Why is Japan so behind in AI?

"Why is Japan so behind in AI? They gave us robots, bullet trains, the PlayStation. And yet in the AI race they barely show up. What actually happened?"

Actually Japan didn't lose the AI race. It owns the layer the race can't run without. The assumption is AI=ChatGPT=LLM. Under that narrow definition, sure, Japan is nowhere. But AI is not one thing. It is a stack.

Private AI investment in Japan was about $0.93B in 2024 (Stanford HAI 2025), while the US and China count in the tens of billions. Rakuten AI 3.0, marketed as Japan's largest homegrown model and funded by a state program (GENIAC, the government generative AI subsidy), turned out to be built on China's DeepSeek V3 architecture. The open source community found the proof in the Hugging Face config files. Japan's flagship "national" model is borrowing another country's spine.

Yet, every Nvidia chip ever made passes through Japanese hands. Japan supplies 100% of the world's EUV photoresist (extreme ultraviolet, the light sensitive chemical used to etch chips below 7 nanometers), and over 70% of all photoresist. Tokyo Electron holds around 90% of the coater developer market. Shin Etsu and SUMCO supply roughly half the world's silicon wafers. And in November 2025, Japan's trade ministry added 12 core semiconductor materials to export controls aimed at Chinese firms. Translation: Japan can choke a rival's entire chip line just by refusing to sell. No language model lab on earth has that kind of leverage.

The robotics and physical AI layer is where Japan is already ahead. Around 38% of the world's robots are made in Japan. FANUC, Yaskawa and Kawasaki dominate industrial robots. More importantly for the humanoid wave, Japan owns the parts you cannot substitute: Harmonic Drive and Nabtesco on precision reducers (the gearboxes that let a robot joint move accurately), Yaskawa on servo motors. When Jensen Huang says robotics is having its ChatGPT moment, the body of that robot is Japanese, even if the brain is American.

So why did Japan lose the model layer?

1. Japan won the age of atoms and lost the age of bits. Frontier AI rewards huge risky capital bets, fast failure, breaking rules. The Japanese system rewards consensus, caution, lifetime employment, slow capital. The exact traits that built the Shinkansen are the traits that cannot build a frontier model.

2. No data flywheel. Japan never had a dominant software platform like Google or WeChat, so it never accumulated the user data, the cloud, or the army of software engineers a frontier model needs.

3. Talent and language. English is the working language of AI research, and Japan's best tend to leave for American labs.

And the cleanest proof of all: Japan's most aggressive AI capitalist chose not to build in Japan. Masayoshi Son sold SoftBank's entire Nvidia stake for $5.83B, went all in on OpenAI (around $30B committed), and became a lead backer of Stargate, the $500B US AI infrastructure project. The man who knows Japanese AI best concluded you cannot win the model race from Tokyo, so he became the financier and deployer instead of the developer.

And it has a survival reason to bet on physical AI over models. Japan's population has fallen for 14 straight years. There are over a million jobs it cannot fill. For Japan, robots are not a productivity tool, they are a demographic necessity. So the state put roughly $65B behind AI and chips, wrote physical AI and humanoids into its national AI plan in late 2025, and is targeting 30% of the global physical AI market by 2040.

If the next decade of AI is physical and embodied, Japan stands upstream of everyone, including the US.

If AI stays pure software, Japan stays a customer.

The question was never "what happened to Japan."

What layer of the AI stack do you think is most undervalued right now?

Source: dotdotnews, Odaily, Hugging Face config (Rakuten/RakutenAI-3.0), , , , ,