Sakana AI’s evolutionary approach to Japanese AI sovereignty

Scaling through collaboration

Sakana AI provides the most effective path for Japan to build a sovereign AI ecosystem without trying to outspend US hyperscalers on massive GPU fleets. The Tokyo-based startup, founded in July 2023 by Google alumni David Ha and Llion Jones, secured $135 million in a Series B round in November 2025. This follows a $100 million Series A in June 2024 and a $30 million seed round in January 2024. The startup’s investment history includes a $30 million seed round in January 2024, a $100 million Series A in June 2024, and a $135 million Series B in November 2025, which brings its valuation to over $2.65 billion. Investors include Salesforce Ventures, Google, Mitsubishi Electric, Citi, Macquarie, and In-Q-Tel. The company focuses on many smaller models working together rather than one large foundational model. This strategy fits Japan because the country relies on energy imports and requires efficient AI solutions. The company targets markets like Japanese language AI where specialized models work better for narrow workflows. You know that frontier labs prioritize bigger datasets and longer training runs, but Sakana AI uses existing open models and cheaper search loops. While the team consists of 20 people, they aim to establish Tokyo as a center for AI research. Co-founder Ren Ito previously ran Mercari Europe, and the founding team includes people Ha worked with at Google and Goldman Sachs. Llion Jones, a co-author of the seminal paper that helped catalyze the current AI boom, works alongside Ha to build the company. The Series A round included investors like KDDI, NEC, ITOCHU, Fujitsu, MUFG, Mizuho, Sumitomo Mitsui, Nomura, Dai-Ichi Life, and SBI Group. These funds allow the company to develop technology that reduces Japan’s dependence on foreign AI.

Efficiency through evolutionary merging

The launch of the RSI Lab this year proves recursive self-improvement works on budgets closer to national research scales. Sakana AI researchers developed the Darwin Godel Machine, which doubled its baseline software-engineering performance on SWE-bench. Their paper-writing agent, The AI Scientist, received recognition in a March 26, 2026, publication in Nature. Other results include ShinkaEvolve, which solved complex optimization problems using only 150 samples, and ALE-Agent, which took first place out of 804 participants in the AtCoder Heuristic Contest 058. The company also produced EvoLLM-JP, a 7B model that beat prior 70B class Japanese models on benchmark suites. This method uses evolutionary techniques to combine different open-source models. They start with a pool of already trained models, splice together their weights and layers, score the children on a concrete task, and keep iterating until a stronger specialist emerges. Successful models from each generation become parents for the next. This process allows the company to discover models with unexpected abilities. The method automatically discovers ways to merge models from diverse domains, including non-English language, math, vision, and language. The company also released EvoVLM-JP and EvoSDXL-JP on platforms like Hugging Face and GitHub. Sakana AI is not trying to outspend OpenAI or Anthropic on massive pretraining runs. The real bet is that better search over existing models can beat brute force training on cost adjusted performance. This approach works with existing open models to serve enterprise customers like banks that want a tuned model for a specific workflow. Can Sakana AI maintain this lead as more companies attempt to replicate its evolutionary merging recipes?

Infrastructure and national computing

The Japanese government supports this development through the GENIAC grant, which provides Sakana AI access to GPU clusters for training foundation models. This access complements the work done on the Fugaku supercomputer, a joint project between RIKEN and Fujitsu. Fugaku, named after Mount Fuji, debuted in 2020 and was the first ARM architecture-based computer to reach the top spot in the TOP500. The system uses 158,976 Fujitsu A64FX CPUs. Fugaku achieved 2.0 exaflops on the HPL-AI benchmark, which exceeded the exascale threshold. By June 2026, Fugaku occupied the 9th place on the TOP500 list. After a hardware upgrade in November 2020, the system’s performance on the mixed precision HPC-AI benchmark rose to 2.0 exaflops, which was 42% higher than the 1.4 exaflops recorded six months prior.

FugakuNEXT, the next flagship supercomputer, targets deployment around 2030. It aims to combine AI and HPC together.