Sakana AI
TL;DR: Sakana AI reached a $2.65 billion valuation in November 2025, becoming Japan’s most valuable private startup after raising $135 million in Series B funding. Founded by former Google researchers David Ha and Llion Jones—co-author of the seminal “Attention Is All You Need” paper—the Tokyo-based company is pioneering nature-inspired AI that challenges Silicon Valley’s compute-intensive approach. Through revolutionary technologies like Evolutionary Model Merge and the AI Scientist system, Sakana has secured partnerships with financial giants MUFG and Daiwa Securities while positioning itself as the architect of Japan’s sovereign AI infrastructure.
The Lightning Rise of a Tokyo Unicorn
Two years ago, Sakana AI didn’t exist. Today, it commands a $2.65 billion valuation and stands as Japan’s crown jewel in the global AI race. The company’s trajectory defies conventional startup wisdom: while most AI ventures chase billion-dollar compute budgets and massive language models, Sakana built its empire on constraints, evolution, and collective intelligence.
The November 2025 Series B round brought in $135 million from heavyweight investors spanning three continents. Mitsubishi UFJ Financial Group led the charge alongside returning backers Khosla Ventures, New Enterprise Associates, and Lux Capital. The round’s most intriguing participant? In-Q-Tel, the venture capital arm serving U.S. intelligence agencies, signaling Sakana’s strategic importance beyond commercial applications.
This latest funding pushed total capital raised past $479 million across four rounds. The valuation jump from $1.5 billion in September 2024 to $2.65 billion represents 77% growth in roughly 14 months—a remarkable pace for an R&D-focused AI company operating outside Silicon Valley’s ecosystem.
The Founders Who Broke From Google
David Ha and Llion Jones didn’t leave Google because the tech giant lacked resources. They departed because bureaucracy had suffocated the very research culture that once defined Google’s AI dominance.
Jones, originally from Wales, co-authored “Attention Is All You Need” in 2017—the landmark paper that introduced transformer architecture and unleashed the generative AI revolution. All eight co-authors of that paper have since left Google, most founding their own ventures. The exodus speaks volumes about where AI’s cutting edge believes innovation happens.
Ha brought his own impressive credentials: former head of Google Brain’s Tokyo research team, subsequent research director at Stability AI, and a career that started at Goldman Sachs before pivoting to machine learning. His doctorate from the University of Tokyo cemented his connection to Japan’s research ecosystem.
The third co-founder, Ren Ito, complements the technical duo with deep operational expertise. A University of Tokyo law graduate who later earned an LLM from New York University, Ito led global expansion at Mercari, one of Japan’s early unicorns, and served in Japan’s Foreign Service. His diplomatic background proves particularly valuable as Sakana navigates government partnerships and defense applications.
Together, they launched Sakana in July 2023 with a provocative thesis: AI’s future wouldn’t be written by whoever could afford the biggest GPU clusters.
Nature’s Blueprint for Machine Intelligence
The company’s name—”sakana,” Japanese for fish—encapsulates its core philosophy. Watch a school of fish navigate the ocean: thousands of individuals moving as one coherent entity, no central commander issuing orders, just simple rules creating complex behavior. This collective intelligence, honed by millions of years of evolution, inspired Sakana’s approach to AI.
While OpenAI, Anthropic, and Google pour billions into training ever-larger models from scratch, Sakana asks a different question: What if we could evolve new AI capabilities from existing models, the way nature creates new species from existing genetic material?
The founders saw compute scarcity not as a limitation but as a design constraint that could spark genuine innovation. Japan, lacking the massive data centers and energy infrastructure of the United States or China, needed a different path to AI sovereignty. Sakana would build that path.
Evolutionary Model Merge: The Breakthrough
In March 2024, barely eight months after founding, Sakana published research that would cement its reputation: Evolutionary Model Merge. The paper, later accepted by Nature Machine Intelligence, introduced a method for automatically combining multiple AI models into new architectures with emergent capabilities.
The technique draws directly from Darwinian evolution. Instead of manually tuning how models merge—a process requiring deep expertise and intuition—Sakana uses evolutionary algorithms to explore thousands of possible combinations. The system treats existing open-source models as “parent” DNA, creates hundreds of “offspring” with different trait combinations, tests their performance, and lets only the fittest survive to the next generation.
