Japan is much further ahead in real AI deployment than most people realize.

Today we talk with David Ha, CEO of Sakana AI, Japan’s clearly most valuable, and arguably most innovative, AI startup, and he explains how Japanese enterprises are using AI in production environments today.

We talk about selling advanced software to conservative enterprises, how Sakana is customizing existing LLMs to fit Japanese sensibilities, and why smaller, more flexible routing models will likely win out over the massive, expensive frontier models.

David also shares some great advice for anyone starting a startup in Japan.

It’s a great conversation, and I think you’ll enjoy it.

Podcast: Play in new window | Download | Embed

Subscribe: Apple Podcasts | RSS | More

Show Notes

  • When it’s best build for Japan instead of (or before) going global
  • How to raise money overseas as a KK and raise large rounds in Japan
  • It’s actually very easy to start a startup in Japan
  • Sakana AI’s three-stage go-to-market strategy
  • The tradeoffs between B2B and B2C strategies in Japan
  • How to drive AI adoption in the Japanese enterprise
  • Designing AI workflows that prevent users from outsourcing decision making
  • Why routing models will win in the market over frontier models
  • What “data sovereignty” actually means in practice in Japan
  • How Sakana AI is adapting existing models to make them reflect Japanese sensibilities
  • David’s best advice to founders in Japan

Links from the Founder

Transcript

Welcome to Disrupting Japan, Straight Talk from Japan’s most innovative founders and VCs.

I’m Tim Romero, and thanks for joining me.

To say that AI is hot right now understates the mania.

AI startups accounted for over 60% of global VC investment in 2025, and things are on track to be even more AI-focused this year. And while investor and media attention remain stubbornly focused on the U.S. and China, things are moving surprisingly fast here in Japan, although things are moving in a slightly different direction.

Today we sit down with David Ha, the co-founder and CEO of Sakana AI, Japan’s most valuable and arguably most innovative AI startup.

Now this episode’s a bit longer than usual, but it is well worth your time, because by the end, you’ll understand why the frontier model companies, the open AIs, the Anthropics, might be pursuing a flawed and very expensive strategy, and why there’s a real chance for Japan to be a global leader here.

In fact, as David explains, enterprise AI adoption in Japan is far more advanced than most people realize. And we’re talking about actual production deployments here, not just POCs and press releases, but actual daily use.

Now David and I talk about raising capital from both Japanese and foreign VCs, the global importance of AI sovereignty, the best way to drive AI adoption in large conservative Japanese companies, and why AI use in Japan is more widespread than you probably expect.

But you know, David tells that story much better than I can. So, let’s get right to the interview.

Interview

Tim: So, we’re sitting here with David Ha, the co-founder and CEO of Sakana AI. So, thanks for sitting down with me. I really appreciate it.

David: Awesome. Yeah, it’s actually really great to have a chat with you, Tim. I mean, as you know, I’ve been following your podcast for more than a decade before I started a company.

Tim: I mean, I love to hear that. I honestly do. Nothing makes me happier.

David: Yeah. I mean, I have no idea that I would start a company like 10 years ago.

Tim: Let’s start with the founding, because the Sakana AI team is really an interesting test case in Japan, I think. So, the founding team itself, it’s a mixture of foreigners and Japanese. Tell me about it. Tell me how it came to be, how the three of you got connected and thought this was a good idea.

David: It’s a really unique, I would say, experiment. I founded this company a little bit less than three years ago. It’s myself, David Ha, and Llion Jones, my colleague at Google, who was one of the co-inventors of the transformer architecture, the T in ChatGPT. And when you start a company in Japan, it cannot just be two foreigner guys, especially if you want to be the frontier AI company in Japan. But luckily, I was able to recruit my friend Ren Ito, who used to work at Mercari. And he was sent to Mercari Europe as well to run Mercari, one of the original unicorns.

Tim: So, how did you know Ren?

David: We worked together at different startup events before. Earlier, also, I worked at another startup briefly called Stability AI. At that time, Ren was working in the UK for that company, and I was working as a contractor remotely in Japan. So, we connected there.

Tim: And you were kind of like, hey, I’ve got a better idea.

David: Yeah, something like that. The best way to meet co-founders and colleagues is when you’re working at other startups. And when you’re working at other companies.

Tim: I think that’s so true because it is being able to work with someone and see how they work and having that experience is invaluable when picking a founding team.

David: Yeah, I totally agree. I think in our founding team, not just the co-founders, many of our employees, the founding employees, are people that I’ve personally worked with at Google, at Goldman Sachs in Japan, also at other companies like Mercali and Yahoo in Japan, like people that we know. And I think when you’ve worked with someone before and you kind of know how to coordinate different tasks and work together well, it’s like a great mix.

Tim: Yeah, yeah. There’s fewer surprises, the better, right? So, tell me about the kind of pre-seed team. Was it just you three?

David: Yeah, it was just us. It was crazy. I think all three of us have different motivations of starting this company. I’ll tell you mine. So, our company is founded in Japan to develop AI technology for Japan. Like when I was looking at your podcast like 10 years ago, there’s like many founders in Japan, like they really want to go international right away. They want to move to the valley and try their luck. But there’s also other companies that are focused on Japan. Remember this, one of your podcast episodes, like you have so many startups in your podcast. One of them are like focusing on developing the product that produced the best invoices for Japan. Like those companies.

Tim: That’s a need. But you guys, I mean, I think what’s unique about Sakana is that you had a product that, I mean, yes, it’s targeted to Japan, but it’s truly a global market. And when you first raised money, you went overseas. Your first round was from overseas investors rather than Japan. So why?

David: Yeah, exactly. I mean, for us, having worked at Google before and Goldman Sachs before, I think like my experience is generally easier to raise money from overseas markets.

Tim: Well, I agree with that for sure.

David: Especially in the VC ecosystem. I think in Japan, the VC ecosystem is growing. There’s a support, lots of government support as well. But at the end of the day, if you want to have a strong funding round, we wanted to go direct to Sand Hill Road. So, I think what happened was when I started this company, I co-founded with Llon at the beginning. We just announced that we’re going to start a company called Sakana AI, an R&D company in Japan. And I just basically sat there working on our tech, working out what we’re going to do. And from there, we had many reverse inquiries. Like maybe on Twitter, because I have some following on Twitter as well.

Tim: So, inquiries from VCs?

David: Yes, from VCs. Basically, all of the VCs reached out to us.

Tim: Both Japanese and foreign or primarily…?

