Sakana AI Partners With DEEP DIVE to Power Japan's Satellite Intelligence

Tokyo-based Sakana AI has signed a partnership with DEEP DIVE, a private intelligence organization, to apply its models to satellite imagery and open-source intelligence analysis. The deal is the clearest sign yet that Sakana, best known for its evolutionary model-merging research, is moving aggressively into defense work alongside its existing finance practice.

The two sides of the deal

DEEP DIVE is not a typical defense contractor. It was founded as a non-profit general incorporated association by Bonji Ohara, a senior fellow at the Sasakawa Peace Foundation, and Yu Koizumi, an associate professor at the University of Tokyo's Research Center for Advanced Science and Technology. Its mission is to deliver early warnings to Japanese society using digital open-source intelligence (OSINT) and satellite imagery, areas historically reserved for state intelligence services.

Sakana brings the models. Under the agreement, DEEP DIVE's analyst expertise and its trove of satellite and open-source data will be combined with Sakana's proprietary AI to perform information analysis at a scale, speed, and resolution that manual workflows cannot match. The two sides plan ongoing joint research to refine analytical methods and push them into operational use.

Why a research lab is chasing satellite imagery

Sakana made its name on efficiency tricks rather than scale. Its research focuses on techniques like Evolutionary Model Merge, which fuses capabilities from different open-source models, and tree search algorithms that orchestrate collaboration between closed models. In plain terms, instead of training a new foundation model from scratch, Sakana evolves and combines existing ones, hunting for combinations that perform better than any parent.

That approach maps cleanly onto intelligence work. OSINT and geospatial analysis tend to involve stitching together specialized vision, language, and reasoning capabilities, which is exactly the kind of problem a model-merging pipeline is built for. Plugging DEEP DIVE's labeled satellite imagery and analyst notes into that pipeline gives Sakana a real-world testbed for monitoring activity around Japan at a cadence that two founders and a small staff could never match alone.

The bigger strategy

The DEEP DIVE deal lines up with a broader pivot Sakana telegraphed six months ago. When the company closed its Series B, it said it would strengthen its applied team to deepen partnerships in finance and expand into defense and manufacturing. That round brought in roughly $135 million at a $2.65 billion post-money valuation, making Sakana one of Japan's most valuable AI startups.

The round's investor list also hinted at where this was going. Participation from In-Q-Tel, the venture arm tied to the U.S. intelligence community, signaled the strategic importance of Sakana's technology for national security applications in Japan and potentially for its allies. A few months later, Sakana secured a multi-year research contract with ATLA, Japan's Acquisition, Technology and Logistics Agency, to build foundational AI for analyzing multi-domain data and supporting command and control. The DEEP DIVE partnership rounds out the picture: government work on one side, civilian OSINT on the other, both feeding the same applied research team.

Industry context

Japan has historically lagged the U.S., China, and Israel in defense AI, and most of its intelligence capacity sits inside government. A private OSINT shop tying up with the country's highest-profile AI lab is unusual, and it reflects two pressures playing out at once. Geopolitical risk in the region keeps climbing, and Japan's shrinking workforce makes scaling any analytical operation by hiring more humans a non-starter.

Sakana also has a model that fits a country with limited compute budgets. Rather than training frontier-scale LLMs, it recycles and recombines existing open-source ones, which keeps GPU bills closer to a research lab's than to OpenAI's. For an OSINT outfit run by two people that needs to process satellite images continuously, that economics matters as much as model quality.

What it means in practice

No financial terms were disclosed, and the announcement is framed as joint research rather than a product launch. Concretely, expect Sakana's vision and language models to be tuned on DEEP DIVE's reports and imagery, with outputs flowing back into DEEP DIVE's public analysis. For anyone building in geospatial AI, OSINT tooling, or sovereign AI for non-U.S. markets, this is a useful data point on how a research-first lab translates model-merging work into a defense-adjacent product line without becoming a traditional defense contractor.