Sakana AI, the AI star unicorn pitched by Jen-Hsun Huang, is valued at over $18 billion after more than 2 years of existence

According to The Information, citing people familiar with the matter, OpenAI's Japanese bidding partnerSakana AIIt was revealed that it will be raised at a valuation of $2.5 billion (roughly Rs. 17.8 billion)US$100 million (equivalent to about 700 million yuan), up 66% from a valuation of $1.5 billion (equivalent to about 10.7 billion yuan) in September last year.

Sakana AI, which is in talks with U.S. and Japanese investors and will be valued at $2.6 billion (roughly Rs. 18.5 billion) after raising this $100 million in new funding, plans to use the new money to expand its team of employees in engineering, sales and distribution, where it now employs about 70 people, people familiar with the matter said.

Sakana AI was founded in July 2023, and its founding team is a big deal.David Ha is a former Google Brain Senior Scientist who also led the research arm of image generation modeling unicorn Stability AI, and the other founder, Llion Jones, is one of the authors of the Transformer paper and a former Google AI researcher.. In August, David Ha also made TIME's list of the top 100 AI people of 2025.

▲Sakana AI founders Llion Jones (left), David Ha (right)

The startup is taking a different route than those chosen by OpenAI, Anthropic, and others.It wants to ditch the Transformer architecture and use natural evolution as inspiration to reduce model computation costs and improve model performance. In terms of product progress, the company has released three Japanese language models, Karamaru, an Edo kozo chatbot, and ShinkaEvolve, an open source framework.

At present, Sakana AI has taken a total of nearly 230 million U.S. dollars (equivalent to about 1.6 billion yuan) financing. 2024 January, Sakana AI obtained 4.5 billion yen (equivalent to about 200 million yuan) financing, Lux Capital, Khosla Ventures, the Japanese telecom company NTT, and the Sony Group to participate in the investment; last September, Sakana AI completed 214 million U.S. dollars (equivalent to about 1.5 billion yuan) A round of financing. Sakana AI completed 214 million U.S. dollars (equivalent to about 1.5 billion yuan) A round of financing, NVIDIA participation in the investment to become a major shareholder, so that its valuation of one year after its establishment reached about 1.5 billion U.S. dollars (equivalent to about 10.7 billion yuan), quickly jumped to a unicorn.

Additionally, the company has received a grant from the Japanese government to purchase computing resources for AI training, according to The Information.

Sakana AI means fish in Japanese, and Ha previously told The Information that he intends to build a “Japanese DeepMind.”

Ha revealed that its algorithms can help companies automate the creation of large models for processing language or images, implemented by mimicking the laws of natural selection when combining, testing and discarding data, thus reducing the amount of manual input required. That's because the amount of power or resources required to train big models consumes a huge amount of power, and he hopes to reduce that time to days or weeks.

Specifically, Sakana AI uses an algorithm that uses three open source primitive AI models as “parents” and “raises” over a hundred ”children” in different ways, and then The best-performing model is selected, and then the next generation is generated, and then the process is repeated to create hundreds of generations of models and select the best.

In March of this year, SakanaAI releasedThree Japanese models, respectively, the large language model EvoLLM-JP, the visual language model EvoVLM-JP, and the image generation model EvoSDXL-JP.

After a month, Sakana AI is releasedKaramaru, an Edo-style chatbot., a chatbot that can answer users with an Edo period worldview and ancient documents of the time, was trained on data from a dataset containing about 25 million characters that the company built from Edo period books.

In September of this year, Sakana AI open-sourced theShinkaEvolve, the framework combines large models with an algorithm that helps it come up with solutions for a variety of potential problems, narrowed down to the most similar ones. The company claims that this approach is more efficient and better at coming up with creative solutions than the Transformer architecture.

▲Overview of the ShinkaEvolve Framework

To date, Sakana AI has signed business agreements with several major Japanese companies.

In May of this year, it announced a multi-year partnership with Mitsubishi UFJ Financial Group, Inc.Development of customized AI models for Japan's largest bank. After this deal was finalized, Ha told the foreign press thatThis will make Sakana AI profitable within a year; Earlier this month, Sakana AI announced a similar partnership with Daiwa Securities Group, one of Japan's largest brokerage firms.

Ha mentioned that every company in Japan that wants to have their own base model can develop their own model in partnership with Sakana AI or other companies like Sakana AI, and that there is room for Japan to be an innovator.

In addition, in Japan, Sakana AI faces competition from many large American model developers.

In February, OpenAI announced a partnership with SoftBank in which SoftBank agreed to spend $3 billion (roughly Rs. 21.4 billion) per year on OpenAI's technology for itself and its subsidiaries, and SoftBank and OpenAI will also set up a joint venture specializing in the sale of OpenAI's suite of tools to Japanese companies; Anthropic, in August of this year, appointed a Japan country head and is set to open an office in Tokyo, Japan; and in August this year, Canadian big modeling startup Cohere hired a Japanese country manager.

Conclusion: Sakana AI bets on big local Japanese models

Sakana AI is similar to French macromodeling startup Mistral AI and Indian macromodeling startup Krutrim, in that it hopes to reduce its dependence on other AI technologies by developing its own macromodels and giving the models a better grasp of their region's language and culture.Ha argues that it makes sense for Japanese companies to have the option of owning their own base model, even if it's not a ultra-high-performance model.

However, compared to OpenAI, Anthropic, Google and other big model core players, Sakana AI such startups are relatively weak in terms of capital and talent strength, and it needs to find a breakthrough path through innovations such as algorithms.