Arlequin AI Takes Defence Money to Map Hidden Networks

Paris AI firm Arlequin AI closed a €28 million Series A on September 10 to scale topological neural networks for investigations that cannot tolerate a made-up answer. Redalpine and OTB Ventures co-led the round, which the company values at about $32 million, and every cheque came from Europe. Bpifrance’s Defence Innovation Fund joined existing backers Vsquared Ventures and 10x Founders, while Xavier Niel and ZEBOX, the CMA CGM accelerator, came in as new names.

The architecture story is the one that travels. The quieter fact is who is paying, and what kind of work already pays the bills: governments and companies that need a map of links they can take back to the file, not a fluent paragraph.

France’s Defence Fund Took a Seat at the Table

The Series A announcement on its site lists the usual venture firms first. The fund that does not usually sit in a consumer-AI round is the one that tells you the customer. Bpifrance created the Defence Innovation Fund in 2021 to take minority stakes, always as a co-investor, in companies whose tools interest the defence sector even when the first market is civilian.

Its own mandate names intelligence artificielle and analyse de données among the fields it will back, and it is written for dual-use civil and defence applications. Cheques run from €0.5 million to €30 million. Most of the book is aimed at later growth rounds; a Series A is the smaller door, used when the first build is expensive. Arlequin walked through that door with a product already in use, which is the proof the fund asks for before it treats a civil tool as defence-relevant.

Oliver Pabst, general partner at redalpine, called the company a “fully productized sovereign tech platform for safety and security in Europe and beyond.” Jeremy Teboul of OTB Ventures put the same idea in plainer words: Europe, he said, needs AI systems it can own, understand and trust, and sovereignty is not only a question of where a model is built or where the disks sit. The fund’s presence makes that sentence operational. A Paris startup selling audit trails to ministries is no longer only a private software story once the state’s defence-innovation vehicle is on the cap table.

Hugo Micheron Mapped Jihadist Networks by Hand

Hugo Micheron, the chief executive, did not come out of a lab that trains language models. He took a political-science doctorate at the École normale supérieure in 2019, then spent September 2020 to September 2022 as a postdoctoral researcher at Princeton’s Institute for Transregional Studies. The fieldwork behind the thesis included neighbourhood studies in France and Belgium and 80 interviews with convicted returnees from ISIS and Jabhat al-Nusra in French prisons, plus interviews in Iraq, northern Lebanon and Turkey.

He later appeared as an expert witness at the trial of the 13 November attacks, on the stand for more than five hours. La Colère et l’Oubli won the Prix Femina essai and the prix du livre de géopolitique in 2023. He has said the gap that became the company opened while he was at Princeton, after the murder of Samuel Paty, when he concluded that a large share of the networks he studied now lived online and that the tools on offer were “full of biases.”

Antoine Jardin, the chief technology officer and a friend of fifteen years, spent a decade as a CNRS research engineer with a doctorate in statistics and machine-learning algorithms. He worked on dimensionality reduction, contributed to France’s national AI strategy, and had a hand in projects tied to the Jean Zay supercomputer and the World Bank. Micheron has said they wanted an architecture with “the precision of scientific expectation,” able to prove and audit its conclusions rather than guess the next word.

You have an investigator looking at a blackboard, and they have pinned a picture, maybe headlines from a newspaper, a phone number, some people, and they are making connections between them with visual links. That’s their investigation. At some point, by connecting all these points, they’re like, ‘Oh, the suspect is that guy.’ This is exactly how a TNN works. It does that connectivity piece.

Hugo Micheron, CEO and co-founder, Arlequin AI

He has used a money-laundering sketch for the same idea. Each transfer can look legal on its own. The scheme appears only when one person is sending funds back to himself through a web of proxies. “The connectivity is the product,” he said.

Topological Models Track Many-Way Links at Once

Large language models learn to continue text. Graph networks learn pairwise links, a line between two nodes. Topological neural networks, as the academic literature now uses the term, add the missing middle: groups of three or more that interact at once, treated as faces or cells rather than as a pile of separate edges. A 2024 paper on the subject describes models that pass messages through those higher-order interactions that graphs miss, using simplicial and cell complexes so the computational graph is no longer chained to the original pairwise map.

WHAT A TOPOLOGICAL NETWORK ADDS

• The pair: A graph model treats a relationship as a line between two points, which is enough for a simple contact list and weak for a cell of people who only make sense as a group.
• The group: A topological model can treat three or more nodes as a single face, so a joint pattern can be learned without being flattened into separate two-way links.
• The audit path: Arlequin’s pitch is that every finding can be walked back to the underlying document, transaction, clip or log, with no context window and no invented prose.