David: Many foreigners as well. I worked at Mountain View for Google for a few years, and many VCs knew me at the time. I also had many personal connections in that community. So basically, if you have a choice, you might as well try to pitch to Deluxe Capitals, the Vinod Khoslas, the Andreessens, and the Sequoias of the world, right?

Tim: So, for the seed round, you raised $30 million. But then the next round, you came back to Japan, and you raised something like $200 million from a series of rounds of Japanese investments. And so why the shift back to Japan?

David: For our seed round, it made sense for us to raise from Silicon Valley investors. I think also the nature of the VCs, especially in the seed round, when you raise from prominent Silicon Valley VCs, they want to dominate the entire round. Because the seed round of a company is your highest cost of capital. It’s the lowest valuation if your company continues to exist. So, if a VC has decided that they’re going to bet on you at the seed stage, it makes sense for them to invest as much as possible. So, there was not a lot of room to involve other companies at our seed stage. We did make room for NTT and Sony and KDDI, but we couldn’t. Basically, the first year of our existence is primarily R&D driven. There’s absolutely no plan for monetization. As we started out, we wanted to build a foundational R&D company in Japan. We recruited many of our top research colleagues from Google, Prefer Networks, Stability AI. And we set out to actually produce pretty amazing research in the first year, like the AI scientists. Something like that didn’t exist before.

Tim: And it’s actually got a published research paper now, right?

David: Yeah. I mean, after two years, it got published in Nature magazine, like the top scientific journal.

Tim: That’s amazing.

David: That’s not going to make me money directly as a company, but as a researcher, it’s kind of proud that a Japanese AI company can publish something in Nature.

Tim: Absolutely amazing proof of concept and proof of suitability.

David: So, our company has all sorts of different research directions, especially my co-founder, Llion Jones. He’s working on trying to replace the transformer architecture with something else. But for the AI scientists, we’re really one of the first companies to look at the concept of recursive self-improvements, getting AI to improve itself using foundation models.

Tim: So, I really want to talk about the RSI Lab, but before I get to that, I want to just…

David: Yeah, how we came up with the Series A…

Tim: Well, I’m just curious. So, are you guys a KK or a Delaware company?

David: Oh, yeah. We are a 100% Japanese Kabushiki Kaisha. I registered this company myself in the Homukyoku in Azabu area.

Tim: So, I love this about you guys, because like the common wisdom, which I mean, I’ve certainly said on this podcast many times in the past is that, hey, if you want to raise from overseas, don’t do a KK, no one’s willing to do due diligence on it. And hey, if you ever need to raise more than like, say, $20 million, it’s really hard in Japan, you’re best to go to Silicon Valley. But you’re a KK, you’ve raised over $200 million in Japan. So, am I wrong? Or are you guys just a special case in some way?

David: Let’s explore this topic. Because I think we broke all of the rules. I have a lot of motivation, because a lot of people give me advice, saying you should have a Delaware company or a Cayman Island company. But at heart, I’ve lived in this country for a while. And I love living in Japan. And I want to build a Japanese company. It’s weird saying that as a foreigner. I want to build a Japanese company. And I think if the investors at that time didn’t want to invest because we’re a KK, then I’ll be like, so be it. And I think this goes back a longer way. Going against conventional wisdom. These are not things that I do intentionally. So, when I first came to Japan, I came over due to a job placement from Goldman Sachs. I didn’t know I was going to come to Japan. I thought the job was somewhere else. And they flew me to Japan. I just decided to stay here. It’s such a lovely city. But I joined here during the Lehman crisis, during a really rough time. And at that time, people don’t talk about this, but in Goldman Sachs, Japan, there’s some sort of founder culture. Somehow, many people who worked with me on the trading floor in Goldman Sachs in Japan, in Roppongi Hills, ended up starting successful Japanese companies. One of them were on your show. Remember Akiko-san from Wantedly?

Tim: That’s right.

David: She was a Goldman Sachs. She was at Goldman Sachs. Mai Shin, who started Genda, they listed recently. She was working with me at Goldman Sachs. There’s also Kaie Shinichiro, who started Folio.

Tim: So, the finance industry, certainly in Japan, is not known for launching entrepreneurial careers. So, what was it? Was it just that team or was there something?

David: Somehow around that time, when we had to live through the Lehman crisis, when everyone got fired except for us, we kind of knew how to survive. This is another guy, I think Russel Cumming, who started Paidy as well on the floor.

Tim: So, was he Goldman Sachs?

David: Yeah, yeah. He was on the same floor. The common theme between all of these companies are companies that are started in Japan, tailored for Japan. Like Paidy is like the buy now, pay later service tailored for Japan. Genda is this amazing concept that Shin Mai started where you notice that all of the arcades in Japan were trading below book value. And she started a company to buy all of the Sega arcades. Now they’re doing really great. Folio started by Kaie, right? And Wantedly, there’s…

Tim: A new generation of recruiting, yeah.

David: In some way that inspired me to start a company. There’s also good training as well. When you’re working at this investment bank during these crazy times, I had to fly to US a lot to talk to investors, like the endowment funds and so on and pitch them.

Tim: So, it gets you ready for a lot, I guess.

David: Yeah, yeah. I think it was good training. And when I started a company, I looked at it from first principles. I didn’t take any advice about you should open a Delaware company, or you should have this advisor, or you should talk to this company. I really think in terms of first principles about what I think would have the most impact long-term in Japan.

Tim: So, did Lux and some of the other investors, were they comfortable with a KK?

David: They had to go along with it. Oh, to be fair, Lux Capital has invested in Japanese companies before. They’re a deep tech investor. And I think Costa Ventures, they invested also in deep tech company. The Fusion, Kyoto Fusion.

Tim: Kyoto Fusioneering?

David: Yeah, that company. I mean, Vinod, he invests all over the world. So, he’s a veteran.

Tim: That is really encouraging. I mean, for the ecosystem as a whole, the fact that US VCs are getting more comfortable and more familiar with Japanese corporate structure is a great thing.

David: Honestly speaking, the process of registering a company in Japan, if you read a bit of Japanese, it’s quite simple.

Tim: It’s not that complicated. Yeah, people make a much bigger deal of it than it really is.

David: It’s like a few pages, and it’s done for. You put in some initial capital and you have a KK. You register a shared office and it’s done. So, when people tell me how complex things are in Japan, I think if people can take a few Japanese lessons and you can open a bank account quite easily, you can start a company, you can rent an office. It’s not that hard, right?

