Japan’s physical AI advantage depends on creating new know-how

In May 2026, the AI and Web3 subcommittee of Japan’s Liberal Democratic Party announced a shift in Japan’s artificial intelligence (AI) policy towards greater focus on physical AI, arguing that seeking domestic production across every layer of the AI stack is neither realistic nor strategic.

While general-purpose AI performs intellectual processing on the internet and in digital spaces, physical AI operates autonomously in the physical world through robots and other machines. Between 2013 and 2025, total private investment in AI reached approximately US$757 billion in the United States and US$132 billion in China. Japan attracted only US$7 billion.

Under today’s AI development model — pouring vast sums into ever-larger models trained on massive datasets — Japan is unlikely to catch up in the race to develop large language models and general-purpose foundation models, which the United States and China already dominate.

But physical AI is a different game. Japan’s manufacturing and healthcare workplaces hold a wealth of untapped practical know-how and tacit knowledge. By contrast, US AI companies have powerful foundation models but virtually no workplace knowledge.

Japan produced approximately 46 per cent of the world’s industrial robots in 2022 and was the second-largest market in 2024 for industrial robots, with 44,500 units installed. Combining tacit workplace knowledge with advanced robotics could give Japan a competitive edge, particularly if AI competition shifts from developing general-purpose foundation models to implementing them in society, and from the digital space to the physical world. AI startups are also emerging in Japan’s healthcare sector.

Yet much of Japan’s manufacturing knowledge is both fragmented across individual companies and tacit, making it difficult to standardise and share. This partly explains Japan’s slow adoption of digital transformation. A skilled worker with tacit knowledge can identify the cause of an inconsistency in the production process or even recognise, from at a product’s engineering drawings, which component might be failing. They can also propose several courses of action by weighing a range of factors such as economic loss, time and cost. Companies developing physical AI would have to convert such tacit knowledge into AI-ready data.

Even if Japanese companies develop sophisticated physical AI incorporating this knowledge, Japan must ensure that physical AI does not turn its companies into mere suppliers of data. Once tacit workplace knowledge becomes AI-ready data embedded in physical AI, it ceases to be confined to the company that generated it and can be used by others, including foreign firms.

This helps explain why Japanese companies reportedly feared that US AI executives visiting Japan might seek access to their workplace data. If so, Japanese workplace knowledge could strengthen foreign models and Japan could become a ‘data tenant’ rather than an AI leader.

If Japan’s physical AI runs on foundation models owned by US or Chinese companies, the danger is even greater. If foreign companies restrict access to the underlying models, as Anthropic did when it suspended access to its most advanced models in June 2026 under a US export control directive, Japanese firms could contribute the data, knowledge and expertise yet later get locked out of the AI systems built upon them.

One response to this risk is to develop Japan’s own foundation models for physical AI. Noetra, an AI development company backed by 44 Japanese companies and supported by the Ministry of Economy, Trade and Industry and other bodies, is one such attempt.

The success of physical AI trained on data converted from tacit knowledge accumulated on the shop floor raises another concern about Japan’s long-term competitive advantage. Japanese workers have accumulated valuable tacit knowledge by observing and solving problems, experimenting and sharing their experience. If physical AI removes people from the shop floor, tacit knowledge creation may weaken. Physical AI can record workplace activity wherever it is deployed. Once the technology is globally available, accumulated operational data may no longer give Japan a distinctive advantage.

The harder question is whether AI can generate new tacit knowledge. Hitachi has developed physical AI that can self-repair production lines. Its developers trained the AI on how skilled workers identify problems when a line stops, enabling it to solve problems using their patterns of thought and tacit knowledge.

But this system creates solutions from existing tacit knowledge, not new patterns of thought or knowledge. In the future, AI may be able to create new tacit knowledge by developing new patterns of thought in solving diverse problems. Japan must either develop physical AI with built-in capabilities for tacit knowledge creation or keep humans on the front line to generate tacit knowledge that machines cannot create.

Japan’s winning formula may be the combination of physical AI and robotics. But the goal should not be a factory without people. It should be a factory that never stops learning. To sustain its advantage, Japan needs something akin to a ‘mother factory’ for knowledge creation — a place where people and AI work together to create new tacit knowledge. That may be Japan’s real AI strategy — not simply automating everything it can, but preserving and strengthening the capacity to create new tacit knowledge.

Shigeru Asaba is Professor at Waseda Business School, Waseda University, Tokyo.