Japan's Fugaku Supercomputer Trains Largest Domestic AI Model

Japan's RIKEN research institute has used its world-renowned Fugaku supercomputer to train the country's largest domestically developed foundation AI model, marking a significant milestone in the nation's push for technological sovereignty. The initiative positions Japan as a serious contender in the global AI race, which has been overwhelmingly dominated by American and Chinese players.

The project, known as Fugaku-LLM, represents a collaborative effort among some of Japan's top research institutions and private companies, signaling a coordinated national strategy to reduce dependence on foreign AI infrastructure. Unlike models from OpenAI or Meta, this model was purpose-built to excel in the Japanese language and reflect Japan's unique cultural and linguistic nuances.

Key Facts at a Glance

• Fugaku-LLM is Japan's largest domestically trained foundation model, featuring 13 billion parameters
• The model was trained on RIKEN's Fugaku supercomputer, which delivers 442 petaflops of peak performance
• A multi-institutional collaboration involving RIKEN, Tokyo Institute of Technology, Tohoku University, Fujitsu, and CyberAgent
• The model has been released as open source, enabling broad adoption across Japan's research and business ecosystem
• Fugaku uses ARM-based A64FX processors from Fujitsu rather than NVIDIA GPUs, demonstrating an alternative path to AI training
• The project aligns with Japan's national AI strategy to build sovereign AI capabilities

Fugaku Proves Supercomputers Can Train Competitive AI Models

The Fugaku supercomputer, located at RIKEN's Center for Computational Science in Kobe, Japan, held the #1 position on the global TOP500 supercomputer rankings from June 2020 through June 2022. While it has since been surpassed by newer systems like the U.S. Department of Energy's Frontier and Aurora supercomputers, Fugaku remains one of the most powerful computing systems on the planet.

What makes this project particularly noteworthy is Fugaku's hardware architecture. Unlike the vast majority of modern AI training infrastructure — which relies heavily on NVIDIA A100 or H100 GPUs — Fugaku is built on Fujitsu's custom A64FX ARM-based processors. These chips were originally designed for high-performance scientific computing, not deep learning workloads.

The research team had to develop specialized software optimizations to make large-scale language model training feasible on this unconventional hardware. This achievement demonstrates that GPU clusters are not the only viable path to training large AI models — a finding with significant implications for nations facing GPU supply constraints due to U.S. export controls.

Japan's Strategic Push for AI Sovereignty

The Fugaku-LLM project is not just a technical exercise — it is a strategic national initiative. Japan has grown increasingly concerned about its reliance on American AI platforms, particularly as geopolitical tensions reshape global technology supply chains.

Several factors are driving Japan's urgency:

• Language limitations: Models like GPT-4 and Llama are primarily trained on English-language data, often producing suboptimal results in Japanese
• Data sovereignty: Japanese enterprises and government agencies are wary of sending sensitive data to foreign cloud providers
• Economic competitiveness: Japan's $4.2 trillion economy needs domestically controlled AI tools to maintain its edge in manufacturing, robotics, and healthcare
• Export control concerns: U.S. restrictions on advanced chip exports to certain Asian nations have heightened awareness of supply chain vulnerabilities
• Cultural preservation: Japanese language, with its complex writing systems (kanji, hiragana, katakana), requires specialized training approaches

The Japanese government has backed these efforts through its AI Strategy 2022 and subsequent policy updates, allocating billions of yen toward domestic AI research and infrastructure. Prime Minister Kishida's administration has positioned Japan as a leader in 'AI with guardrails,' emphasizing safety and trustworthiness alongside capability.

How Fugaku-LLM Compares to Global Foundation Models

At 13 billion parameters, Fugaku-LLM is considerably smaller than frontier models from Western labs. OpenAI's GPT-4 is estimated to contain over 1 trillion parameters, while Meta's Llama 3 offers variants up to 70 billion parameters. However, raw parameter count does not tell the full story.

The Fugaku-LLM team focused on training data quality over model size. The model was trained on a carefully curated corpus of Japanese-language text, including academic papers, government documents, news articles, and web content. This targeted approach means the model can outperform much larger multilingual models on Japanese-specific benchmarks.