Tim: I totally agree. I’m sort of two minds. So, when I’m talking with Medi or the Shibuya government, yeah, any friction you can smooth off is good. That benefits everybody. But when you’re talking to an individual, it’s like, come on, it’s not that hard. It really isn’t. I mean, if you’re having trouble with this, you have no idea what’s coming at you when you try to actually build and grow this.

David: If someone comes to Japan and tries to start a company in Japan, they should at least know how to open a bank account here. They should at least try to rent an office here.

Tim: So, what about the experience raising in Japan? I mean, you had some fantastic investors from your seed round and you raised over 200 million in different rounds from Japanese investors, right?

David: Yeah, we raised, now we’re at a series B, but let’s go back to the series A when I started raising series B from Japanese investors. Like from our seed stage, we raised mainly from, say, Cosla and Lux Capital, and we showed to the world that we can produce meaningful research to the point where when people think of Sakana AI internationally, they don’t think of us as a Japanese company anymore. They think of us as, oh, you guys made this AI scientist thing, or you guys made this evolutionary model merge thing. But as a startup, we have to have a go-to-market strategy. Like the first step is developing a tech and the second step is monetizing it.

Tim: Well, eventually you got to sell something.

David: Exactly, exactly, right? Or get acquired, but we don’t plan to do that. So, my plan is given the success of the first six months to 1 year of R&D, I wanted to raise a series A to have a go-to-market strategy for the company. So, we went up to the Silicon Valley again to get another term sheets to get priced. That’s how the game works. So, you need to have a term sheet and someone needs to price your series A. So, we talked to Lux Costa again. They’re kind of in, but they already led our seed round. So, they cannot really lead our series A. It’s customary to find another lead. So, we pitched to New Enterprise Associates and they decided to leave our series A. So now I have a term sheet.

Tim: So, then you can come back and fill it out.

David: Exactly. But before I did that, we also pitched to NVIDIA. So, I had to pitch to Jensen. And I think NVIDIA had a long history in Japan. And I think they’ve also made previous investments in Japanese companies. The strategy at the time is we wanted the beginning of our business model is to service Japanese enterprises. So, this is kind of a very contrarian approach. Usually for Silicon Valley type company, we should be product first. Develop your own foundation model, develop your own chatbots, make everyone able to use it online and grow your user base organically and cross your fingers. And hopefully that more monthly active users every day. So, that approach might work, but it’s a risky strategy in Japan. Because I feel in Japan, the consumer market might not be as strong.

Tim: I think that’s interesting, particularly in your case, where you have OpenAI and Anthropics, really well-funded competitors running that strategy. You want to be doing something different.

David: Yeah. I mean, I don’t want to go against Google as well at the beginning. So, that was a strategy. What we did instead is we wanted to work with large Japanese enterprises to actually help them build a platform to deploy customized AI workflows to actually really help the productivity of Japanese companies.

Tim: So, how do you avoid getting into the consultant trap?

David: You can’t. That’s the point, right? This is the contrarian strategy. Many people view the SI-er consulting solution as a nasty, unscalable thing.

Tim: I mean, there’s some truth to that.

David: Right. So, where you’re kind of always at the helm of the customer overlords. But you know what? So be it. We wanted to have a very strong enterprise business at the beginning.

Tim: I mean, looking at kind of like a meta model here, are you thinking of something like Salesforce, for example? It’s very consulting heavy, but all the consulting, all of that linearly scaling stuff is done by partners and resellers. Is that kind of what you’re thinking?

David: In my case, the vision I have, this is getting back to our experiment with the AI scientists as well. Back then, I knew that eventually or really soon, many of the AI workflow or even the software engineering workflow or the forward deployment workflows would be heavily done by AI itself. So, I was actually challenging this sort of SIR enterprise model with the conventional view is, okay, in Japan, you have all these SI-er consulting solution. It’s not scalable. You have all of this annoying things you have to do to add value. But from my point of view, the danger is actually the SaaS market at the time where you have these kinds of border plates products that everyone will use and people would demand customization because once software is cheap, everyone wants to have a customized software. So, the silver lining for Japan is this sort of SIR model augmented with AI might actually be the model that the rest of the world might use.

Tim: And so in a sense, actually, the AI itself will make the SI-er model more scalable.

David: Exactly, exactly. We’re still not there yet.

Tim: That makes sense as a direction.

David: I knew back then because we were starting to play with AI agents to do scientific discovery. If a system of AI agents can produce a machine learning research paper, surely it might be able to produce a credit analysis agents or a slide generation agents at the time.

Tim: And so that’s why it’s great to have NEC and Fujitsu.

David: Exactly. And this is a contrarian because many people did not appreciate the enterprise business. They think it’s not scalable. It’s a lot of like a handholding. But for me, it’s not just technological. I think in Japan, to get AI to have any impact, you really have to work closely with the top and the bottom of a company.

Tim: You know, I think this is one thing that Japan is uniquely strong at. Japanese enterprises and Japanese startups are able to work in these long-term partnerships that both sides would consider way too risky in the US, which is a much more transactional type of a culture. But Japan, you see these long-term co-development, co-marketing. There are startups that happily rely on enterprise partners for distribution in ways that makes me as an American nervous. But it works here. And it really is different the way enterprises will work with you.

David: It’s certainly a different operating mode. People are optimized for coexistence rather than winning. But it’s not easy working with large Japanese companies. Even when I was at Goldman Sachs, like the bar is very high. The expectations of the quality is very high. And one way to make it easier to work with large Japanese traditional companies is to have them invest in you as a strategic investor.

Tim: Then you’re on the inside.

David: It’s kind of, not yet, but kind of.

Tim: But you know what I mean? You’re closer, right?

David: You’re closer, right? So, getting back to the Series A story, after a year, we had the term sheet in hand. We got NVIDIA on board. Then me and my co-founder, especially Ren Ito, we pitched to many of the Keidanren companies. And we eventually had strategic investments from Mitsubishi, SMBC, Mizuho, all the three largest banks.

Tim: Big banks.

David: Because in Japan, they kind of run the show. We had NEC, Dai-iichi, Itochu, Fujitsu, you name it. So, we had many of the large Japanese companies invest and then we’re in business. So, we have a research portfolio. We have a strong R&D team. Last year, I started building an applied research engineering team. And we cannot work with all of our strategic partners. We focus on a few of them to get the ball rolling. So, we were able to secure strategic contracts with many customers, but namely Mitsubishi Bank, SMBC, and Daiwa Securities.