The results astounded the AI community. Sakana evolved Japanese-language models capable of mathematics by merging a Japanese language model with an English mathematics model—solving a problem that would normally require training a new model from scratch on a dataset that didn’t exist. The evolved models discovered “unintuitive” ways to combine capabilities that human engineers wouldn’t have conceived.
The implications extended beyond Japanese-specific applications. With over 500,000 open-source AI models now available, Evolutionary Model Merge democratized advanced AI development. Small teams and startups could now create specialized, powerful models without the compute budgets of tech giants. The technique has since been implemented in popular frameworks like mergekit and Optuna Hub, enabling thousands of developers worldwide to create unique model combinations.
The AI Scientist: Machines Doing Science
If Evolutionary Model Merge proved Sakana could revolutionize how AI models are built, the AI Scientist demonstrated something more unsettling: machines conducting the entire scientific research process autonomously.
Released in August 2024 through collaboration with Oxford and the University of British Columbia, the AI Scientist system executes the complete lifecycle of scientific research without human intervention. Given a broad research direction, it brainstorms hypotheses, searches academic literature for novelty, designs experiments, writes and debugs code, runs tests, analyzes results, generates figures, writes complete manuscripts in LaTeX, and even conducts peer review.
The economics proved as striking as the capability. Each paper cost approximately $15 to produce—several orders of magnitude cheaper than human-led research. The system could run continuously, never sleeping, never needing sabbaticals, generating dozens of research papers in the time a human researcher produces one.
In March 2025, Sakana achieved another milestone: the first fully AI-generated paper accepted through peer review at a major AI conference. “Compositional Regularization: Unexpected Obstacles in Enhancing Neural Network Generalization” received an average reviewer score of 6.33 at an ICLR 2025 workshop—just above the acceptance threshold but good enough to pass blind peer review where reviewers didn’t know which papers were AI-generated.
The achievement sparked intense debate. Reviewers noted “definitional imprecision” and “contradictions between claims and data,” criticisms that could apply to many human-authored papers. The paper wasn’t brilliant, but it cleared the bar. More importantly, it demonstrated that AI-generated research could fool expert reviewers—a watershed moment with profound implications for scientific publishing.
The Enterprise Validation
While academic achievements built Sakana’s reputation, enterprise partnerships proved the technology could solve real-world business problems. In May 2024, Sakana announced a multi-year partnership with Mitsubishi UFJ Financial Group—Japan’s largest bank and one of the world’s most influential financial institutions.
The deal involved developing custom AI models tailored to MUFG’s specific operational needs: Japanese financial regulations, tacit business knowledge, professional culture, and the nuances of Japanese client relationships. Generic LLMs from OpenAI or Anthropic couldn’t capture these domain-specific requirements. Sakana’s evolutionary approach could.
CEO David Ha told media the MUFG partnership would enable Sakana to achieve profitability within a year—remarkable for an R&D-focused startup barely 12 months old.
October 2025 brought a second major financial partnership. Daiwa Securities Group, one of Japan’s largest securities firms, contracted Sakana to build an AI platform for asset consulting. The three-and-a-half-year agreement would create tools enabling Daiwa advisors to deliver personalized financial services to retail investors, from first-time investors to high-net-worth clients.
These partnerships exposed a critical gap in the AI market: general-purpose models struggled with domain-specific professional applications requiring deep cultural and regulatory knowledge. Sakana’s ability to evolve specialized models from existing architectures—without massive retraining costs—offered a practical solution.
MUFG CEO Hironori Kamezawa articulated the vision: extending AI benefits beyond banking operations to “Japan’s diverse industries.” The bank planned to hire over 350 AI specialists by March 2027, positioning itself as an “AI native company” with Sakana’s technology at its core.
Beyond Finance: Defense and Manufacturing
The Series B funding signaled Sakana’s expansion beyond financial services. In-Q-Tel’s participation telegraphed interest from defense and intelligence communities, both in Japan and among U.S. allies.
