David Ha is co-founder and CEO of Sakana AI, based in Tokyo.

The world is obsessed with frontier AI models. Billions are poured into training ever-larger monolithic systems, on the assumption that scale alone guarantees dominance. But this strategy is becoming economically irrational. Open-source alternatives now trail by only a few months, and the inference costs of many frontier models exceed the hourly wages of the employees they are meant to augment. The "one model to rule them all" approach is a path to diminishing returns.

Sakana AI, the Tokyo-based lab I co-founded in 2023, is not building the Japanese version of OpenAI. We are building a different architecture for intelligence, rooted in constraints, collective action and cultural sovereignty.

The real strategic value is moving up the stack, from the raw model weights to the intelligence that coordinates them. The winners of the next phase of AI will not be those who build the largest monolith, but those who own the orchestration.

The school of fish Some of David Ha's early conceptual logo designs for Sakana AI. He says that after many attempts, he eventually settled on the right one.

Sakana means fish in Japanese, and the name reflects our central bet: collective intelligence. Just as a school of fish behaves in ways no single fish can, we believe no single company or model will dominate AI. Every model carries its own biases, strengths, and blind spots. The future belongs to the model orchestrator: a system that learns how to coordinate a pool of models to solve complex, multistep problems. Intelligence is not just an individual fish.

It's the school. Our orchestration model, Fugu, is itself a foundation model. Unlike a hard-coded router that follows static human rules, Fugu is trained end-to-end through reinforcement learning on thousands of multistep reasoning chains. It learns which underlying model to invoke at each step, how to synthesize conflicting outputs, how to balance cost, performance and compliance, and when to route sensitive data to a self-hosted model rather than a cloud API.

The difference between an individual foundation model and this learned orchestration layer is the difference between a mainframe and a personal computer.

In a world moving toward hybrid local and cloud inference, the orchestrator becomes the operating system. Some tasks require the brute force of a frontier model. Others can be handled by a small, efficient model running on a local device at 150 watts. The intelligence lies in knowing which to use, when and why. And if one model goes offline or changes its policies, the system simply routes around the damage.

Sovereignty is a supply chain, not a wall The debate around AI sovereignty is often misunderstood. It is not about building a perfectly isolated, purely domestic model. No country, not even the U.S. or China, owns the entire stack. Japan needs Nvidia graphics processing units (GPUs); American models are trained on global data. True sovereignty is supply-chain resilience: the know-how to develop, adapt and maintain AI domestically, combined with the ability to orchestrate across global resources without depending on any single vendor. This is why post-training matters.

Our Namazu models, which power Sakana Chat and Sakana Translate, are not trained from scratch. They start from open-weight base models and are then post-trained on our own datasets for Japanese language, culture and business contexts, with additional tuning to reduce unnecessary refusals and bias.

Namazu outperforms its base models on benchmarks for neutrality in politically sensitive Japanese contexts, on comprehensive Japanese-language evaluations, and on translation. Cultural alignment does not require sacrificing capability.

But the ultimate insurance policy is orchestration. If one provider shuts off API access, the routing model fails over to others. Sovereignty is continuity through diversity. In an era of geopolitical fragmentation, this is not an academic concern. It is an existential requirement for any nation that wishes to control its own digital infrastructure.

Building under constraint Japan does not have the capital or energy to train trillion-parameter models from scratch every quarter. That constraint shapes how we do research and how we go to market.

In research, our focus is recursive self-improvement, or RSI: AI systems that help design and train the next generation of AI systems. RSI is often framed as science fiction, an AI rewriting its own code toward god-like intelligence. For us, it is an economic imperative. Our AI Scientist framework, published in Nature, demonstrates that systems of AI agents can already generate scientific ideas, write code, run experiments and compile findings.

Applied to model development itself, we estimate this could reduce pre-training costs by one to two orders of magnitude. The goal is not to remove humans from the loop, but to use human judgment as a sanity check before committing millions of dollars to compute. Alignment, error-checking and efficiency go hand in hand. This is how a smaller player out-innovates a better-funded rival.

In the market, many in Silicon Valley go consumer-first. We chose the opposite path. Our first deployments were with Japan's most conservative institutions: megabanks, securities companies and government agencies. If you can deploy AI in a Japanese megabank, you can deploy it anywhere. Our partnership with MUFG has produced the AI Lending Expert, a production system that generates detailed credit memos that a human loan officer can affirm or override.

A document agent for M&A cuts pitch-deck production from weeks to hours, but the banker retains ownership. This human-in-the-loop design is not a technological limitation; it is a cultural feature that ensures accountability.

These deep co-development partnerships function as our laboratory. Japan's enterprise culture, often criticized as slow, is a competitive advantage. These companies demand production-grade reliability. They do not celebrate laying off staff for unproven technology; they reward results. That discipline forces us to build software that actually works.

The need for optimism There is one final, nontechnical ingredient to Japan's success: hope. If I had a magic wand, I would make everyone in the country optimistic. With optimism comes change. And change leads to more optimism.

I have lived in Japan for years, and I have seen its capacity for rapid, collective mobilization when people believe change is possible. The post-war economic miracle and the rise of the likes of Sony and Panasonic were not driven by abundant capital alone, but by a collective sense of purpose. Japan has a history of turning scarcity into an advantage.

The conditions are ripe. Japan is transitioning from deflation to inflation. Its banks hold vast capital that must be deployed. Its enterprises are adopting AI in production faster than their global peers realize. What is needed now is a narrative of optimism. Moving from monolithic models to collective systems, from imported technology to sovereign infrastructure, requires a society willing to believe that a different architecture is possible, and that Japan is the right place to build it.

Japan has a history of turning scarcity into an advantage. Today, the constraints of compute, energy and capital are forcing us to build smarter, more agile, and vastly more efficient systems. The next decade of AI will not be won by those with the most GPUs. It will be won by those with the most intelligent architecture: systems that route, adapt, improve themselves and respect the cultural and operational realities of the societies they serve.