMs ever trained
• Hardware: Trained on RIKEN's Fugaku supercomputer, which features over 158,000 ARM-based A64FX processors developed by Fujitsu
• Purpose: Sovereign Japanese-language AI for government, enterprise, and research applications
• Training data: Curated Japanese-language corpus spanning academic papers, government documents, web data, and cultural texts
• Comparison: Roughly comparable in scale to GPT-3's 175B parameters but nearly 3x larger, though smaller than GPT-4's rumored 1.8 trillion parameters
• Significance: Japan's largest domestically trained LLM to date

Fugaku Brings Unique Hardware to LLM Training

Fugaku first claimed the top spot on the TOP500 supercomputer rankings in June 2020 and has remained one of the world's elite computing systems since. Built by Fujitsu in partnership with RIKEN, the system is located at the RIKEN Center for Computational Science in Kobe, Japan.

What makes Fugaku unusual in the LLM training landscape is its architecture. While most large language models — including those from OpenAI, Google, and Anthropic — are trained on clusters of NVIDIA GPUs (typically A100 or H100 accelerators), Fugaku relies on ARM-based A64FX processors. These CPUs were originally designed for high-performance scientific computing, not the matrix multiplication workloads that dominate deep learning.

RIKEN's engineering teams had to develop custom software optimizations to make the A64FX architecture efficient for transformer-based model training. This included novel parallelism strategies and memory management techniques that could prove valuable for future non-GPU AI training approaches. The effort demonstrates that GPU dominance in AI training, while overwhelming, is not absolute.

Why Japan Is Betting Big on Sovereign AI

Japan's push to develop its own large language model reflects a growing global movement toward AI sovereignty — the idea that nations should control their own AI infrastructure rather than depending entirely on American tech giants. Several factors are driving this strategy.

First, existing Western LLMs consistently underperform on Japanese-language tasks. Japanese is a morphologically rich language with 3 writing systems, extensive honorific structures, and contextual meanings that English-trained models struggle to capture. Even multilingual models like GPT-4 show measurable accuracy drops when handling complex Japanese text compared to English.

Second, Japan faces significant data privacy and security concerns. Government agencies, healthcare systems, and financial institutions need AI tools that can process sensitive Japanese-language documents without routing data through foreign cloud providers. A domestically trained and hosted model addresses these concerns directly.

Third, Japan's AI strategy aligns with its broader economic revitalization goals. The Japanese government has allocated over $13 billion toward semiconductor and AI initiatives as part of its effort to reclaim technological leadership in key sectors.

How Japan's Model Compares to Global Competitors

Japan is far from alone in the sovereign AI race. Multiple nations and regions are now developing their own large language models tailored to local languages and cultural contexts.

• France: Mistral AI has raised over $600 million and developed competitive open-weight models with strong European language support
• UAE: Technology Innovation Institute built Falcon 180B, one of the largest open-source models
• China: Baidu's ERNIE, Alibaba's Qwen, and DeepSeek have all developed massive Chinese-language models
• South Korea: Naver's HyperCLOVA X targets Korean-language applications with hundreds of billions of parameters
• EU: The EuroHPC initiative is funding multilingual European LLMs across member states

RIKEN's 500 billion parameter model places Japan among the top tier of sovereign AI efforts globally. However, parameter count alone does not determine model quality — training data curation, alignment techniques, and inference optimization all play critical roles in real-world performance.

Compared to leading commercial models, the RIKEN model occupies an interesting niche. It may not match GPT-4 or Claude 3.5 on general English benchmarks, but its Japanese-language specialization could make it significantly more accurate for domain-specific Japanese applications in law, medicine, and government administration.

Technical Challenges of Training on Non-GPU Hardware

Training a 500 billion parameter model on CPU-based architecture presents formidable engineering challenges that RIKEN's team had to overcome. The dominant paradigm in AI training relies heavily on the massive parallelism offered by GPU tensor cores, which are purpose-built for the matrix operations underlying transformer architectures.

Fugaku's A64FX processors compensate with extremely high memory bandwidth — each processor features 32GB of HBM2 memory with bandwidth exceeding 1 TB/s. This high-bandwidth architecture, originally designed for scientific simulations, proves surprisingly useful for the memory-intensive operations in large model training.

RIKEN's researchers implemented several key optimizations:

• Custom distributed training frameworks adapted for Fugaku's Tofu interconnect network topology
• Mixed-precision training techniques modified for the A64FX's SVE (Scalable Vector Extension) instruction set
• Novel data pipeline architectures to keep the massive processor array efficiently fed with training data
• Checkpoint and recovery systems designed for training runs spanning weeks across 158,000+ nodes

These innovations could have implications beyond Japan. As the global AI industry faces GPU shortages driven by insatiable demand from companies like OpenAI, Google, and Meta, alternative training hardware pathways become strategically important.

What This Means for Developers and Businesses

For developers and businesses operating in or with Japan, RIKEN's model could reshape the Japanese AI ecosystem in several important ways.

Enterprise applications stand to benefit most immediately. Japanese corporations have been slower to adopt generative AI compared to their American and European counterparts, partly because existing tools handle Japanese poorly. A high-quality, domestically hosted Japanese LLM removes a major adoption barrier.

Government digitization efforts could accelerate significantly. Japan's bureaucratic systems still rely heavily on paper-based processes, and a powerful Japanese-language AI could help automate document processing, citizen services, and regulatory compliance at scale.

For international developers, the model may eventually become available through APIs or open-weight releases, providing a powerful tool for building Japanese-language applications without fine-tuning English-centric models. This would be particularly valuable for translation services, cross-border business tools, and cultural content applications.

The model also creates opportunities in the academic research community. Japanese researchers across disciplines — from materials science to social sciences — could leverage the model for literature review, data analysis, and hypothesis generation in their native language.

Looking Ahead: Japan's AI Ambitions Extend Beyond a Single Model

RIKEN's 500 billion parameter model is likely just the beginning of Japan's sovereign AI journey. The Japanese government has signaled plans to develop next-generation computing infrastructure specifically designed for AI workloads, potentially including GPU-based supercomputers that could train even larger models more efficiently.

Fujitsu, RIKEN's hardware partner, is also developing dedicated AI accelerator chips that could complement or eventually replace the A64FX for machine learning tasks. The company's roadmap includes next-generation processors with native support for transformer workloads.

Several Japanese tech companies — including NEC, Preferred Networks, and SoftBank — are also investing in their own AI model development, creating a competitive domestic ecosystem that could produce specialized models for industries like automotive manufacturing, robotics, and healthcare.

The broader implication is clear: the era of AI monoculture dominated by a handful of Silicon Valley companies is giving way to a more distributed global landscape. Japan's investment in Fugaku-trained LLMs demonstrates that meaningful AI development can happen outside the NVIDIA-centric, American-dominated paradigm — even if it requires significant engineering creativity to do so.

As nations worldwide recognize AI as critical infrastructure on par with energy and telecommunications, expect more sovereign AI initiatives to emerge. Japan, with its combination of world-class computing hardware, deep research expertise, and a linguistically complex language that demands specialized solutions, is positioning itself as a leader in this new chapter of global AI development.