Japan’s defense posture has evolved dramatically in recent years. The government committed to raising defense spending to 2% of GDP amid regional tensions. AI-driven intelligence analysis, cybersecurity, and autonomous systems represent critical capability gaps. Sakana’s culturally-attuned, Japanese-language models could enhance these applications while maintaining data sovereignty—keeping sensitive information within Japan’s borders rather than flowing through U.S. cloud infrastructure.
The company also won a government contract in June 2025 from Japan’s Ministry of Internal Affairs and Communications to develop technology combating online misinformation—a growing concern in an era of deepfakes and AI-generated disinformation.
Manufacturing represents another expansion vector. Japan’s industrial base—automotive, robotics, electronics—faces acute labor shortages as the population ages. AI automation could maintain productivity with fewer human workers. Sakana’s efficient models, optimized for resource constraints, align perfectly with Japan’s limited energy and compute resources compared to the United States or China.
The Sovereign AI Doctrine
Sakana’s rise cannot be separated from Japan’s broader push for technological sovereignty. As U.S.-China tensions escalate and AI emerges as a strategic technology comparable to semiconductors or aerospace, nations increasingly view dependency on foreign AI infrastructure as a national security vulnerability.
Japan watched as ChatGPT and other U.S. models struggled with Japanese language nuances, cultural context, and regulatory requirements. Chinese models, while sometimes more effective with Asian languages, raised different security concerns. The solution? Build domestic AI capability tailored to Japanese needs.
Sakana positioned itself as the architect of this sovereign AI ecosystem. Its founding philosophy—that “intelligent life has arisen not from an abundance of resources but rather from the lack of it”—resonated with Japan’s resource constraints. The country lacks massive hydroelectric dams or cheap natural gas to power energy-intensive data centers. Solar and nuclear capacity remains limited. Training frontier models from scratch would require enormous capital and energy Japan couldn’t sustainably deploy.
Evolutionary Model Merge offered an alternative path. By combining and optimizing existing models rather than training new ones from scratch, Sakana could deliver advanced AI capabilities with a fraction of the compute and energy requirements. This approach aligned with Japan’s strengths: sophisticated engineering, attention to efficiency, and the ability to do more with less.
The government supported this vision through programs like NEDO GENIAC, providing Sakana access to government-sponsored GPU clusters for rapid experimentation. Partnerships with Sony, NTT, and KDDI further embedded Sakana into Japan’s tech ecosystem.
The Global Investor Coalition
Sakana’s Series B cap table reads like a who’s who of global venture capital and strategic investors, revealing the company’s positioning at the intersection of research, enterprise applications, and geopolitical strategy.
Mitsubishi UFJ Financial Group’s repeated investment across rounds signals conviction that Sakana’s technology will transform financial services. MUFG’s CEO personally championed the investment, stating: “As a business leader myself, I feel a responsibility to lead the implementation of AI across all Japanese companies.”
Khosla Ventures, Lux Capital, and New Enterprise Associates—Silicon Valley’s elite AI-focused funds—validated Sakana’s technical approach. These investors backed the original wave of transformer-based companies; their support for Sakana’s evolutionary methods suggests conviction that post-transformer paradigms will define AI’s next phase.
In-Q-Tel’s participation deserves special attention. The CIA-linked venture arm invests in technologies with national security applications. Its presence in Sakana’s Series B indicates the U.S. intelligence community sees value in Japan developing robust AI capabilities—potentially as a counterbalance to China’s AI ambitions in the Asia-Pacific region.
MPower Partners, Japan’s first ESG-integrated global VC fund, brought a sustainability lens. General Partner Kathy Matsui, former Goldman Sachs Japan Vice Chair, noted: “Sakana AI’s unique determination to develop frontier AI technology sustainably, through innovation, aligns with our core values.”
The investor mix—Japanese financial institutions, Silicon Valley venture capital, U.S. intelligence funding, and ESG-focused capital—reflects Sakana’s unique positioning. The company isn’t just building AI technology; it’s constructing infrastructure for Japan’s technological independence while maintaining strategic ties to Western allies.
The Contrarian Bet Against Scale
Sakana’s philosophy directly challenges the dominant AI development paradigm. OpenAI reportedly spent over $100 million training GPT-4. Anthropic raised $7 billion across multiple rounds to fund Claude’s development. Google and Meta measure their AI investments in tens of billions annually.
