David Ha's bet is simple: the next serious AI race may be won by systems that know how to use many models well, not by one giant model trying to answer everything alone.
The biggest AI labs have spent the past few years teaching you to watch one number: model size, benchmark rank, compute budget, valuation. Sakana AI is asking a different question. What if the smarter move is not to build the largest model in the room, but to build the conductor that knows which model should play which part?
That is the point David Ha, Sakana's co-founder and CEO, made in his Disrupting Japan conversation with Tim Romero. It is also the argument sitting behind Fugu, the Tokyo company's new model orchestration system. Sakana's technical report, submitted to arXiv on June 19, 2026 and revised on June 23, describes Fugu as a family of language models trained to coordinate a team of other models. Fugu handles faster, lower-latency work. Fugu-Ultra is built for harder tasks where quality matters more than speed.
This is current for a simple reason: Fugu landed last week, and it goes directly at the industry habit of treating scale as destiny. The report says the system can route, delegate, verify, and synthesize across a pool of frontier workers, including coding and reasoning agents. You don't have to believe every benchmark claim to see the strategic move. Sakana is not trying to outspend OpenAI, Anthropic, or Google on one enormous training run. It is trying to make their category of advantage less absolute.
That suits Sakana's history. The company was founded in Tokyo in 2023 by Ha, Llion Jones, and Ren Ito. Jones is one of the authors of the 2017 "Attention Is All You Need" paper, which introduced the transformer architecture behind modern language models. Reuters reported in 2024 that Sakana raised about $200 million from investors including Nvidia and Japan's three megabanks, Mitsubishi UFJ, SMBC, and Mizuho. That investor list is not decoration. If you're selling AI into Japanese institutions, having the country's megabanks around the table changes the conversation.
Ha's sharper claim is about enterprise Japan. The easy foreign story is that Japanese companies are slow, cautious, and trapped in pilot programs. That story is too lazy. Japan has plenty of old systems and slow procurement, but it also has banks, telecoms, manufacturers, and trading houses with real reasons to automate work carefully rather than theatrically. A model orchestration layer fits that culture better than a single black-box chatbot dropped into every workflow.
Fugu is interesting because it treats AI work as a management problem. A user's task comes in, the system decides which worker model or models should handle it, and the final answer is assembled from that process. The report frames this as "collective intelligence," which is Sakana's long-running theme. Even the company's name points there: sakana means fish in Japanese, and the image is a school of fish moving as one.
Here's the thing: this is not just a nice metaphor. In software engineering, finance, legal review, scientific research, and operations, most useful work is not a single clean question with a single clean answer. It is a chain of smaller judgments. One model may be better at reading code, another at planning the fix, another at checking whether the output actually satisfies the task. A system that can assign those jobs well has a practical advantage, even if none of its component models is the world's most famous one.
Sakana has been circling this idea for a while. Its 2024 AI Scientist paper, co-authored by Ha and others, described a system that could generate research ideas, write code, run experiments, produce figures, and draft a paper for less than $15 per paper. That was an audacious claim, and it deserved scrutiny. A 2025 independent evaluation on arXiv found real weaknesses, including failed experiments, thin literature reviews, and hallucinated numerical results. Good. You want that pressure on claims this large.
Still, the direction is clear. Sakana is not presenting AI progress as a straight climb toward one machine that knows everything. It is arguing for systems that improve by combining, testing, and reusing pieces. That is a very different industrial model from the frontier-lab race, and it may be more useful to companies that need controllable systems before they need a cosmic demo.
For founders, the lesson is not "go build a model orchestrator." Don't bother copying the surface. The useful lesson is that you can attack a giant's strength from the side. If the market assumes bigger always wins, build around coordination, cost, locality, compliance, or workflow fit. Sakana's bet is that the strongest model is not always the most valuable thing in the system.
That is why Ha's argument is worth taking seriously. If Fugu works in production as well as it reads in the report, Sakana has found a way to make the model race less monolithic. The future of AI may still belong to the giants in many areas. But for the companies actually trying to put AI inside banks, labs, and offices, the winner may be the system that knows when to call in help.
Also read: Karnataka's Bidadi AI City is already a land test before it is a tech test, Firmus puts its ASX float behind a larger Nvidia bet, and Pimco is turning AI data centers into a private debt power play
