The founders of Sakana AI named the company after the Japanese word for fish, and they have described the work as breeding models rather than building them: take open systems that already exist, merge them, mutate them, keep whatever performs best. It is a bet that a small team with a clever method can stand beside laboratories spending billions on computing power.
That bet is now headed toward a stock exchange. Sakana has hired Goldman Sachs to advise on an initial public offering, according to people familiar with the matter, with a listing possible as early as the autumn of 2027 in Tokyo or New York. The company has not chosen a venue, and the people cautioned that the timetable could change.
A listing two years out would be an unusual thing to announce for a startup founded in 2023, with a staff measured in the dozens and a revenue base that has not been disclosed. It is also a measure of how badly Japan wants a homegrown artificial-intelligence champion, and how thin the field of candidates remains.
Sakana was started by David Ha and Llion Jones, both former Google researchers. Mr. Jones is a co-author of the 2017 paper that introduced the transformer, the architecture behind most of the large language models now in commercial use. The pair argued that the industry’s race toward ever-larger training runs was wasteful, and that intelligence could be assembled from existing parts through evolutionary search, the way nature optimizes without a plan.
The approach has produced models that run cheaply and papers that other researchers cite. It has not produced the kind of revenue that justifies a frontier-lab valuation. That gap is the central question any offering would have to answer, and Goldman’s first job will be to find investors willing to price method rather than scale.
Japan’s policy apparatus has reasons to want the answer to be yes. The government has spent on domestic computing capacity and framed AI as an industrial priority, while its most prominent AI startup, Preferred Networks, has stayed private for more than a decade. SoftBank has assembled pieces of an AI empire mostly by investing abroad. For a country with Japan’s engineering base and its shrinking workforce, the absence of a listed model developer looks like a gap in the capital markets as much as in the technology.
Tokyo has been working on that gap. The Tokyo Stock Exchange has pushed companies to lift their valuations and has tried to make its growth segment more attractive to young firms. Still, the deepest pools of capital for unprofitable technology companies sit in New York, and a Japanese AI startup that lists at home would be relying on domestic institutions and retail investors who have historically been cautious about companies without earnings.
There is a strategic argument for listing in Tokyo anyway. A domestic listing keeps ownership local, keeps the company inside the government’s industrial policy orbit and gives Japanese savers a way to own a piece of the sector. It also exposes the shares to a market that trades technology at lower multiples than Nasdaq does, which is why some founders in Japan have looked west.
Competitors will be watching the process for signals about valuation. Companies building applications on top of other companies’ models have found public markets receptive when revenue grows quickly. Companies that train their own models have found investors asking harder questions about capital intensity and defensibility. Sakana sits in between: it does not build foundation models from scratch, and it does not simply resell someone else’s.
The competitive position carries its own risk. Merging open-weight models is cheaper than pretraining, which means rivals can copy the technique. The advantage, if there is one, lies in the accumulated engineering knowledge and in the data and evaluation work that surrounds it, none of which is easy for an investor to verify from a prospectus.
Japan’s regulators have taken a light approach to AI so far, preferring guidance to statute, and that has been part of the pitch to founders. A listed company would face the disclosure obligations of any public issuer, including quarterly reporting on a business whose products change every few months.
The timeline invites skepticism for a simpler reason. Two years is several model generations in this field. A company that lists on the strength of one method may find that the method has been superseded by the time the shares trade, and the founders who argued against scale will be judged by a market that rewards it.
None of that has stopped the preparation. Hiring Goldman this early suggests the company wants the option rather than the obligation, and investment banks are typically retained long before a filing is drafted. What remains unsettled is whether the first serious test of Japanese AI in public markets will be a company that builds models, one that merges them, or one that sells something else entirely.