The conventional wisdom: bigger models trained on more data with more compute will always outperform smaller, more efficient alternatives. The scaling laws discovered in the late 2010s suggested that doubling model parameters consistently improved performance. Why would anyone bet against scale?
Sakana saw cracks in this logic. First, the environmental cost: training a single large language model can emit as much carbon as five cars over their entire lifetimes. Energy consumption for AI training doubled every 3-4 months between 2012 and 2024. This trajectory isn’t sustainable—literally.
Second, the economic cost: AI companies were burning capital faster than they could generate revenue. OpenAI reportedly lost $5 billion in 2024 despite massive ChatGPT adoption. The path to profitability remained unclear for companies spending hundreds of millions on compute before earning a dollar of revenue.
Third, the diminishing returns: improvements from scaling began plateauing. Going from 1 billion to 10 billion parameters produced dramatic gains. Going from 100 billion to 1 trillion showed smaller improvements at exponentially higher cost.
Sakana’s evolutionary approach offered a fundamentally different value proposition: leverage the collective intelligence of 500,000+ existing open-source models rather than reinventing capabilities from scratch. Focus innovation on post-training optimization—how models are combined, fine-tuned, and deployed—rather than pre-training scale.
The company’s research validated this approach. Japanese mathematics models evolved through merging showed competitive performance to models trained specifically for mathematics—at a fraction of the cost. The four-step diffusion models for image generation ran fast enough for real-time applications while maintaining quality.
Technical Innovations Beyond Evolution
While Evolutionary Model Merge garnered headlines, Sakana advanced multiple research frontiers simultaneously.
The Darwin Gödel Machine represents self-improving AI architecture where models modify their own code. Unlike traditional models with fixed architectures, DGM can rewrite its inference pathways, potentially discovering optimizations human engineers wouldn’t conceive.
ShinkaEvolve, released as open-source software, provides a framework for evolving LLM-generated programs with computational efficiency in mind. The system applies evolutionary pressure not just to model parameters but to the algorithms and code the models generate.
The AB-MCTS (Asynchronous Best-First Monte Carlo Tree Search) algorithm enables collaborative reasoning across multiple AI models. Instead of relying on a single model’s output, AB-MCTS orchestrates multiple models with complementary strengths, combining their insights for more robust results.
These techniques share a common theme: getting more intelligence from less compute by cleverly combining and evolving existing capabilities rather than brute-forcing larger training runs.
The Japanese Advantage
Sakana chose Tokyo as its headquarters despite most AI talent clustering in San Francisco, London, or Beijing. The decision proved strategic rather than sentimental.
Japan offers world-class technical infrastructure and a highly educated workforce. The country graduates over 100,000 engineering students annually. Universities like Tokyo, Kyoto, and Tohoku maintain strong AI research programs. Cultural factors—Japan’s engineering excellence, attention to detail, and long-term thinking—align with Sakana’s research-intensive approach.
The regulatory environment favors innovation. Japan’s government actively supports AI development through grants, supercomputer access, and favorable policies. Unlike Europe’s restrictive AI regulations or China’s heavy-handed government control, Japan provides breathing room for research while maintaining reasonable guardrails.
The market opportunity is enormous. Japan’s economy, the world’s third-largest, desperately needs AI solutions. An aging population (28% over 65) and declining workforce create acute labor shortages. Manufacturing, finance, healthcare, and logistics all face productivity crises that AI could address.
Crucially, Japanese companies understand they need AI tailored to their specific context. Generic English-language models fail at tasks requiring Japanese business knowledge, cultural nuance, or regulatory compliance. Sakana’s Japan-first approach positions it to capture this underserved market.
The founders also noted that being outside Silicon Valley’s echo chamber fostered creative thinking. Ha and Jones deliberately sought distance from the “bigger is better” mentality dominating AI research in the Bay Area.
Competitive Landscape and Positioning
Sakana operates in a complex competitive environment spanning multiple dimensions.
Against global AI giants like OpenAI, Anthropic, and Google, Sakana doesn’t directly compete for general-purpose model supremacy. Instead, it targets the specialized model market—domains where cultural specificity, regulatory requirements, or efficiency constraints favor tailored solutions over generic frontier models.
