速報｜小魚は、サメを使えるか——Sakana AI上場準備報道、「国産」の強み・依存と後発の勝負｜Small Fish, Big Ocean
Can the Little Fish Coordinate the Sharks? Sakana AI’s Reported IPO Preparations TokyoBrainie · Quick Cut · 9 November 2026 JST. Reported preparations are distinct f
Can the Little Fish Coordinate the Sharks? Sakana AI’s Reported IPO Preparations
TokyoBrainie · Quick Cut · 9 November 2026 JST. Reported preparations are distinct from approved listing terms. The sea creatures are fictional business satire, not a performance or market-share ranking.
Three findings
Nikkei reports IPO preparations, a possible Japan/U.S. listing as early as autumn 2027 and Goldman Sachs as adviser. These are reported plans, not an approved offering. Sakana is a Japan-based developer, but some products use overseas foundation models. Being Japanese does not establish wholly domestic technology or domestic-only data processing. Late entry creates disadvantages, yet the relevant races include adaptation, orchestration and enterprise integration—not only training the largest model.
1. What is the news?
The Nikkei report concerns preparations. Listing applications, exchange approval, issue size and offer price have not been established in the company materials reviewed here. An adviser is not automatically the formally appointed lead underwriter.
An IPO can fund the company and provide liquidity to existing shareholders; the split depends on primary issuance versus secondary sales. Preparation proves neither profitability nor international competitiveness. The next evidence is a prospectus, financial statements, revenue mix, cash burn, use of proceeds and dilution.
2. What does Sakana actually do?
Founded in Tokyo in July 2023, Sakana’s co-founders are David Ha, Ren Ito and Llion Jones. Ha led Google Brain’s Tokyo research team; Jones co-authored the Transformer paper, “Attention Is All You Need.” Their experience is relevant evidence of research capability, not a guarantee of future earnings. Company information
The company researches collective intelligence, evolution-inspired techniques, combining models and AI-assisted research and algorithm development. Its strategy is broader than building a single ever-larger model. Research/business strategy
| Product | Function and qualification |
|---|---|
| Fugu | A trained coordinator delegates work to models and combines results. More calls may add cost and latency |
| Namazu | Adapts an existing model to Japanese and local business needs. The August 3 version uses Moonshot AI’s Kimi K2.6 |
| Chat / Translate | User-facing chat and translation. Interface quality and underlying model capability are separate |
| Marlin | Research and report production; assess source accuracy, omissions and correction time |
| Aquarium | Connects organizational knowledge, rules and actions; requires permissions, human approvals and auditability |
| API Platform | Enterprise model access and administration. October 28 early access was not general availability for every product |
Fugu / Namazu / Aquarium / API platform
“Orchestration” means controlling roles and collaboration among AIs. An “ontology” makes entities such as customers, contracts and products—and their relationships and rules—explicit. The business test remains fewer errors and less work.
3. What does “Japanese” establish?
| Dimension | Potential advantage | What it does not guarantee |
|---|---|---|
| Local R&D | Closer feedback from Japanese users and operations | Best global performance, lowest price or uninterrupted service |
| Language and workflow adaptation | Focused tuning for terminology, documents and approvals | Elimination of hallucinations or superiority on every Japanese task |
| Local implementation and accountability | Easier contractual and operational coordination | Automatic compliance, security or compensation |
| Data and deployment control | Contractual control over processing, storage and access | Domestic-only processing merely because the supplier is Japanese |
| Technology supply | More interchangeable models can reduce single-vendor exposure | Independence from overseas chips, clouds and licensing |
Using downloaded model weights is not equivalent to sending every request to that model developer’s foreign API. The actual deployment, external calls and logging determine data movement. Namazu
Sovereignty involves separate controls over model rights, compute, maintenance, talent, data and contractual change. A domestic headquarters satisfies one condition, not every condition.