Tim: As a research organization with a mind to the future where you know you’re going to have to productize, how did you hire? You’ve got a mix of Japanese engineers and foreign engineers.

David: As the founder of a company, there’s really three most important things you have to do. The most important thing is recruiting. Second most important thing is fundraising. And then business development for the startup.

Tim: That’s the big three, yeah.

David: In that order, actually. So, recruiting is always a challenge because AI talent is in demand. It’s very scarce. For me, it’s really important that our company builds a distinct brand of research. We’re not using our investor money to train the LLM from scratch on day one. We would do that on a smaller scale, but we’re not going to do exactly the same thing as what Google is doing.

Tim: Right. It doesn’t make sense, right?

David: Yeah, we have specific research goals of AI-driven discovery, recursive self-improvement. So, this itself attracts certain researchers to want to work for us. So, it’s not like they’re attracted by any company. They’re attracted by Sakana AI’s research direction and the R&D philosophy. And I feel like having this vision allows us to attract people who are aligned with our R&D strategy naturally.

Tim: Especially since you’re taking a very contrarian approach. You’re going to get people who are aligned. Is that pool big enough? Or do you find you have to reach out to more of an AI generalist?

David: It’s not small. Even though we’re contrarian in a way, but we’re not that contrarian. Think about it this way. When we develop the AI scientists, we’re not using totally new things. We’re using frontier LLMs. LLMs we built ourselves. We’re building agents. These agents are discovering new algorithms. So, this kind of skillset, the people who are able to do this, they’re also able to develop systems to analyze the credit worthiness of a loan applicant or to create automated slide deck for a bank. The skills are not that different. The technology is not so different.

Tim: So, for the team, what’s the mix between Japanese engineers and foreign engineers who you’ve recruited in Japan and foreign engineers who’ve come to Japan to join you?

David: The mix changes from time to time. I would say it’s like 30% internationally and 70% domestically. And I think in the product developments and applied engineering teams, there’s more emphasis on Japanese nationals or people who can speak business level Japanese, like JLPT-N1. Whereas for the research team, it’s more of a foreigner bias because there’s less of a Japanese language requirements. It is like a challenge as well because to steer from a company that has been doing research 100% into applied direction where we’re working on enterprise solutions for like the mega banks. So, you need to build different teams and have those teams work together. So, we started working with many customers, but three in particular, like Mitsubishi Bank, SMBC Bank, and Daiwa Securities. I think the problems we’re solving are quite significant because they’re related to Japan’s change. Japan has gone from deflationary to inflationary economy. The stock market is finally coming up. From that standpoint, Japanese banks have been sitting on all this cash and there’s a need for them to start deploying the capital via lending. So, one of the important projects we work with Mitsubishi Bank is to deploy AI agents on loan analysis. We actually have a deployment called AI Credit Officer with Mitsubishi Bank, where we’re using this technology to not replace humans, but to augment a credit officer’s role so that they can assess.

Tim: What is it looking at? What is it giving the loan officer that he would take a long time to do himself?

David: I mean, the process of processing a loan, it’s quite long. You have to assess the entire application documents. There’s many things in the whole process. The credit officers has to write a credit memo. It’s not a simple yes or no decision. It has to have a reasoning into that decision. And I think this is similar to, in the AI scientists, we have AI reviewer components where you would say, here’s a scientific paper. As a peer reviewer, I would recommend rejection of this paper because of X, Y, Z reasons. And it’s very detailed. So, this kind of technology transfers over to credit reports.

Tim: It’s a preliminary analysis and a speeding up of the workflow in general.

David: Exactly. The thing is, we’re not using AI to automate the acceptance of rejection. We’re actually trying to alleviate the work behind writing the credit memo and also provide the reasoning so that the actual bank officer can agree or disagree with that reasoning.

Tim: I find one of the biggest challenges in designing an AI workflow like this, and by AI workflow, I mean an AI workflow within an enterprise, is it is so tempting for us human beings to just say, yeah, that looks good. To just offload that especially, especially when the AI is right most of the time. So, how do you design these workflows to help fight against that?

David: This is really, really challenging. I think it’s the same with software engineering as well. You can’t simply just accept all of the AI-generated code. There has to be many tests, checks and balances. And in our system, there are also the equivalent of tests, whether you agree or disagree. I think having a human in the loop is very important. The task may be automated with the AI, but the job and the responsibility is owned by the human.

Tim: Well, I know, but I think there is, and this is not so much a technology issue, but I think there’s a real danger of if you have a tool that’s right most of the time. We humans, we get tired, we get busy, we want to get home early and go to our kid’s soccer game. How do you make sure that we humans stay engaged and really are part of the process and not just someone to blame when something goes wrong? You know what I mean?

David: That’s a great question. I think it depends on the culture. In Japan, I really respect the work culture here. I feel that, this may be a generalization, but employees and workers in Japan take pride in their work and their sense of responsibility. So, when you’re assessing a loan, this affects someone’s livelihood. And it’s very important for the loan officer to really do their best to make sure they give the right assessment. So, I think having this emphasis on the responsibility is paramount. It’s not a technology problem, as you said, but this has to be communicated really well.

Tim: So, it might also be like making sure that the loan officer is not just presented with that memo, he’s supposed to check something he’s presented with something he can check it against.

David: Exactly, yeah. I’ll give you another example of this. The other problem we’ve worked on with Sumitomo, SMBC Bank, many banks are tasked with trying to emanate, sell and buy businesses. And especially with the aging population in Japan, there’s many like nice businesses in Japan where the owners have no heirs. There’s no sons and daughters to take over. There could be like a profitable ramen chain somewhere that SMBC might need to find a buyer for. So, one of their issues is many of these owners of these businesses they cannot write their own pitch documents, like slide decks and so on. And they need a megabank to help them with the whole process of writing all the financials, the pitch documents. So, one thing that we work with SMBC on is a document creation agents that’s tailored for SMBC standards.

Tim: Okay, so it’s the document collection and organization.

David: Not just that, the generation as well, generating slide decks in SMBC colors with all of their standards. And this actually helps reduce the time it takes to generate these documents from two to three weeks to two to three hours. And then to your point earlier, we cannot simply use these documents right away because AI is not perfect. It may not convey the same information that the owner or the banker may want to convey. So, the banking staff would tweak the 5% of the document before actually using them for pitching. So, this is an example of where you actually need this human worker to take ownership of the role to actually check what’s being written on the pitch deck to make sure that these are everything you want to convey.