Within Japan, Sakana faces competition from U.S. companies establishing local presence. OpenAI partnered with SoftBank in February 2024, agreeing to supply $3 billion annually in technology to SoftBank and its subsidiaries. Google and Microsoft both operate AI research labs in Tokyo. These giants possess capital and brand recognition Sakana can’t match.
However, Sakana’s advantages include:
- Deep Japanese integration: Native understanding of language, culture, and business practices rather than afterthought localization
- Efficiency focus: Solutions designed for Japan’s resource constraints rather than assuming unlimited compute
- Sovereign control: Models and training data remaining within Japan rather than flowing to U.S. servers
- Government alignment: Close partnerships with ministries and public institutions seeking domestic AI capability
Against other Japanese AI startups, Sakana’s pedigree stands out. The founding team’s combination of Google research credentials, landmark paper authorship, and operational experience at Stability AI and Mercari surpasses most competitors. The $479 million in funding provides runway unavailable to smaller players.
The company’s positioning as both frontier research lab and commercial AI provider creates dual competitive moats. Academic credibility attracts top research talent. Enterprise partnerships generate revenue and validate practical utility. This combination—publishing Nature papers while deploying production systems at megabanks—distinguishes Sakana from pure-play research labs or purely commercial vendors.
Challenges and Criticisms
Despite impressive achievements, Sakana faces significant challenges.
Technical Limitations: Independent evaluations of the AI Scientist revealed problems. A February 2025 study found 42% of experiments failed due to coding errors. Literature reviews produced poor novelty assessments, misclassifying established concepts as novel. Some papers contained hallucinated results or flawed comparisons. While the system showed promise, the quality resembled “a rushed undergraduate paper” rather than leading research.
Scaling Questions: Evolutionary Model Merge works brilliantly for combining existing models, but can it produce genuinely frontier capabilities? The technique optimizes what exists; it doesn’t create fundamentally new architectures the way Transformers represented a breakthrough in 2017. Skeptics argue Sakana will hit ceilings that only massive training runs can break through.
Commercial Viability: While MUFG and Daiwa partnerships validate demand, Sakana must prove it can scale beyond pilot projects. Enterprise AI sales are notoriously slow. Implementation takes years. The path from technical demo to production deployment at a thousand-employee organization involves countless obstacles.
Talent Competition: Despite Japan’s engineering strength, the global war for AI talent favors companies offering Silicon Valley compensation packages and access to massive compute resources. Sakana must retain its founders and early employees while competing for new hires against offers from Google, OpenAI, and well-funded competitors.
Geopolitical Risks: As AI becomes increasingly tied to national security, Sakana’s position straddling U.S. (via In-Q-Tel) and Japanese interests could complicate certain applications. Export controls, technology transfer restrictions, and alliance politics may constrain commercial opportunities.
The Sustainability Narrative: While Sakana champions efficiency, critics note that combining models still requires compute. Training the “parent” models Sakana merges consumed enormous energy. The company optimizes model deployment but doesn’t eliminate the broader industry’s environmental footprint.
The $2.65B Question: Justified Valuation?
Sakana’s $2.65 billion valuation invites scrutiny. What justifies pricing a 27-month-old, pre-profit startup above many established tech companies?
Bulls point to:
- Technical moats: Evolutionary Model Merge and related techniques represent genuinely novel approaches with published research backing their effectiveness
- Marquee customers: MUFG and Daiwa aren’t taking flyers on unproven technology; these institutions conducted extensive due diligence
- Strategic positioning: As Japan’s sovereign AI champion, Sakana enjoys government support and preferential access to domestic enterprise customers
- Founder pedigree: Jones co-authored the most influential AI paper of the past decade; Ha built Google Brain’s Tokyo presence
- Market tailwinds: Japan’s AI market is projected to grow from $8 billion in 2024 to $40 billion by 2030
- Profitability trajectory: Ha claimed MUFG partnership alone could enable profitability—rare for AI startups
Bears counter:
- Unproven scale: Enterprise pilots don’t guarantee production deployments at scale
- Competitive threats: OpenAI, Anthropic, and Google have vastly more resources and could replicate Sakana’s techniques
- Small domestic market: Japan’s economy, while large, is a fraction of the U.S. or China’s size
- Limited international traction: Beyond Japan, Sakana hasn’t demonstrated it can compete globally
- Research risk: Evolutionary methods might hit fundamental capability ceilings requiring traditional massive training
- Valuation compression: AI valuations across the board declined through 2024-2025 as hype moderated
Comparable valuations provide context. Cohere, another transformer-paper-author-founded startup, raised at a $5 billion valuation in 2024. Anthropic reached $60 billion after its Series D. Mistral AI hit $6 billion valuation in 2024. By these standards, Sakana’s $2.65 billion seems reasonable for a company with proven technology, marquee customers, and strategic government backing—even if not yet dominant.
