Japan's Sakana AI Bets on a Swarm of Small Models Instead of One Giant AI
- Japanese startup Sakana AI has unveiled a system called Sakana Fugu.
- The design lets multiple small models divide the work instead of one giant model.
- US and Chinese rivals compete on capital and compute for ever-larger models.
- Lower compute and power requirements are the main selling point.
Most AI news from the US and China is about scale: bigger models, more chips. Japan's Sakana AI is asking a different question with Sakana Fugu — can a swarm of small models, each doing part of the job, replace one giant one?
The choice reflects Japan's position. Training frontier models takes capital, power and advanced silicon, none of which Japan leads in. What Japan does have is a large base of manufacturers and regional firms that want AI but cannot send data outside the company. Small models can run on owned servers or at the edge, which keeps data in-house and power bills manageable.
Judging it should not turn on parameter counts. The questions are how much answer quality is lost when work is split, what the orchestration overhead costs, and how much staffing operations require.
For readers in Taiwan there are two takeaways. Small and mid-sized manufacturers here face the same constraints, so the approach is worth tracking. For investors, few listed Japanese "AI names" own model technology; most sell integration and hardware.
Watch for named deployments and third-party benchmarks rather than launch events.