Tim: Business culture, I think, is a big part of the answer to that of making sure that the employees are empowered and respected enough that they have no trouble overruling the AI when they see fit and they have the information to do so. And it’s really interesting to see how these types of things are rolling out. So, what stage are these projects at? Are they still proof of concept or just something they’re in production?

David: It’s quite exciting. So, usually these contracts have three phases. The POC phase, and then the initial enterprise deployment phase, and finally the firm-wide deployment phase. Of the three projects that we have, the main ones, all three of them have moved beyond the POC stage.

Tim: That is so impressive just because, I mean, I’ve worked with Japanese banks. To say they are conservative organizations is understating it. But the fact that they moved from POC to initial deployment is amazing. I mean, they’re clearly seeing real value in it.

David: For this, I also have to give a lot of credit for to the banks themselves. I would actually say that the Japanese mega banks, they’re not behind at all. They’re, in fact, quite ahead, even compared to some of their American peers.

Tim: It’s really an interesting mix, isn’t it?

David: Yeah. I mean, they use cloud technology. They use frontier models. They’re actually heavily investing in AI. Mitsubishi Bank has a chatbot app. So, they’re not that behind. If anything, they might be ahead of their global peers. And I think this is also part of the change that Japan is going through right now. I feel when Japan needs to change, it will change.

Tim: Yes.

David: We’re seeing the start of that happening.

Tim: After exhausting every other possibility, it will change. And when change comes, it comes really fast.

David: Yeah, exactly. So, at the first stage of our go-to-market strategy, we secured large Japanese customers, built solutions that work for them. But the enterprise solution is only one part of the story. But given our success in enterprise, within six months, we’ve already secured many of these contracts and delivered them. We’ve attracted another round of investors. So, seed stage, R&D, mission complete. Series A, enterprise solutions. We showed some success there, mission complete. It’s like a video game. And then the Series B is the next stage. I want to start off, maybe it’s the big boss at some point, but we’re at the Series B stage. So, from the initial success of enterprise deployment, we were able to raise a Series B. This time around, many of the investors are actually not traditional large Japanese companies, but instead they’re like Salesforce, Google, Datadog, Citibank, Santander Bank, and In-Q-Tel, the CIA.

Tim: And for the same strategic reason, to drive that enterprise collaboration.

David: It’s an interesting angle. They view that, okay, you guys had some success with Japanese enterprises, and you kind of did things that people didn’t expect. They actually made AI deployments work in large Japanese companies. And I think if the technology is good enough for them, we may have a path to international deployment. So, the path I took was start in Japan first, take care of your domestic market before going international. And I think from Series B onwards, which we closed a few months ago, our company really has three go-to market strategy, three revenue lines. One is our expanding enterprise solutions business, which I just described, kind of like the Palantir style forward deployment engineer, making things work. The other one, which is also growing, especially with the new administration in Japan, Takeuchi administration is the government and defense business. Even since the founding of our company, our mission is to help Japan improve its defense infrastructure. I think this is really important, especially as someone who lives in Japan. That’s something I believe strongly about. And thirdly, is we’re finally producing our models and products aimed at the general audience in Japan and abroad. So, now after Series B, we’re going full on, everything, enterprise, government products.

Tim: So, let’s talk about some of your products, because, I mean, you guys are doing so many cool things, but there’s a couple I want to dig down on. So, one is Fugu, which if I understand right, just, it coordinates pools of different models and can learn from itself. Can you tell me a bit about what it does?

David: Our company, we’ve worked on this research direction called collective intelligence. We don’t believe that a single company or a single model would dominate all of AI.

Some companies believe that, but we don’t believe that.

Tim: Well, the ones building the big models kind of have to believe that.

David: We think that every model has its own uniqueness to it, its own biases, its own strengths and weaknesses. And being in Japan, we’re not China. We could use US models. And we could also use Chinese models as well. We can use our own models. We should try to build the technology to combine the strengths of as many models as possible.

Tim: So, when you’re doing that architecturally, or maybe philosophically, when you’re setting that up, are you trying to pick the right model for the right task? Or are you more of give a diverse set of models a similar problem and then work through the different answers?

David: That’s the philosophical direction. We want to develop the technology that would take advantage of the strengths of different models and kind of fuse them in creating the model that works in Japan. And there’s many approaches to do that. But one project we did that ended up to be our Fugu flagship product is to actually develop a foundation model. So, Fugu is a foundation model, is an LLM. But this LLM has the capability to call other LLMs when needed to actually achieve a superior result.

Tim: How does it know when it’s needed? Because LLMs are notoriously bad at this.

David: This is a good question. If you played around with all of these models, say you’re an AI engineer and you have to implement workflows for your customer. And at your helm, you can have Gemini, DeepSeek, Kimichu 2.5, Opus, and GPT. You kind of know which models are good at this, which models suck at that. Maybe Gemini is good at creative writing, GPT at reasoning, Cloud at code. And when your client or your boss gives you this project, you say, this part I use Gemini, that part I use Cloud and so on. This only works up to some extent. When we’re getting these models to do multi-step reasoning, there’s a lot more that we can get out of using different models. So, our approach, which is the same approach as any AI research, is whenever you don’t need to rely on human heuristics, but you can learn those heuristics from data, you should. So, our approach is actually to use reinforcement learning on a variety of different tasks to train an LLM how to do this orchestration and at which step you should do this orchestration. So, given the multi-step reasoning nature, maybe the first step you might assign to this kind of model. Maybe the second step, you might actually use another model. Maybe the third step you use the final model. So, all of this has to be learned from data using reinforcement learning. So this Fugu LLM itself is trained with reinforcement learning. It’s not a simple router. It doesn’t simply assign the task to one model and give the response to the user and call it a day. It’s a multi-step approach.

Tim: So, does the responses from those models feed back into its own training and eventually gets better and more self-reliant?

David: When we’re developing, we collect a lot of data internally on our own use cases. So, we can actually train this model and improve it by verifying our own users’ feedback. And this is very common for coding models as well, where you would have traces of coding data to improve on. So, in our case, when we wrote it out, we have a beta testing. When we have the users blessing, we can use data to improve the model. But at the same time, there’s opt-out as well. So, users can opt out of data collection. So I think it’s fair. We also use synthetic data to train the model better. There’s all sorts of things we can do to improve the quality of this model. But at the core, I think having the technology to be able to reliably orchestrate between numerous different foundation models is an important technology. I think this allows, if one model provider suddenly ceases to exist or suddenly decides to change their policy towards Japan, then you can continue to use a routing model.