The valuation ultimately reflects investor conviction that Japan needs sovereign AI infrastructure and Sakana represents the best vehicle to build it. Whether that conviction proves justified depends on execution over the next 3-5 years.
Cultural Impact and AI Nationalism
Sakana’s success carries implications beyond its balance sheet. The company represents Japan’s reentry into the global AI conversation after years of perceived decline.
Japanese researchers led neural network breakthroughs in the 1980s-90s. Companies like Sony and NEC pushed the frontier. But the deep learning revolution of the 2010s centered on North America and China. Japan watched as Google, Facebook, Baidu, and Tencent hired AI talent, built massive infrastructure, and deployed products that reshaped industries.
Government officials worried Japan was missing another transformative technology wave—similar to how Japanese tech giants ceded smartphone dominance to Apple and Samsung despite pioneering mobile internet. The fear drove increased AI funding, university programs, and startup support.
Sakana became a symbol of Japanese AI resurgence. Local media covered the company extensively. David Ha appeared on NHK explaining his vision. The unicorn valuation triggered national pride—proof that Japanese startups could compete globally in cutting-edge technology.
More broadly, Sakana exemplifies “AI nationalism”—the recognition that AI infrastructure determines national competitiveness. Countries increasingly view reliance on foreign AI as similar to dependence on foreign oil or semiconductors. Developing domestic AI capabilities represents strategic imperative, not just commercial opportunity.
This nationalism manifests in multiple countries. France champions Mistral AI. The UAE backs Falcon models. China restricts data flows to ensure local models dominate domestically. The U.S. debates export controls to maintain technological lead. Sakana fits this pattern: a national champion building AI that reflects Japanese values, operates under Japanese control, and serves Japanese interests.
The Road Ahead: 2026 and Beyond
Sakana enters 2026 with ambitious expansion plans funded by the fresh Series B capital.
R&D Acceleration: The company will deepen research into self-evolving architectures and model merging techniques. Expect additional Nature publications and open-source releases that keep Sakana at the research frontier.
Enterprise Expansion: Beyond finance, Sakana targets manufacturing, government, defense, and intelligence sectors. The company aims to replicate its MUFG and Daiwa success across multiple industries.
Geographic Growth: While Japan remains the core market, Sakana indicated it will “actively pursue strategic investments, partnerships, and M&A for long-term global growth.” Expect expansion into other Asian markets with similar cultural-AI-alignment needs—potentially Korea, Taiwan, or Southeast Asia.
Defense Applications: In-Q-Tel’s investment presages deeper engagement with defense and intelligence communities. Sakana’s Japanese-language models could analyze intelligence, support cybersecurity operations, or power autonomous systems—all while maintaining data sovereignty.
Product Commercialization: Transitioning from research lab to product company requires shipping production-ready software, building sales and support teams, and navigating enterprise procurement. Sakana’s announcement that it will expand engineering, sales, and distribution teams signals this shift.
Potential IPO: At $2.65 billion valuation with strong revenue growth, Sakana becomes a potential IPO candidate within 18-24 months. A successful listing would provide liquidity for early investors and cement Sakana’s status as Japan’s AI champion.
The company faces a critical test: proving that efficiency-focused, evolution-based AI can compete with scale-focused approaches from better-funded rivals. Success would validate a new paradigm; failure would reinforce conventional wisdom that AI advancement requires massive capital and compute.