Benchmark results show Fugaku-LLM achieving competitive scores on Japanese language understanding tasks, including:

• JCommonsenseQA: Common-sense reasoning in Japanese
• JNLI: Natural language inference for Japanese text
• JSQuAD: Japanese reading comprehension
• MARC-ja: Sentiment analysis on Japanese product reviews

In several of these benchmarks, the 13B model matches or exceeds the performance of larger multilingual models that were not specifically optimized for Japanese. This validates the strategy of building language-specific foundation models rather than relying on one-size-fits-all solutions.

Open Source Release Accelerates Japanese AI Ecosystem

Critically, RIKEN and its partners chose to release Fugaku-LLM as open source, making it freely available for research and commercial use. This decision mirrors the open-source strategies employed by Meta with Llama and Mistral AI in France, and it could have an outsized impact on Japan's AI ecosystem.

Japanese companies have historically been slower to adopt generative AI compared to their American counterparts. A domestically developed, open-source model removes several adoption barriers. Enterprises can deploy Fugaku-LLM on their own infrastructure, fine-tune it for specific industry applications, and maintain full control over their data.

Early adopters span multiple sectors. Financial institutions are exploring the model for regulatory document analysis. Healthcare organizations are testing it for medical record processing. Manufacturing companies see potential in using it to analyze technical documentation written in Japanese.

The open-source approach also enables Japan's vibrant startup ecosystem to build on top of the model. Companies like CyberAgent, which participated in the training effort, are already integrating insights from the project into their own commercial AI products.

The Broader Trend: Nations Building Their Own AI Models

Japan's Fugaku-LLM project reflects a global trend toward national AI sovereignty. Across the world, countries are recognizing that dependence on a handful of American AI companies poses strategic risks.

France has championed Mistral AI, which has raised over $400 million to build European-made large language models. The UAE developed Falcon through its Technology Innovation Institute. China has produced numerous domestic models, including Baidu's Ernie and Alibaba's Qwen, partly in response to U.S. technology restrictions.

Even within Europe, the EU AI Act has sparked discussions about whether the continent needs its own foundation models to ensure regulatory compliance and data protection. Germany, through its LAION project, has contributed to open-source AI training datasets.

Japan's approach is distinctive in its use of a government-funded supercomputer rather than privately financed GPU clusters. This model could be replicated by other nations that have invested in supercomputing infrastructure but lack access to massive GPU farms.

What This Means for Developers and Businesses

For developers and businesses operating in or with Japan, the implications are significant. Fugaku-LLM provides a high-quality, open-source alternative to proprietary models for Japanese-language applications. Companies can avoid vendor lock-in with major cloud AI providers while achieving strong performance on Japanese NLP tasks.

The project also validates an important technical principle: domain-specific and language-specific models can outperform larger general-purpose models within their target area. This has implications beyond Japan. Organizations worldwide should consider whether fine-tuned, specialized models might serve their needs better than the largest available general-purpose systems.

For the global AI industry, Fugaku's ARM-based training success opens conversations about hardware diversity in AI infrastructure. As NVIDIA GPUs face supply constraints and rising costs, alternative computing architectures could become increasingly relevant.

Looking Ahead: Japan's Next Steps in AI

RIKEN and its partners have signaled that Fugaku-LLM is just the beginning. Plans are underway to scale up future models with more parameters and larger training datasets. Japan is also investing in next-generation supercomputing — Fugaku's successor is already in the planning stages, with potential deployment by 2030.

The Japanese government continues to increase AI-related funding, with reports suggesting additional budget allocations exceeding $1 billion over the next several years. Partnerships between academia, government, and industry — a hallmark of Japan's technology strategy — will likely deepen.

The critical question is whether Japan can translate this foundational research into commercial products and services that compete globally. The country's track record in robotics and semiconductor manufacturing suggests it has the industrial base to do so, but the speed of AI development means the window for establishing competitive advantage is narrow.

As the AI landscape continues to fragment along national and regional lines, Japan's Fugaku-LLM project stands as a compelling case study in how nations can leverage existing supercomputing assets to build sovereign AI capabilities — without relying on the dominant GPU-centric paradigm that defines Silicon Valley's approach.