Tim: Well, I do think this is one of the biggest risks for the big frontier model companies, is that there really is very little switching costs. The different LLMs have different strengths and weaknesses, different flavors, but it’s not that big. And the models you’re working with, are they the open models? Are they the big frontier models? All of the above?

David: Yeah, it’s all of the above. We have our own open source models. We have our own models. We can use the US frontier models as well. And when we roll this out, the user can actually choose which models they want the underlying to be. So, for example, if the user in Japan does not want to use DeepSeek, they can turn it off. Or if someone does not want to use Anthropic, they can turn that off. So they can choose which underlying models they want to use.

Tim: So, I’m curious, the frontier models are always six months ahead of the open source models, but that’s not a lot of time. So, that means anything the frontier models are doing today, the open source models will be able to do in six months. To me, it looks like that’s going to put tremendous pressure on any margins that Anthropic or OpenAI wants and would really favor a system that is a self-learning scaffolding type of a platform.

David: That’s what I think too. I mean, of course, the frontier players, they also know this.

Tim: Yeah, they’re smart people.

David: For example, if they have something like the Mythos, they can kind of emphasize its unique applications and only sell it at an extremely premium cost to some customers. But getting back to this type of product, I think strategically it makes sense because as you mentioned, open source models or the second tier models are only six months behind. And in reality, most of the tasks that people need to do might be done by a cheap model. Maybe 10% of the time you may need an Opus model and maybe 90% of the time, maybe like a Genma 4DB or something will do the job. So, having the orchestration technology to actually decide where to route the task, you can also optimize this for cost performance rather than pure performance. And we’re getting at the stage where cost of tokens per hour exceeds like multiple hourly wage of real employees.

Tim: It’s getting really crazy.

David: So, you’re going to enter into the world where the CFO of a company might actually want to have orchestrator or a router so that the task where you’re simply creating a landing page for the website, it should not go to a very expensive model. Whereas if you’re trying to prove Fermat’s Last Theorem, then it will go to the more expensive model. So, having this technology makes sense. Also, there will be in the future, I envision there will be not just performance-related reasons to send, but also data sovereignty.

Tim: Yeah. So, this kind of routing can also determine whether you want to be routing to a self-hosted model as well.

David: Yeah. So, these are all things in our vision that I think it’s important to be the company that’s very focused on building the R&D towards orchestration. This is like a very complicated advanced R&D topic.

Tim: No, but it’s fascinating. And I think it’s going to be absolutely essential for rolling out at enterprise scale.

David: Exactly. And the first product we roll out is a simple API. It’s open AI compatible, but many things we’re talking about, they’re needed in the future.

Tim: Oh, hell yeah. But so for something like this, would it require, I mean, it’s got to be a mixture of like hard-coded rules and the AI inference. Would that have to be trained up uniquely for each individual customer to fit their policies? Or do you think it’s like 80% just intelligent routing will be the same across everywhere? How do you think this will roll out?

David: I think there will be different needs for different orchestration models. The initial models that we’re rolling out are tailored for software engineering and for research tasks, but I can see demand for custom enterprise tailored orchestration. Whereas like this particular say mega bank or this particular strategy consultant or the government may need to have a unique orchestration strategy that’s trained on their own use case data.

Tim: Or I mean, but are some things just like personal information has to be handled on site?

David: Exactly. Those will also require more guardrails as well. For example, if you’re using an orchestration model in your local Mac and it’s using your own personal email, maybe your agent will simply have a hard guardrail where it simply cannot send any data to a cloud model. So, you’re going to have all of these features that will be required in the next few years. I think we’re going to enter into a world where we’re moving from mainframe to PC again. Everything right now is mainframe on a cloud.

Tim: The history of computing since the sixties has been in and out of the cloud.

David: Like back in the eighties, people kind of laugh at the hobbyists using their Apple twos, but now you’re at a stage where, we’re getting to the stage where you can run more advanced local models that will do 90% of what people need to do, but people still want that extra 10% performance, right?

Tim: And that’s where that routing comes in.

David: We’re going to have some sort of hybrid, I think.

Tim: That makes a lot of sense. I mean, just from a pure financial perspective, it almost has to go that way.

David: I mean, I think the RAM shortage issues is very challenging, but it will be addressed sooner than the data center capacity issues where you actually have to build more nuclear reactors, build actual transmission lines. Whereas like you can buy like a Mac mini and it’ll run at like 150 Watts in your home power line.

Tim: That’s exciting. Let’s talk about sovereign AI, Namazu in particular. So, this is an LLM adapted to the Japanese needs, right?

David: Yeah. So, this is kind of a weird story of why we did this. We’re also having our R&D effort on developing our own foundation models. We have numerous efforts in the pipeline. We’re even exploring training models from scratch end to end, but being the scrappy startup before devoting a huge amount of resources to train our models from scratch, we want to see how far we can get by working with open source models. They’re already available. People have already invested massive sums of money and compute to train these models. So, our early projects is actually to explore how to adapt open source models around the world to Japan’s needs.

Tim: So, what are Japan’s needs? What are they worried about?

David: We’ve actually did a systematic study of this. We looked at the most biased models actually in particular, say DeepSeek is produced in China. But of course we really respect the researchers and engineers at DeepSeek. But given the development in China, they also have to adhere to the regulations there. They have to say the right things. We looked at like LLAMA, it’s also quite biased towards the US narrative. We looked at GPT-OSS that was produced by OpenAI. We did a systematic study of how neutral these models are in Japanese culture and Japanese history. So, we’re not looking to make a model that’s biased towards Japan, to be clear. We want to have a model that is neutral. So, we have two different benchmarks. One is neutrality and we also have a factual accuracy. So, we measured these in the context of things that Japan cares about. We actually work with scholars and academic institutions to create our own benchmarks for neutrality and factual accuracy.

Tim: So, looking across from DeepSeek and LLAMA and GPT-OSS, it doesn’t look like they’re that different in terms of the amount of bias.

David: Actually, GPT-OSS is not so biased, but LLAMA is actually quite biased.

Tim: Surprising.

David: And DeepSeek, we want to choose the most biased model to do the study. We want to develop a post-training technology to solve this problem. This is actually a very challenging problem. There are four things you need to do to post-train open source models to adapt them for Japan. One is what I call a de-censor the model so that it will say things that maybe the government in their country does not want them to say. At least that’s the first step. The second step is de-biasing the model. So, it’s not like it’s not allowed to say it, but this is the narrative they’re trained to spread.