Lessons for the Global AI Industry
Sakana’s rise offers broader lessons transcending one company’s trajectory.
Constraints Spark Innovation: Resource limitations forced Sakana to innovate rather than outspend competitors. This echoes historical patterns—many breakthrough technologies emerged from resource-constrained environments where brute force wasn’t an option.
Post-Training Matters: The industry’s fixation on pre-training scale obscured opportunities in post-training optimization. Sakana demonstrated that how models are combined, fine-tuned, and deployed matters as much as initial training.
Open Source Acceleration: By building on 500,000+ open-source models rather than starting from scratch, Sakana achieved in months what might take years with proprietary approaches. The collective intelligence of the open-source community provided enormous leverage.
Cultural Specificity Creates Value: Generic models miss opportunities in non-English, non-Western contexts. Companies that deeply understand specific cultures, languages, and regulatory environments can capture markets global giants overlook.
Research-Product Balance: Sakana’s dual identity as research lab and commercial vendor created competitive advantages. Academic credibility attracts talent; enterprise revenue funds research. The flywheel effect accelerates both activities.
Government as Catalyst: Japan’s strategic support—supercomputer access, contracts, favorable policies—helped Sakana scale faster than purely commercial funding allowed. Other nations seeking to build AI champions should note this partnership model.
Frequently Asked Questions
What makes Sakana AI different from OpenAI or Anthropic?
Sakana focuses on evolving new models from existing open-source architectures rather than training massive models from scratch. This approach requires far less compute and energy while enabling rapid experimentation. The company optimizes for Japanese language and culture rather than building generic global models.
How does Evolutionary Model Merge work?
The technique treats existing AI models as genetic material. Evolutionary algorithms automatically explore thousands of ways to combine model parameters and layers, test the results, and select the best performers for further evolution. This mimics natural selection, discovering model combinations that human engineers wouldn’t intuit.
Can Sakana compete with companies that have 100x more funding?
Sakana’s efficiency-focused approach means it doesn’t need equivalent funding to achieve competitive results. By leveraging existing models and focusing on post-training optimization, the company can deliver specialized AI solutions at a fraction of the cost required for training frontier models from scratch.
What is the AI Scientist system?
The AI Scientist automates scientific research from hypothesis generation through manuscript writing. It designs experiments, writes code, runs tests, analyzes results, generates figures, and produces complete research papers—all without human intervention beyond specifying the research topic.
Why did the founders leave Google?
Both Ha and Jones cited bureaucracy and organizational inertia that prevented them from pursuing speculative long-term research. The startup environment offered freedom to explore unconventional approaches outside the scaling-focused paradigm dominating large tech companies.
What role does In-Q-Tel’s investment indicate?
In-Q-Tel, the CIA’s venture capital arm, invests in technologies with national security applications. Its participation signals U.S. intelligence community interest in Sakana’s efficient AI capabilities for defense and intelligence applications—likely through allied partnerships with Japan.
How does Sakana plan to become profitable?
Enterprise partnerships with MUFG, Daiwa Securities, and other Japanese corporations provide direct revenue. CEO David Ha stated the MUFG deal alone could enable profitability within a year of announcement. The company charges for custom model development and licensing.
What are the limitations of evolutionary model merging?
The approach optimizes existing capabilities but may not produce fundamentally new architectures equivalent to breakthrough innovations like Transformers. Critics argue there’s a ceiling to how much can be achieved by combining existing models without massive training runs.
Is Sakana really Japan’s most valuable startup?
Yes, as of November 2025. The $2.65 billion valuation exceeded all other Japanese private startups. Previous leaders included Preferred Networks (robotics/AI) and SmartNews (content aggregation), both valued below $2 billion.
What’s next for Sakana AI?
The company is expanding beyond finance into manufacturing, defense, and government sectors. It plans to deepen R&D, pursue strategic M&A, and potentially expand geographically across Asia. An IPO could occur within 18-24 months if growth continues.
How accurate are papers produced by the AI Scientist?
Independent evaluations found significant limitations. About 42% of experiments failed due to errors. Papers sometimes contained hallucinated results or flawed methodology. The system produces work resembling undergraduate-level research rather than leading scientific contributions—impressive for full automation but requiring human oversight.
