Arlequin AI raised a €28 million Series A on September 10, 2026, to build proprietary models on topological neural networks and push its investigation software beyond France. redalpine and OTB Ventures co-led a round the company said is backed only by European investors. Bpifrance’s Defence Innovation Fund joined. Existing backers Vsquared Ventures and 10x Founders added to their stakes. Xavier Niel and ZEBOX, CMA CGM’s startup accelerator, came in as new investors.

Hugo Micheron, a specialist in jihadism and geopolitical instability, and Antoine Jardin, a former CNRS research engineer, founded the Paris company in 2024. They sell an analytics platform (redalpine names it HuDEx) that links people, places, communications, and events across documents, transactions, video, and operational files. The new money is for the next step: models that learn the structure of those links, including relationships among more than two entities at once, and that can still point a user back to the underlying evidence.

The company says that approach is meant for work where a guess is not enough: counterterrorism, criminal and judicial investigations, fraud and money laundering, information integrity, cybersecurity, and AI system safety. It says the software is already in use with governments and large organizations in Western and Eastern Europe, across four countries. Research partners named in the post include INRIA, CNRS, the Max Planck Institute, and teams at Oxford, Cornell, Princeton, and UC Santa Barbara.

Jardin argued that the next gains will come from different architectures, not from larger transformers and more energy. The company said TNN models are being designed to need less compute than large language models. It has opened offices in London and Berlin and said it will add a Silicon Valley lab later in 2026.

“AI sovereignty is not only about where models are built or data is hosted. It is about having control over the technologies that increasingly underpin our most critical decisions.” Jeremy Teboul, OTB Ventures

Decoded Take

European Series A checks that fund a new architecture, rather than another frontier LLM, are still rare. Arlequin is selling explainable graph work to ministries that cannot put a hallucinated edge in a court file, which is a real buyer. The risk is that “topological neural networks” stays a research slogan while HuDEx remains a well-funded investigation tool. Four unnamed government deployments are not the same as a public tender or a published accuracy number against an LLM baseline. Watch whether the Silicon Valley lab ships a named TNN checkpoint, and whether a second country beyond the current four puts the platform on a defense or justice contract before the next round.