Tim: Sure, sure. I mean, just based on the training data, that’s going to be in there, right?

David: Exactly. So, you have to de-bias it, make it as neutral as possible. So, uncensoring the model so that the model might say things that its government might not allow it to say. De-biasing it, making sure that the narrative is neutral between the country of development and Japan. The third part is actually instilling Japanese specific knowledge and values into the model. So, as a foreigner coming to Japan, I spent many years learning about Japanese culture and so on. Same as an LLM. You have to have the data sets to actually instill Japanese culture, knowledge, and values.

Tim: And we end up learning the same way. It’s just by lots of training. No one will lay out the rules for you.

David: The fourth one is the most technically challenging. Because once you start post-training LLM, there’s a concept called catastrophic forgetting. So, it will forget what it already knows when you train it. So, these models are trained to do really well at their original tasks with their benchmarks like math, reasoning, coding capability. Once you start de-biasing it, uncensoring it, and teaching it what it means to be Japanese, it starts forgetting how to code. It starts forgetting how to reason. So, you actually have to continue training it on all of these difficult tasks to maintain or exceed its original performance. So, these four things are very important in order to post-train an open-source LLM into something that we believe is usable within Japan.

Tim: So, when you’re doing this, is it primarily giving it new training data? Is it primarily looking at fine-tuning responses or a combination of all?

David: It’s a combination, yeah. We spent a lot of work on our own data curation. We worked with many sources in Japan as well, like academic institutions, scholars, and so on. And I think once we made it work, the models are quite capable. To showcase this model, we released our first public service called Sakana Chat.

Tim: I was playing with that the other day.

David: Yeah, yeah. It’s kind of fun, right? I think it’s a good launch because last year we’re known as, okay, you guys are just doing enterprise work. How can I use Sakana AI technology? So, I wanted to really release something that anyone in Japan could use. Even though it might not make revenue, I think it’s important for us to have this commitment and attachment to Japan. So, we released Sakana Chat serving our Namazu LLMs that are tuned not just for Japanese values and culture, but it’s also very capable agentic-wise as well. It’s able to use many agentic tools. It’s able to use web search agents. It has reasoning capabilities. So, when we’re using Sakana Chat, it’s not a simple chatbot. You can ask it factual things like, what’s the weather tomorrow? Whether I should buy Japanese government bonds or not? What’s the price of the Nikkei? And so on and it’ll tell you. To add some mix on it, we even incorporated the Osaka mode so that you can kind of spoke in broken AI Kansai dialect. And people somehow really enjoyed that.

Tim: But I could see that just because it shows the attention to Japanese culture. It shows you’re going the extra mile to make it work. Are some of the models more retrainable than others, or are they all fairly similar?

David: From our experience, we chose the models because we’re able to post-train them. Not every single model is easy to post-train. In particular, some models that have gone through extensive post-training, especially small models, they might be more difficult to post-train. The GPT-OSS series, the Llamas and DeepSeek, they’re relatively easy to post-train. But we also chose DeepSeek because it’s the most biased model to showcase that we’re able to align the model. It’s also a controversial topic as well. I think many people in Japan and U.S., I think these are getting geopolitical.

Tim: I mean, that’s totally understandable. Not just Japan. Every culture is unique. Every society has their own history and their way of looking at things. And how important do you think this kind of sovereign AI or AI that’s tailored to a particular culture is going to be in the future?

David: Sovereign AI in a country like Japan, in my opinion, it’s not simply about training a Japanese LLM. At the end of the day, Gemini is pretty good at Japanese. Cloud is pretty good at Japanese. Even DeepSeek is pretty good at Japanese as well. You can always tune these models for the value. Sovereign AI is about a country that has the know-how and expertise to develop AI technology and even work with technologies in AI around the world so that it’s not reliant on individual models or technologies from a particular vendor or country to really own the sovereign of the technology. It’s like a supply chain.

Tim: It definitely makes sense from a supply chain and industrial policy point of view, but isn’t there a real need of like the Japanese government would want to be using an AI that is hosted in Japan, that is tuned for Japanese needs? And I imagine most major governments would feel the same way.

David: Yes, definitely. But what I’m saying is like in Japan, there are many efforts, including our companies, to actually develop LLMs even from scratch to adhere to the needs of Japanese governments or to preserve culture. But that alone does not solve the sovereign AI’s needs. If we only do that, then you might get into this like, maybe like the, if Japan only stuck with iMode phones and did not use iPhones and so on, like you’re only allowed to use technology developed in the country. It’s not going to work either. This technology is so global. There’s a global supply chain interdependence. Japan needs, for example, NVIDIA GPUs to do inference, so it cannot be totally self-reliant. China is trying to be self-reliant, but maybe their models are using cloud models outputs to train them using distillation. And in the US, they’re also using contents from around the world, including Japan’s to train their model. They’re probably using Japanese parts in their data centers. It’s a global supply chain. So, I think this concept of sovereign AI, whether we like it or not, AI is a complex global supply chain ecosystem. And you can’t fully have a self-contained AI ecosystem. It’s important to have the capability to develop and maintain and run models domestically. So, if other countries turn off their API access, your infrastructure continues to run. That’s very important.

Tim: Okay, that makes sense. So, AI sovereignty is much more about supply chain resiliency than about owning the stack.

David: Yeah, I don’t think any country can truly own the entire stack. Even the US and China cannot own the entire stack.

Tim: Then that gets back to why that routing is so incredibly important, because it gives you that resiliency.

David: This is, I think, precisely the reason why we really want to emphasize building out the R&D on the orchestration technology. Even philosophically, humans are a collective intelligence. We’re not a big brain. So, we’re like billions of little brains working with each other and communicating. The entire humanity is one big collective intelligence. The same is going to be like that for AI.

Tim: I love that. This post-training, this Namazu post-training to set tone and reduce bias, is that something that only makes sense on a national cultural level, or is that something that could be tuned up for specific enterprise customers to fit their feeling and needs and biases?

David: That’s a good question. I’m sure that cultural alignment is not restricted to individual countries. There’s different levels of culture. Maybe your individual family has its own culture, a company has its own culture, or even people in Tokyo behave very differently, or even people in Kyoto behave very differently than Osaka. So, I feel there’s definitely the possibility of this happening. Weirdly enough, I think in Japan, Japanese companies used to have company songs. It’ll be interesting to have a unique LLM tailored for a Japanese traditional company and train on the history of what the previous Satchos have said.

