When Llion Jones left Google in 2023, it raised eyebrows. Jones wasn't just any researcher—he'd co-authored "Attention Is All You Need," the 2017 paper that birthed the Transformer architecture now underpinning everything from ChatGPT to Google's own models. Walking away from the epicenter of the AI revolution to launch a startup in Tokyo? That demanded explanation.
The explanation arrived November 17, 2025, in the form of a ¥20 billion Series B round (roughly $129 million) that values Sakana AI at $2.6 billion. It's a striking figure for a company barely two years old, and it positions Sakana as Japan's most serious homegrown contender in the race to build what insiders are calling "sovereign AI"—infrastructure tailored explicitly for Japanese enterprises and government, not adapted from Western models as an afterthought.
The funding mix tells its own story. Returning investors MUFG, Khosla Ventures, New Enterprise Associates, and Lux Capital were joined by Macquarie Capital, Mouro Capital (the venture arm of Spain's Santander Group), In-Q-Tel—the CIA's strategic investment vehicle—and Shikoku Electric Power's STNet subsidiary. More names are expected in coming weeks, according to the company. It's a roster that reads less like typical Silicon Valley backing and more like a consortium preparing for something geopolitically consequential.
An Unusual Founding Trio
Jones brought David Ha and Ren Ito along for the ride. Ha, formerly head of research at Stability AI and a Google Brain alum, handles the CEO role. Ito—an ex-official at Japan's Ministry of Foreign Affairs who later served as COO at Stability AI UK—provides the diplomatic and strategic glue. It's an eclectic mix. A Transformer pioneer, a generative AI veteran, and a former diplomat. The chemistry matters, particularly when you're trying to navigate Japan's notoriously consensus-driven corporate culture while pitching AI that sounds, on the surface, less like conventional machine learning and more like evolutionary biology.
That's the hook: "nature-inspired" AI. Sakana draws from principles of evolution and collective intelligence rather than the brute-force compute scaling that defines most frontier labs. In practice, this translates to techniques like Evolutionary Model Merge, which optimizes combinations of pre-existing models instead of training massive new ones from scratch. The method earned a publication in Nature Machine Intelligence this past January—a meaningful validation, though skeptics note that publication doesn't equal commercial viability.
Corporate Japan Goes All-In
If Sakana's approach sounds niche, its investor roster suggests otherwise. The company's Series A in September 2024 pulled in approximately $200 million from Japan's financial and industrial heavyweights: MUFG, Sumitomo Mitsui Banking Corporation, Mizuho, NEC, SBI Holdings, Dai-ichi Life, Itochu, KDDI, Fujitsu, Nomura, ANA Holdings, Tokio Marine. NVIDIA joined too, with an ongoing collaboration focused on data centers and AI research infrastructure in Japan.
That corporate backing has already converted into revenue. In May 2025, MUFG inked a ¥5 billion deal to automate banking document creation using Sakana's AI Scientist tool. Daiwa Securities followed in October with a 3.5-year partnership aimed at building an AI-driven wealth advisory platform. These aren't pilot programs. They're multi-billion-yen bets.
Perhaps more telling is Sakana's pivot toward defense and intelligence work. In March 2025, the company was named a finalist in both themes of the US-Japan Global Innovation Challenge 2025, a joint initiative between Japan's Acquisition, Technology & Logistics Agency and the US Defense Innovation Unit. CEO Ha has been unusually direct about Japan's need for indigenous AI defense capabilities. "Japan should produce its own AI defense solutions," he told the Japan Times in May—a statement that would sound routine in Washington but carries weight in a country still navigating its post-pacifist security posture.
The Post-Training Bet
Sakana's technical strategy acknowledges Japan's limitations. The country doesn't have America's hyperscale compute infrastructure or China's state-backed GPU farms. Its language and cultural specificity means Western-trained models often miss the mark. So rather than compete on training ever-larger models, Sakana focuses on post-training optimization—taking existing frontier models and adapting them for Japanese contexts through inference-time scaling and multi-model cooperation.
Take AB-MCTS, which orchestrates multiple large language models—ChatGPT, Gemini, DeepSeek—to collaborate on complex tasks. The company open-sourced TreeQuest, the codebase behind this approach, and published results on the ARC-AGI-2 benchmark. It's clever. Whether it's enough is another question.
After all, Sakana faces well-funded rivals: OpenAI, Anthropic, and domestic challengers like NTT (with its "tsuzumi" models) and SoftBank (backing "Sarashina"). And the company has already stumbled. In February 2025, Sakana walked back bold claims that its AI CUDA Engineer could deliver 10-100× training speedups after users uncovered bugs and reward hacking issues. Then there was AI Scientist v2's assertion of producing "the first fully AI-generated paper to pass peer review"—a claim that unraveled when it emerged the paper was withdrawn before meta-review at an ICLR workshop, not published in a peer-reviewed journal as implied.
These missteps reveal the tension between startup hype cycles and scientific rigor. They also underscore the difficulty of Sakana's position: too small to out-compute the giants, too ambitious to settle for incremental improvements.
Sovereign AI's Moment
The Series B capital will fund hiring across engineering, sales, and distribution in Japan, along with strategic investments, partnerships, and potential acquisitions. Sakana's stated goal is to "democratize AI in Japan" with an emphasis on energy efficiency and domestic data sovereignty—buzzwords, maybe, but ones that resonate in a country acutely aware of its dependence on foreign technology infrastructure.
With cumulative funding now at roughly ¥52 billion (around $347 million), Sakana has the runway to test whether post-training optimization and nature-inspired techniques can hold ground in an AI market increasingly dominated by compute-rich incumbents. The Japanese corporate and government sectors are clearly betting it can.
Whether Jones's gamble pays off remains to be seen. But in an AI landscape where most innovation flows from California or China, a Tokyo-based startup with Transformer credentials and deep ties to Japan Inc. represents something rarer than another language model: an alternative pathway. One that, for better or worse, dozens of Japan's most powerful institutions are now funding with real money and real expectations.
