Sakana AI’s $100M at $2.5B: Japan wants a local model shop that ships, not just demos

Reports say Sakana AI is raising $100M at a $2.5B valuation. In Silicon Valley that’s Tuesday. In Japan it’s a statement: regulated enterprises want a domestic lab that can deliver models with latency, privacy, and provenance guarantees—without punting their data across the Pacific.

What investors think they’re buying

Scarcity premium: There aren’t many credible non-US/non-China frontier labs. If you can train at scale and speak the language—literally and regulator-wise—you get paid.
Enterprise over hype: The money is in predictable inference: on-prem or VPC, retrieval bolted on properly, guardrails that pass audits, and a support team that answers the phone.
Access to hardware: Cloud partnerships and national programmes blunt the GPU shortage if you have the right agreements and workloads.

The only metrics that matter in the next 12 months

Unit economics at p95 latency: Cost per 1K tokens with retrieval and safety on. If they can quote it and still close, the pipeline is real.
Renewals over pilots: Seat expansions in banks, manufacturers, and insurers beat every leaderboard brag.
Domain agents that close tickets: Contact centre, underwriting, factory QA—if they don’t have at least one of these working at customer scale, the margin story collapses.