Using overseas technology can be rational if it delivers suitable quality faster and at lower cost. Dependence becomes problematic when price changes, restrictions or service interruptions cannot be managed, or contractual obligations are unclear.
4. Is it late?
| Race | Disadvantage | Possible route—and its falsifier |
|---|---|---|
| Largest general foundation models | Compute, capital, talent and accumulated data | Adaptation and coordination can save resources, but retain upstream dependence |
| Everyday consumer assistants | Existing habits, brands and distribution | A compelling specific use case; patriotic trial alone is not retention |
| Enterprise operations | Incumbents’ systems, sales channels and bundled contracts | Domain data, workflow knowledge and deployment track record; bespoke work can consume margins |
| Multi-model coordination | Incumbents can add similar functions | Better outcomes at comparable total cost; comparisons must use matched conditions |
This is structural inference. Later entrants may avoid recreating existing technology, while still paying upstream suppliers. The difference between customer receipts and the full delivery cost determines economics. Small does not automatically mean efficient; late does not automatically mean defeated.
Models can compensate for one another, but can also share false assumptions or fail at verification. Benchmark comparisons need model versions, attempts, total compute, latency and human intervention. More agents do not guarantee truth. Technical report
5. The monsters are not all enemies
Claude is a model/service, Copilot a product family, Gemini a model/service brand and Meta a company. They are not identical comparison units.
Sakana announced a Google investment and strategic partnership in January 2026, explicitly including Gemini/Gemma use. Earlier Namazu prototypes also used Meta’s Llama. Competitors can be suppliers and partners. Google partnership / Namazu prototypes
TokyoBrainie’s Cut: A little fish need not grow a shark’s teeth from scratch; it may coordinate a school. But if calling the school incurs charges, it must swim through the invoice too.
6. Japan already has commercial connections
SMBC’s April 30 announcement says it is starting deployment of a jointly developed proposal-generation application. That is evidence beyond a research demo, not proof of bank-wide adoption or a quantified investment return. Customer release
The September 10 Sakana–SCSK–Sumitomo partnership combines research, systems integration and industrial distribution. A partner’s addressable customer base is not Sakana’s paying-customer count. Joint release
Local users may gain implementation help while incurring data preparation, training, verification and governance costs. Those are transmission hypotheses, not measured benefits. An investor’s listed shares also do not automatically revalue: ownership, carrying value, lockups and significance to the parent matter.
7. Which numbers are established?
The Series B announcement, originally November 2025 and updated in April 2026, states approximately JPY 32bn raised, JPY 432bn post-money valuation and JPY 66bn cumulative funding, using JPY 160/USD. Those are historical financing figures, not a current tradable share price or IPO price. Funding disclosure
Valuation is not revenue or cash. Cumulative funding is not the remaining cash balance. Private financing securities may carry different rights from public ordinary shares.
Company benchmarks help explain the products but do not prove blanket superiority over overseas rivals. Check tasks, versions, cost and disclosure. Namazu’s FairPoliticsQA is an internal assessment; a score about response bias cannot be relabelled general business accuracy. Evaluation methodology
The reviewed announcements do not establish all the revenue, recurring sales, concentration, unit economics and operating cash-flow figures necessary for valuation. This article therefore makes no cheap/expensive, profitability-date or post-IPO price claim.
8. Five numbers to watch
- Paying recurring contracts and renewal rates, separate from free users or partners.
- Total cost per successfully completed task, including retries, retrieval, verification and human corrections.
- Implementation effort and support cost per customer, testing product reuse versus bespoke development.
- Gross profit and operating cash flow, accounting for models, cloud and labour.
- New shares, secondary sales and post-offering share count, separating company financing, shareholder liquidity and dilution.
Provisional assessment: Local development and workflow proximity can matter. Overseas models offer efficiency alongside dependencies. Judge the specific task, comparable cost, safety requirements and recurring economic value. Supporting a Japanese company does not exempt it from demonstrating sound economics.
