When Frontier AI Emerges from Outside Silicon Valley
Japan’s Sakana AI is changing the playbook on how to build frontier AI—not by scaling, but through evolution
Frontier AI used to be bound by geography and compute. Tokyo-based Sakana AI is challenging both.
For years, there has been an assumption that to be a serious AI player, one must be rooted in Silicon Valley, where the capital, talent, and network to build ambitious ventures reside. But the risks of geographic concentration are increasingly clear. If all frontier AI models come from one region, they will inevitably reflect the priorities, values, and needs of that particular region. Diversity gets sacrificed in the process.
Sakana AI’s recent funding round—the company raised a US$135 million Series B at a US$2.65 billion valuation last month—signals shifts in AI innovation. First, that you no longer need to be in the Bay Area to build frontier models. Second, that being outside the Silicon Valley bubble can be an advantage. For Sakana, distance has been liberating: it emboldened the company to develop models in ways that break from the brute-force scaling paradigm dominating U.S. labs.
Sakana AI is forming a new axis in the global AI landscape, challenging the gravitational pull of the U.S. and China tech clusters. A broader set of innovators is now contributing to a more diverse future for AI models—a development not to be taken for granted, given the influence frontier AI players have in shaping a new world.
A Tipping Power Scale
Sakana AI wants architectural sovereignty, not just representation. The company’s co-founder and CEO, David Ha, has been vocal about the need to build AI ecosystems outside the influence of the West and China. “I think it is time for Japan to shine in the AI space,” he posted on Twitter (now X) in June 2023. Ha launched Sakana AI the following month.
Ha saw AI rapidly gaining traction across Japan, even being adopted for municipal governance in Yokosuka. “I have mixed feelings about this,” he said at a public forum in 2023. “Japan needs to develop its own strong R&D ecosystem. The problem is the big companies might not want other countries to develop their own AI ecosystem.”
Around the same time, OpenAI CEO Sam Altman faced backlash after downplaying the possibility of foundational models emerging from regions such as India. Although he later clarified his remarks, the sentiment implied in his “hopeless” comment captured why innovators outside Silicon Valley often feel dismissed. The backlash galvanized many—including Ha—who publicly pushed back.
The critique of concentration has only intensified. More recently, OpenAI’s dominance has come under sharp scrutiny in Karen Hao’s Empire of AI, which interrogates the power the company wields over the industry. With mounting concerns over such influence, the public is waking up to the need for a more diversified playground for innovation—and for architectural plurality.
Architectures Beyond Scaling
Sakana AI is challenging Silicon Valley not only on ideology but on method. While the Valley has long operated on a doctrine of brute-force scaling—pursuing ever-larger models that demand exponentially greater computing resources—Sakana is charting a different path. One inspired by evolution.
Through this evolutionary lens, Sakana is developing competitive frontier models with far lower computational overhead, positioning itself as a credible alternative at a time when leading AI figures, including Ilya Sutskever, have begun signaling that the era of pure scaling is nearing its limits.
Central to Sakana’s strategy is its evolutionary model-merging technique, which iteratively combines and refines smaller models to produce architectures with emergent capabilities. Instead of training one massive model from scratch, Sakana generates populations of models, tests them, recombines their strongest traits, and repeats the cycle—much like biological systems evolving toward greater fitness. This approach conserves compute and fosters diversity.
Sakana’s work on character-level language modeling further shows how innovation is shaped by context. Llion Jones, Sakana’s co-founder and CTO—who contributed to the foundational Transformer model at Google—has explored character-level modeling for years, but says the wider AI community has been slow to adopt it. He contends that cultural friction is impeding the shift. It’s easier to move from word- to character-level modeling in Japanese than in English, because far more meaning is packed into each written unit, making the language especially well-suited to this approach.
Frontier AI advances not just through more power, but through smarter architectures rooted in the particularities of a region. Local linguistic and structural constraints at Sakana have produced design choices that never emerged inside Silicon Valley labs.
Structures Beyond the Valley
Sakana AI’s rise isn’t happening in isolation. Japan is assembling a parallel support stack—compute access, grants, and institutional partners—that enables frontier work outside the Bay Area’s gravitational field.
One cornerstone is the Generative AI Accelerator Challenge (GENIAC), a government initiative launched in February 2024 to support companies building foundation and generative AI models. As part of this push, Sakana AI received a supercomputing grant that allowed it to train new foundation models for several months in 2024. While the company operated in “GPU-poor” mode—using constraints as a creative engine—it was able to develop its core ideas on nature-inspired AI models through the provision of computing power from the Japanese government.
Beyond government support, Sakana is also constructing its financial and commercial ecosystem deliberately. It is assembling a cap table that doubles as a client base, providing early revenue pathways while building a network of partners invested in its success.
Together, these structures form a genuine alternative center of frontier AI gravity, and Sakana now sits near its core. Its rise shows that frontier AI is beginning to diversify—away from the geographic concentration and architectural paradigm of the Valley.
If the Valley built AI through scale, Asia’s contribution may well come through evolution.