Skip to content
ImpactArcher

Sakana AI

Tokyo AI lab building nature-inspired, model-agnostic AI — Japanese LLMs, multi-model orchestration and enterprise agents for finance, defense and industry.

Lean Canvas

Read the write-up

Canvas by Kenji HatanakaAs of Oct 12, 2026

Customer segments

  • Japanese megabanks, securities firms and regional banks (MUFG, SMBC, Daiwa, Hokkoku)
  • Japanese government and defense bodies (Ministry of Defense, ATLA, MIC)
  • Large manufacturers and trading-house networks (Mitsubishi Electric; Sumitomo Corp. x SCSK alliance)
  • Developers and businesses buying Japanese-tuned or multi-model LLM APIs (Fugu, Namazu)
  • Strategy, consulting and research teams using Marlin for long-form research

Early adopters

  • MUFG Bank corporate-lending teams (AI Lending Expert in live-case verification since Mar 2026)
  • Daiwa Securities wealth-management staff (production development from Aug 2026)

Problem

  • Regulated Japanese institutions need generative AI in Japanese with data control, but frontier models are US-built and single-vendor
  • Knowledge work such as loan memos, client proposals and research reports takes days to weeks amid labor shortages
  • Pretraining frontier models from scratch is unaffordable for Japanese players

Existing alternatives

  • OpenAI via SoftBank JV SB OAI Japan ('Crystal intelligence'), Google Gemini, other US labs sold directly
  • Japanese domestic LLMs: NTT tsuzumi, PFN PLaMo, NEC cotomi, Fujitsu Takane, SB Intuitions Sarashina
  • In-house bank IT teams and large SIers/consultancies building on hyperscaler APIs

Unique value proposition

A Tokyo frontier-research lab that turns nature-inspired, model-agnostic AI into production agents for Japan's most regulated institutions, without pretraining giant models itself.

High-level concept

“A Palantir-style applied AI lab for Japan's banks and government”

Solution

  • Japan-tuned post-trained models (Namazu) and an orchestration model (Fugu) that routes across many LLMs via one API
  • Multi-agent applications co-built with clients: AI Lending Expert (MUFG), proposal generator (SMBC), wealth-management AI (Daiwa)
  • Research-driven efficiency (evolutionary model merge, AB-MCTS, AI Scientist) instead of brute-force scale

Channels

  • Direct multi-year partnerships with enterprises and government, delivered by the Applied Team and forward-deployed engineers
  • Strategic investors as distribution: MUFG, Citi, Google, Salesforce Ventures, Mitsubishi Electric
  • SI and trading-house alliance with SCSK and Sumitomo Corp. (about 900 consolidated companies)
  • Self-serve API console and subscriptions (Fugu, Namazu, Marlin); free Sakana Chat for awareness

Revenue streams

  • Enterprise co-development and deployment contracts (values not disclosed)
  • Government commissioned research and demonstration contracts (MoD, ATLA, MIC)
  • API usage and subscription tiers for Fugu and Namazu; Marlin pay-per-use plus Pro/Team/Enterprise plans
  • Total revenue not disclosed; one unverified third-party estimate is about US$30M for 2025

Cost structure

  • Research and engineering talent in Tokyo (CEO said headcount would double in 2026)
  • Inference costs, including third-party frontier model APIs routed by Fugu
  • GPU compute for post-training and research (partly government-supported via GENIAC)
  • Forward-deployed delivery teams for bespoke enterprise and government projects

Key metrics

  • Series B ¥32B (~US$200M); post-money ~¥432B (~US$2.7B), company figures as of 2026-04-09
  • Cumulative funding ~¥66B (~US$412M), company figure as of 2026-04-09
  • Marlin closed beta ~300 professionals (Apr-Jun 2026); Fugu beta close to 500 users (Jun 2026)
  • Daiwa project in production development from 2026-08-01; MUFG live-case verification since 2026-03
  • Headcount not disclosed (~157 third-party estimate, Jun 2026)

Unfair advantage

  • Research brand and talent density: Transformer co-author as CTO, Schmidhuber as advisor, AI Scientist published in Nature (2026)
  • Embedded with MUFG, SMBC and Daiwa: rare references in a conservative sector
  • Trusted-vendor position with Japan's Ministry of Defense and backing from In-Q-Tel

Our take

Take the meeting: Sakana is the strongest bet on Japan's sovereign-AI integrator, with bank projects moving into production and repeat defense contracts. At ~¥432B post-money with undisclosed revenue, diligence must centre on revenue quality and margins.

Biggest risk
Squeeze from both sides: OpenAI/SoftBank's Crystal intelligence and Google sell directly into the same banks, while Fugu's margins depend on third-party model APIs. Sakana could end up a services-heavy integrator that cannot justify its valuation.
Why it could win
Its model-agnostic stack benefits from frontier progress without pretraining costs. Meanwhile, multi-year bank and defense deployments build switching costs and Japanese references that no foreign lab has.
What to watch
  • Any disclosure of revenue/ARR, or IPO filing (careers page lists IPO preparation)
  • Decision on nationwide rollout of MUFG's AI Lending Expert
  • Paid adoption of Fugu and Marlin outside Japan
  • First results from the SCSK x Sumitomo alliance in manufacturing and cybersecurity
  • Independent benchmarks of Fugu against frontier models

Research notes by our members, not investment advice.

Sources (10)
  1. sakana.ai/series-b/
  2. techcrunch.com/2025/11/17/sakana-ai-raises-135m-series-b-at-a-2-65b-valuation-to-continue-building-ai-models-for-japan
  3. sakana.ai/mufg-ai-lending/
  4. sakana.ai/daiwa-shoken-full-scale/
  5. sakana.ai/defense-integrated-analysis/
  6. sakana.ai/scsk-sc-partnership/
  7. sakana.ai/fugu-release/
  8. sakana.ai/marlin-release/
  9. sakana.ai/namazu-api/
  10. research.contrary.com/report/sakana-ai

Write-ups about Sakana AI