Tim: If it can be made economically, I can completely see that happening.

David: I can see that happening. I’m not sure what the consequences will be. Would you have large Japanese companies have their own LLMs that’s trained on their own corporate culture, and all the employees will learn to adapt to that corporate culture?

Tim: Let’s talk about startups in Japan again, because I love the fact that you guys broke all the rules. You raised money overseas as a KK, you raised significant capital here in Japan. What advice do you have for founders in Japan now, both foreign founders and globally-minded Japanese founders? Is that the right model to follow?

David: We broke many rules. We didn’t break all the rules. I think the most important thing is to think from first principles. Think whether every action actually makes sense or not, rather than blindly taking anyone’s advice, including my own. Being able to think rationally from first principles is important. My advice to founders is not to easily take other people’s advice actually because a lot of the advice that people give that you should do everything that Silicon Valley companies are doing. I feel in Japan, especially for Japanese founders, there’s an idolization of Silicon Valley.

Tim: Do you think that’s still the case? I mean, it was definitely true five, ten years ago, but I find there’s a more Japanese self-confidence.

David: I hope so, yeah. I think the tide might be turning, because in the course of me starting this company, I do think that many Japanese founders kind of idolize the way the Valley works. They think it’s so great, but I live there, man. I don’t like Silicon Valley. I like Japan.

Tim: That’s why we’re here.

David: That’s why we’re here, right? So, I do think try to understand what you’re trying to achieve, what the goals you have in mind. And if it does make sense for a founder in Japan to move to the Bay Area, then that’s fine as well. Vice versa, I think there’s many foreigners now coming to Japan to start companies. Many of them are well-intentioned, have a clear business strategy, but many of them might be like, oh, Japan’s kind of cool. I like anime, and I went to Japan for a vacation many times. Why don’t I go there and visit Shibuya support center? You have to be prepared to start companies here.

Tim: I get contacts from overseas at least a couple times a week. People saying, hey, I want to move to Japan to start a company. What should I do? I’m like I don’t know, man. But yeah, starting a company, you need to be close to your customers and understand the problem you’re solving. But it makes sense what you’re saying. So, listen to the advice, but just make your own decisions.

David: Exactly. I think we really have to think for ourselves. And if we do things that everyone else is doing in startups or even in trading and investing, it’s less likely that it’ll be very successful.

Tim: Yeah, absolutely. Especially in startups. There’s so many people trying to solve the same problems. You’ve got to be doing your own thing.

David: You got to be doing something different, but maybe not too different.

Tim: Well, listen, David, before I let you go, I want to ask you what I call my magic wand question. And that is, if I gave you a magic wand, and I told you that you could change one thing about Japan, anything at all, the education system, the appetite towards risk, how rapidly people adapt new technologies, anything at all to make it better for startups and innovation in Japan, what would you change?

David: This is a good question. There’s many typical answers like, let’s teach everyone English at a younger age or something.

Tim: That’s a common one.

David: That might actually make people less innovative, in my experience. Japan will cease to be Japan, in a sense. So, we have to be very careful if we have this power. You asked about increasing the amount of innovation in Japan. There’s all this talk in Japan about should we teach kids more English or should we introduce STEM programs, give everyone laptops, or encourage more daycare centers in Japan so people can work. All of those are great. But to really move the needle in terms of having this innovation, like the era in Japan when you have companies like Sony or Panasonic come out, there’s a few factors. One is the nation needs to have a collective sense of hope. It needs to feel hopeful. It could be poor. Everyone can be dirt poor, but they have hope. And that hope leads to things being done. So, what creates hope in Japan? Is it more money? Is it less taxes? Or is it more benefits? Probably not. I think it’s stories and narratives. So, if I have a magic wand, I think if the population in Japan is more optimistic about the future in this country, suddenly they’re persuaded, brainwashed to be more optimistic. They believe things are going to happen.

Tim: And the thing is, there’s plenty of reason to be optimistic here. There really is. But yeah, you’re right. There is this unfortunate cultural tendency to, when faced with good news, saying, well, it’s not really that good, right?

David: Exactly. People should be optimistic. With optimism comes change. And change leads to more optimism. So, if I have a magic wand, make everyone in the country be very optimistic.

Tim: And that’ll lead to even more optimism.

David: Exactly.

Tim: Fantastic. Well, listen, David, thank you so much for sitting down with me. I really appreciate it.

Tim: I appreciate it, Tim.

Outtro

And we’re back.

Sakana AI’s orchestration strategy of owning the router, of being the router, is a good one. The low-cost models are only six months behind and gaining quickly. And with inference costs certain to rise in the future, in the long run, this kind of orchestration is the only financially rational approach.

The other thing I appreciated from our conversation was just how fast Japan is changing, both in terms of AI adoption and in terms of startup innovation. In terms of AI adoption, it makes sense that Japanese enterprises talk about their AI experimentation a lot less than their American counterparts do.

The U.S. market rewards bold risk-taking, and the Japanese market rewards results. An American CEO announcing that he is firing 20% of his staff and replacing them with new, unproven AI technology is hailed as a visionary and rewarded by the market with higher stock valuations.

A Japanese CEO doing the same thing would be rebuked for recklessly betting the whole company on unproven technology and severely punished by the markets. Japanese companies want to make sure something actually works before betting the firm on it.

Now, it’s a fair debate which approach is better, but if you want to operate in both markets, you need to understand this difference.

The fact that Sakana was able to raise global capital as a KK and then raise around $200 million from Japanese VCs is incredibly encouraging. For years, the common wisdom was not to try to raise foreign capital as a KK and to look to the U.S. if you’re planning on raising tens of millions of dollars. Well, Sakana AI has shown us that neither of those rules are as solid as we thought. They received very little pushback as a KK, and they were able to raise around 200 million from Japanese investors.

Now, granted, Sakana is a hot startup in a very hot sector, so things might not be quite as smooth for other founders, but still, global investors understood the KK structure. Sakana was not their first KK investment.

Also, Sakana still had U.S. VCs lead their rounds and set the valuations, but there was no shortage of Japanese capital willing to invest. And that, that is encouraging for the whole startup ecosystem here.

Individual investment trends and amounts ebb and flow, but the landscape is steadily improving for Japanese startups and globally-minded Japanese founders. Building startups is never without challenges, but as David and I discussed, there are just so many reasons to feel optimistic about startups in Japan.