Sakana AI bets small nature-inspired models beat brute-force scaling; raises $100M+

Tokyo AI lab founded by ex-Google researchers builds small, cheap, evolutionary models and raised $100M+ with NVIDIA, then ¥20B more in 2025.

Sakana AI

它在赌什么That small, efficient, evolving models can rival monolithic giants — making top-tier AI cheap, sustainable and Japanese-built without megascale compute.在扩

做的是什么生意

A Tokyo AI lab building small, low-cost generative models (image, language, coding) whose algorithms create models by mimicking evolution and natural selection, instead of scaling one giant model.

启动资金：Seed $30M in Jan 2024 led by Lux Capital with Khosla Ventures, backed by Sony, NTT, KDDI and Google's Jeff Dean; Series A over $100M in Sep 2024 led by NEA/Khosla/Lux with NVIDIA joining; about ¥52B (~$347M) cumulative after the 2025 Series B.

起因

Sakana AI was founded in Tokyo in July 2023 by David Ha (ex-Google Brain), Llion Jones (co-author of the paper behind the transformer architecture used by ChatGPT) and Ren Ito (ex-Stability AI COO and ex-Mercari Europe CEO). The name comes from small fish that coordinate to hunt — the thesis that many small models working together can outperform one giant one.

经过

In Jan 2024 it raised a $30M seed led by Lux Capital and Khosla Ventures, with Sony, NTT, KDDI, Google's Jeff Dean, and Hugging Face and Scale AI co-founders joining. In Sep 2024 it closed a Series A of over $100M led by NEA, Khosla and Lux, with NVIDIA participating and partnering on research, data-center access and Japan AI talent. In Nov 2025 it raised a Series B of about ¥20B ($135M), lifting cumulative funding to about ¥52B ($347M).

新闻Forbes Japan: $100M+ Series A with NVIDIA 报道财联社: Sakana A-round with NVIDIA

结果

Still live and scaling: after the Series B it is building Japan-market-optimized foundation models and deploying AI in manufacturing and other core industries while expanding globally — funded without following OpenAI-style megascale spending.

新闻創業手帳: ¥20B Series B, cumulative ¥52B

背景

Sakana AI was founded in Tokyo in July 2023 by three insiders: Google Brain's David Ha; Llion Jones, co-author of the transformer paper that underpins ChatGPT; and Ren Ito, ex-COO of Stability AI and ex-CEO of Mercari Europe. The company's name means 'school of fish' in Japanese — the bet is that many small models coordinating and evolving can outperform one monolithic giant, making world-class AI possible without megascale compute.

In Jan 2024 Sakana raised a $30M seed led by Lux Capital with Khosla Ventures, joined by Sony, NTT, KDDI, Google's Jeff Dean and co-founders of Hugging Face and Scale AI. In Sep 2024 it closed a Series A of over $100M led by NEA, Khosla and Lux, with NVIDIA participating and partnering on research, data-center access and Japan AI talent. It had already shipped models for image generation and Japanese-language tasks, trained on small datasets.

The efficiency thesis kept raising money: in Nov 2025 Sakana announced a Series B of about ¥20B ($135M), lifting cumulative funding to about ¥52B ($347M), earmarked for Japan-market-optimized foundation models, deployment in manufacturing and core industries, and global expansion. The company is still scaling, having built its reputation as Japan's flagship AI startup without following OpenAI-style megascale spending.

这件事要成立，得有什么

Sakana attacked the dominant scaling paradigm: instead of bigger data and compute, it bet on small, nature-inspired models that evolve and cooperate, making efficiency itself the differentiator.
Japan's constraints became the pitch: limited compute and energy pushed the founders to design for sustainability, turning a weakness into a thesis investors could fund.
Founder pedigree mattered: a Transformer paper co-author and a Google Brain researcher gave a small Tokyo lab credibility to raise from top US VCs and NVIDIA.
Strategic partnerships with NVIDIA, Sony, NTT and KDDI substituted for scale, plugging the startup into research, compute access and Japan's corporate ecosystem.

可借鉴之处

Compete with giants by attacking their cost base: Sakana reframed AI as an efficiency problem, turned Japan's limited compute into a thesis, and investors funded it.

后续进展

As of 31 Aug 2026, Sakana AI remains private and scaling: after the Nov 2025 Series B (~¥20B, cumulative ~¥52B) it is developing Japan-market-optimized foundation models, pushing AI into manufacturing and other core industries, and expanding globally. Public disclosures do not yet include revenue, profit, or a confirmed valuation; the company has not followed OpenAI-style compute spending, staying true to its small-efficient-models thesis.

资料来源

发现哪里写错了？告诉我们。

轮到你了

你刚读完一家。说说你在做什么，看看谁在赌同一件事。

免费账号 · 3 次免费提问 · 不用绑卡