OpenAI’s Japanese Rival Gets $1 Billion Valuation From Silicon Valley Investors
The Takeaway
• NEA, Lux and Khosla are co-leading a new round for Sakana AI
• The Tokyo-based startup wants to be the OpenAI of Japan
• It was valued at $200 million in January
In their hunt for artificial intelligence investments, U.S. venture capitalists are looking past Silicon Valley at foreign AI startups, betting that locally developed large-language models will outperform imports from the U.S.
The latest to win their affections is Sakana AI, a Tokyo-based large-language model developer. Founded last year by alumni of Google’s DeepMind, Sakana is raising around $100 million in a new financing co-led by New Enterprise Associates and existing investors Lux Capital and Khosla Ventures, according to three people with direct knowledge of the deal. The round values the company, which does not yet generate revenue, at approximately $1 billion, the people said.
The funding has not closed and the terms could change. Sakana cofounders David Ha and Llion Jones did not respond to a request for comment.
Sakana’s new funding comes on the heels of a breakthrough announced in March, when it released research showing how to develop large-language models for less money, as well as an LLM that solves math problems in Japanese and models that can generate and understand images and text in Japanese. It has released some of its models as open-source software, but not a product that consumers can access similar to OpenAI’s ChatGPT. It’s still unclear how the company will make money.
U.S. AI developers are already establishing a presence in Japan. In April, OpenAI announced that it opened up an office in Tokyo as a way to expand Asia operations. It also hired Tadao Nagasaki as president of OpenAI Japan and gave local businesses access to a model that is optimized for the Japanese language.
The new funding for Sakana comes amid a rash of investment interest for AI startups developing models focused on specific geographies.
Earlier this month, the Paris-based AI startup Mistral, which has styled itself as the “OpenAI of Europe,” raised $650 million at a $6 billion valuation from U.S. investors including Lightspeed Venture Partners and General Catalyst, for instance. Nikkei Asia earlier reported Sakana was in talks with the U.S. VC investors for a $1 billion round.
‘DeepMind of Japan’
Sakana was founded last year by Ha and Jones, who co-authored the Google research paper that outlined the transformer model architecture that underlies today’s most popular LLMs like OpenAI’s GPT and Anthropic’s Claude. Ha previously told The Information that he intended to build a “DeepMind of Japan.”
“If we had started Sakana AI in the Bay Area, it would have been a strategic blunder, because we would have looked more like everyone else and it would have been very hard to differentiate,” Ha said.
In January, Sakana raised a $30 million seed round led by Lux Capital and included a mix of Japanese investors, including Sony Group and NTT Group as well as Japan-focused venture firms such as Geodesic. The round valued the company at $200 million, according to two of the people with direct knowledge.
Sakana also has received a boost from the Japanese government, which chose Sakana as one of a handful of companies and labs to receive grants to gain access to a large amount of computing power.
The startup is also part of a growing number of upstarts looking to challenge the traditional transformer architecture. Transformers, which are particularly capital and compute-intensive to train and run, are notorious for producing errors known as hallucinations, which make them difficult to use in high-stakes business settings. Recent research has also suggested that as transformer models get better, they will require increasingly more training data—and, therefore, capital—to eke extra performance out of them.
Instead, Sakana is building foundation models inspired by concepts found in nature. (Its name is derived from a Japanese word “meant to invoke the idea of a school of fish coming together and forming a coherent entity from simple rules,” a reference to the company’s unique approach to building “nature-inspired intelligence.”)
For example, the company has experimented with “model merging,” in which researchers combine two or more models trained for different purposes to create a single model that exhibits the strengths of the original model without needing additional training—a similar concept to evolution in nature.