A (0:02)

There has been so much attention paid to the AI boom in the United States and China. Japan has sort of been ignored. Should it be? Japan has in recent years launched a leading edge AI lab, a leading edge semiconductor foundry, and several integrator startups. Over the past week I've had another set of conversations with various companies and people in Tokyo. Mostly I explored the question how is the AI boom going in Japan? I recognize that talking to a few people does not a country profile make, but in today's video, a small vibe check on the Japanese AI boom. We can split up AI progress in Japan into two categories, deployment and development. Let us start with deployment for Japanese consumers. Is Generative AI being Used? I can certainly say that the AI buzzword marketing is prolific, much like in the United States. It's been leaking into everything. Everything is AI now, often with ample side helpings of ict. They run these video adverts and taxis. One is an intriguing but inscrutable series featuring a very attractive actress interviewing a rabbit puppet. Maybe I should adopt that format. There are also several commercials blaring out the importance of AI. One is run by this organization called guga, which stands for association to Generalize Utilization of Generative AI. It apparently administers tests for generative AI Passport, a certification for avoiding generative AI risks. I cannot attest to its effectiveness. I do think that ordinary consumers are using generative AI tools. ChatGptified Ghibli style images can be seen on social media here. LLMs are used for translation regularly. Corporate reports mention that some percentage of companies use AI to generate images for marketing, saving time. And it does seem like programmers in Japan, like their peers in the United States, are, are rapidly adopting coding AI tools like Claude code. I expect that adoption to continue the extent to which they are being used. However, I need to spend more time on in a future visit. Anecdotally, my impression is that Japanese programmers are more limited in how much Claude code they can use. Something that I cannot help but notice, however, is that all of these generative AI tools are foreign made. Which makes me ask where are the popular domestic AI models? Let's get back to that one later. The state of corporate deployment of AI and generative AI in particular, remains a mixed bag. In certain domains, AI remains underutilized. In fact, general information technology and processing remains underutilized. I visited an HR outsourcing company here, and not a small one. And a big topic was how competitors struggle to automate internal processes. It's all people work. For instance, there is a significant tax form that Japanese have to fill out and file with the government and they told me how they were the only ones to have automated the form with an online wizard, which feels like something that should have been done 20 years ago. I know it's unfair to pick on tax stuff, but yeah. I asked them if they use AI to automate their internal work processes and they told me about how they trained a small classifier to categorize the tens of thousands of emails they receive from various clients each month. This is part of the challenge of discussing AI in corporate Japan. In the workforce, AI still often means this sort of small AI, basic classifiers and image recognition models. I then clarified that I was asking about generative AI like ChatGPT and was told that that stuff was not even close to being on their radar. In their eyes, the human should always be in charge. Underline that last part. I suspect that bringing generative AI to all of Japan's corporates might take a generation. A few forward thinking companies are pushing employees to use AI in a top down manner, but the smaller businesses still lag far behind. The second part of the equation is AI development, and in this I worry that Japanese firms are making the same mistakes in AI that they made in software some generations ago. Today there is an armada of system integrator businesses that deploy AI into corporate data and government workplaces. I was shown a case study of how a large power company worked with an AI system integrator to train a model to replicate the button presses of a human operator at a particular power plant. The model works very well and human labor was saved. This is a good business for the system integrators. Some of Japan's most valuable startups are actively pursuing these consulting gigs and I do not blame them for taking these jobs. Money is money and isn't for deployment system integration what Palantir is doing. I'm just joking guys don't come after me. But these are small AI models like the email classifier I mentioned earlier. I'm reminded of how Japanese software companies heavily niche themselves down with extremely custom software solutions. I discussed this in a prior video. With this excessive customization, software companies cultivated Galapagos syndrome, making software that fell behind global standards and got increasingly more expensive to maintain. My biggest worry is that the big LLMs or the agent layers wielding them eventually enable competitors to vastly outperform these little AI models. I lean to big general solutions, not these small niche ones. Another thing is revenue. Richard Katz, who writes a substack