Can Sakana’s approach work outside Japan?
The evolutionary model merging techniques are universally applicable. However, Sakana’s competitive advantage lies in deep Japanese cultural and regulatory knowledge. Replicating this advantage in other markets would require similar localization investments.
What environmental benefits does Sakana’s approach offer?
By avoiding massive training runs from scratch, evolutionary methods consume far less energy and produce lower carbon emissions. However, critics note this doesn’t eliminate the industry’s overall environmental impact—it optimizes around existing resource-intensive models.
How does Sakana attract talent against better-funded competitors?
The company offers researchers autonomy to pursue speculative ideas, opportunities to publish in top venues, and meaningful equity stakes in a high-growth startup. For engineers seeking impact on Japan’s technological sovereignty, Sakana provides mission-driven work unavailable at foreign tech giants.
What’s the relationship with Sony, NTT, and KDDI?
These major Japanese corporations invested in Sakana’s seed round and collaborate on research projects. The partnerships provide access to data, compute resources, and domain expertise while giving these incumbents strategic exposure to frontier AI research.
Could OpenAI or Google simply copy Sakana’s techniques?
The research is published openly, so techniques can be replicated. However, Sakana’s competitive advantage includes cultural knowledge, government relationships, and deep integration with Japanese enterprises—moats that can’t be easily copied even with equivalent technical capabilities.
How does the valuation compare to global AI startups?
Sakana’s $2.65 billion falls between mid-tier and top-tier AI startups. It’s well below Anthropic ($60B), OpenAI ($157B), and Mistral ($6B) but substantial for a company focused on a single national market rather than global scale.
What government support has Sakana received?
The company received supercomputer access through NEDO’s GENIAC program, won contracts from the Ministry of Internal Affairs and Communications, and benefits from favorable AI development policies. This represents strategic partnership rather than direct subsidies.
Is the AI-generated paper acceptance a big deal?
Yes. It marked the first fully AI-generated paper passing peer review at a major conference without human editing. While the paper’s quality was modest, the milestone demonstrated AI systems can produce work indistinguishable from human researchers to expert reviewers.
What risks does Sakana face?
Key risks include scaling commercial operations beyond pilot projects, retaining talent against better-compensated competitors, potential technical limitations of evolutionary approaches, geopolitical complications from U.S.-Japan defense partnerships, and execution risk inherent in young startups.
Conclusion: The Evolution of AI
Sakana AI’s journey from zero to $2.65 billion in 27 months represents more than one company’s success. It signals a potential paradigm shift in how AI development could evolve.
The conventional wisdom—that AI progress requires exponentially growing compute budgets and ever-larger models—may be giving way to approaches emphasizing efficiency, evolution, and intelligent combination of existing capabilities. Just as biological evolution produced human intelligence through iterative improvement rather than wholesale design, Sakana’s methods suggest AI progress might accelerate through clever recombination rather than brute computational force.
Whether this approach ultimately proves as capable as scale-focused methods remains uncertain. The next 2-3 years will test whether evolutionary model merging can produce genuine frontier capabilities or merely optimizes within existing capability bounds.
Regardless of the technical outcome, Sakana has already demonstrated that Japan can produce globally competitive AI companies, that resource constraints can drive innovation rather than merely limiting it, and that alternatives to Silicon Valley’s AI development paradigm deserve serious consideration.
For Japan, Sakana represents hope that the country can remain technologically relevant in the AI era. For the broader AI industry, it offers a potential roadmap for sustainable development that doesn’t require unsustainable energy consumption and capital expenditure.
The real question isn’t whether Sakana can reach $3 billion, $5 billion, or eventual IPO valuation. It’s whether the company’s approach—nature-inspired, efficiency-focused, evolution-driven—can compete with and potentially surpass the scale-obsessed methods dominating today’s AI landscape.
Two years in, with Nature publications, major enterprise contracts, and In-Q-Tel backing, the bet looks increasingly credible. But the story is just beginning. The evolution continues.
Disclaimer: This article is for informational purposes only and should not be construed as investment advice. AI technology markets remain highly speculative, and startup valuations can fluctuate significantly.