that I like called Japan Economy Watch, points out that Japanese companies tend to look at information technologies as more way to cut costs and improve productivity. This is in contrast with executives at American companies, which seem to see IC technologies as opportunities to generate more revenue. I don't know if a system integrator can help clients with that. I would like to see more Japanese companies trying to use AI to grow the pie in this I feel like Japan might be well suited for the AI materials and drug discovery spaces. Big Japanese companies already have their own systems to discover new items and probably also have a wealth of data lying around. There are AI materials discovery startups all over the world, but in Japan a few that have caught my eye are Matlantis and MI6. Someone in the semiconductor space also told me about Ixtel, which spun out of Nagoya University. I tried to reach out to them but got a nada. Guys read my message. Sakanaai's AI scientists might also have some potential here though, though I've not heard much about its achievements. Hopefully that changes soon. One of the key ingredients to creating new AI models is compute. Japan is building a few AI data centers, but nothing like what the US and China are doing. Reuters last month reported a data center cluster in Toyama with future capacity in total of 3.1 gigawatts, which doesn't feel big compared to Stargate's 10 gigawatts. They might be restricted by the power supply situation, though with Japan there is a conceptually easy fix. About half of Japan's nuclear power plants are currently idle because of Fukushima. Turning them back on, which I know can be quite difficult, can help. And then there are the chips and systems for that. I think Japan's hardware strengths can help. Google's TPU chips show the potential benefits of not needing to pay the Nvidia tax. There's another AI company building their own hardware, Preferred Networks, one of Japan's largest and most valuable AI startups. Founded in 2014, they started with convolutional neural networks and the like, before pivoting into transformers and now LLMs. In 2017, they worked with a Kobe University professor to design their own chips. The MN Core series, the first and second of which came out surprisingly quickly. After ChatGPT, they began working on a new line of 3D stacked processors focusing on LLM inference. The Mncore L1000. These L1000s will be ready by 2027. I reckon that they will build a few data centers with these guys, and while it seems like this hardware will mostly run their own models, they said they will make it possible to run other companies models too. And I think that's an intriguing twist because it opens the door for them to run inference on various closed and open source models, and that can be a legitimate business. How big? That's dependent on many other things, of course, but it's intriguing. Another of Japan's big issues when it comes to AI development has been talents. While there are indications that Japanese on the whole are not as digitally savvy as their East Asian peers, it does seem like Japanese have the raw skills. Their high school students Math, science and problem solving scores rank quite high. And the top tier of Japanese talent should be able to hold their own on the world stage. Japan is certainly capable of producing people who can do groundbreaking AI work. I was surprised to learn that the popular AI framework Pytorch drew strong inspiration from the ideas of an older Japanese framework called Chainr. Chainr was developed by the aforementioned preferred networks. Someone also pointed out to me the Japanese seem to perform well in Kaggle, the data science and machine learning community. Kaggle's skills and success do not quite map 1 to 1 to AI algorithm research capability, but I would say it is indicative. That said, it is true that Japanese are not well represented in the tight community of X centric AI theorists and researchers, which is interesting because X Twitter is popular in Japan. I do also think that it is inevitably true that the absolute top tier Japanese AI thinkers, programmers, will want to go to the United States, considering how much more they can get paid there. Japanese companies, even the top, most highly valued startups like Sakana AI, can never offer the same tier of salaries that the American giants can. But there will always be some who would prefer to live in Japan rather than the United States. And I think that can be of benefit. The role of government in Japan has mixed views here. On the one hand, people seem to believe that the Japanese government listens to what the industry is saying and tries to respond. Moreover, the government is a rich source of business. There's some anticipation for the new Prime Minister's fiscal expansion plans, which signal a lot of government and defense business coming down the pike. Yet at the same time, the government has its own way of doing things, their own priorities. And I can sense that frustrates some in the Japanese AI community. I'm sure that one thing that that the government probably hears a lot from industry is that private companies lack the resources to train big LLMs and make new products with them. The government has been willing to help with the former. They're putting efforts into a few programs to promote Generative AI, A prominent one being the Geniac project, which stands for Generative AI Accelerator Challenge. It is a program by Medi where the government shares some of the cost of training an LLM and that has allowed companies like Rakuten to train new models. Korea is running a variant of this game too, eliminating entrants every six months. Bloomberg and a few other netizens are calling it the AI Squid game. Games and programs like this are a way to demonstrate fairness, especially when using taxpayer funds. They want to spread things out across various players. So the government is plenty fine with putting public money into R and D and development, not so much into productization and commercialization which they see as for private venture. But Japanese companies don't always have the margins to fund this, leading thus to iffy products. The Japanese government has been pushing the concept of sovereign AI. So the phrase comes up often in my talks. My understanding of it is that when a Japanese sovereign AI model is trained, it is trained with Japanese data controlled by Japanese entities. In addition, those models are trained and inferenced in data centers within Japan. The idea is that Japan can maintain resilience in its AI infrastructure. I get and do not want to downplay the benefits, especially considering the ongoing rare earths kerfuffle, but there does seem to be an over focus on sovereign AI. It is a variant, not the end goal, and I think Japanese companies should be aiming higher. Once the data is collected and data centers built. You train by essentially running a script. What new competencies or leading edge knowledge are we building here? Who will use a sovereign AI unless they are mandated to by a regulation or restriction? True, I can see times when the regulation is right. For instance, preferred networks in December 2025 announced that their Playmo Translate model is being adopted by the Japanese government for translating administrative documents. You would probably want a sovereign AI for that, but for many situations you want the best. And if your companies are to be globally competitive, and Japan's top companies are indeed that, then they need to use the best. And I don't think a Japanese sovereign AI will ever be the best, if only because I doubt there is enough good Japanese data out there. Like I can't imagine a modern LLM ever being world class in math or science problems without training. Data bought from mainland China. Japan's semiconductor industry of the 1980s is Exhibit A of this. One of the major reasons for their decline was that the Japanese company stuck with Japanese tools even as the latter declined. It was ride and die Japan and they died. The Japanese chip guys have learned their lesson? Rapidis is Japan's semiconductor champion, the new one, and funded with taxpayer dollars. And even so, they went out and bought an ASML lithography machine rather than a Nikon machine because ASML have the best. Sorry to the Nikon folks, they make a great machine. I visited the Nikon Museum and had a blast. Side note, I know I was a bit skeptical about Rapidis, especially at the start, but they got their 2nm fab up in Hokkaido, are feeling out customers and seem to have a focused strategy all in just three years. That deserves some props. Japan really does know how to do hardware. In a way, Rapidis shines a path forward for Japanese AI. Go straight for the leading edge. Go out and find the best technology you can get your hands on. IBMs. In this case, bring it in and start improving on it. It's how Japan kickstarted itself in the Meiji Revolution days and thereafter. Why not do the same for AI? Anyway, just my two cents. Before we conclude, I want to say that I hope to have more conversations with people in the Japanese AI community in the future. If you are in that community and want to swap dots, please contact me. We can talk next time I am in town. So, reflecting on this trip, the thing that I think Japan most lacks compared to Silicon Valley is ambition. A recurring topic in my conversations has been Japan's digital deficit, the significant amount of Japanese GDP flowing out of the country to foreign software providers, mostly American ones. This digital deficit is substantial, measured at about 24 billion in the first half of 2025 and it is still growing. Note that for all of 2023 the deficit was only just $37 billion. It's not clear how they're going to close this gap unless Japanese companies somehow make software products that can go head to head with those of the American tech companies. Most Japanese people considering this notion sort of just think, well then that is impossible and sort of give up before they start. Except the inevitable, right? Meanwhile in the valley, there's an immense flush of ambition and entrepreneurial energy kicked off by a variety of things, including ChatGPT's success and growth. They're guys who want to disrupt both ASML and TSMC at the same time. Think about that. It would be silly to ask the Japanese to have the same rah rah as the Bay, but I think the Japanese AI boom can do with a huff or two of their ambition. SoftBank Preferred Networks and Sakana AI shouldn't be the only big Japanese non government organizations with a spine. All right, everyone that's it for tonight. Thanks for watching. Subscribe to the channel. Sign up for the Patreon and I'll see you guys next time.