That last point is the company’s claim, not a published score. Its site says it is building unsupervised and topological AI models for large-scale analysis, and it markets the working platform as HuDex. The product page says client data is not stored and is not used to train models, and that the stack can be put on the customer’s own machines. Jardin has said further gains will have to come from different architectures, not from larger models fed more data and more chips. The weights Arlequin actually ships are proprietary. The papers describe a family of methods; they do not certify this product.

What the Platform Is Already Doing in Europe

Micheron said revenue is about 50 percent public and 50 percent private, spread across France, Germany, the United Kingdom and an unnamed Eastern European country. Coverage of the round puts the live book at around 30 clients and the payroll at around 50 people, about two-thirds of them researchers and engineers. Offices are open in Paris, London and Berlin. The company says it is already working with research teams at Oxford, Cornell, Princeton, CNRS and Inria.

ARLEQUIN AI AT THE CLOSE

• The split: About half of revenue comes from public bodies and half from private customers, Micheron said.
• The map: Live work sits in four European countries, with one Eastern European buyer left unnamed.
• The bench: Around 50 staff, most of them researchers and engineers, with a Silicon Valley lab planned in the coming months.
• The use case: In a counterterrorism file, the company says the platform can take millions of items from several seized devices and draw links among people, places, communications and events.

The same method is being sold into fraud and money-laundering desks, criminal investigations, cybersecurity, information integrity, and what the company calls AI safety. Banks and media groups were in the early mix; ministries are the reference customer the new investors want to clone. Micheron said a drop of data into the platform is meant to return causality links and patterns in minutes, work that would take researchers months if they did it by hand, and that the point is to scan the whole haystack rather than spend a fortune of compute hoping to find one needle.

WHERE THE COMPANY SAYS THE MAP APPLIES

• Security and defence: Networked threats, including extremist graphs and, in a defence reading, sensor and order-of-battle relationships.
• Financial crime: Flows that look ordinary one transfer at a time and illegal once the proxies are drawn as a single scheme.
• Forensics and comms: Fragmented messages, devices and logs reconstructed as an event, not as a summary.
• Cyber and integrity: Signals across systems, plus the information-operations work the seed round was sold on.

None of those buyers is named in the Series A materials, which is normal for this kind of contract and inconvenient for anyone trying to test the accuracy claim from the outside.

From a €4.4 Million Seed to a Defence Round

Fifteen months ago the same company was raising a much smaller cheque on a different poster. On June 10, 2025, Arlequin closed a €4.4 million seed led by Vsquared Ventures, with 10x Founders, Kima Ventures, Better Angle, and angels including former Meta executives Julien Codorniou and Julien Lesaicherre. The public line then was disinformation and “certified” insight. Micheron said Europe had to step out of the shadow of American and Chinese giants, and that Arlequin was “not building yet another LLM” but comprehension models that “require no training data” and use far less resource.

THE TWO CHEQUES

Kima, Xavier Niel’s early-stage vehicle, sat on the seed. Niel came in personally on the Series A. The Defence Innovation Fund did not. That sequence is the tell: first a European sovereignty story about noise in the information space, then a larger round once governments were already paying and a defence fund could treat the same software as dual-use. Maria Juesas Portelés, a venture partner at Vsquared, had called it the kind of sovereign technology Europe needed to navigate information chaos. The new money is aimed at a harder room, where a wrong link can move a warrant.

The Hallucination-Free Pitch Still Lacks a Public Test

Arlequin’s site says it is “the only company in the world scaling unsupervised systems for large-scale data analysis,” with “unbiased, hallucination-free results,” no black-box reasoning and no context window. Those are the words a ministry wants to hear if it has already been burned by a chatbot that invents a citation. They are also words that, in work that can touch a prosecution file or a targeting decision, still need a test someone else can repeat.

No independent benchmark of HuDex sat in the round materials, and no named government customer went on the record with a miss rate. The academic TNN papers measure tasks such as distinguishing structures and locating signals on synthetic complexes; they do not measure whether this commercial stack, on seized phones or bank wires, beats a well-run graph tool plus a human analyst. Until that comparison is public, “hallucination-free” remains a product line, and the €28 million is a hypothesis with paying users rather than a published result.

Micheron has claimed the benefits could run tens or a hundred times past those of language models. That figure is his. The more careful version, the one the fund can underwrite, is narrower: some investigative work is a topology problem, language models are a poor fit for it, and Europe would rather own the alternative than rent an American one.

Engineers, GPUs and a Silicon Valley Lab

Micheron said the new capital will hire engineers, buy the compute needed to train and scale the models, and put the product in front of more customers. The company also plans an AI lab in Silicon Valley in the coming months, a California outpost for a firm that still sells European data residency and on-premise installs as a feature. That is not a contradiction so much as a recruiting plan. The models may be meant to need less silicon than a frontier chatbot; they still need people who can train them, and those people are concentrated in a few cities.

WHAT THE €28 MILLION IS FOR

• The team: Expand the international group that develops and trains the proprietary models, from a base of around 50.
• The machines: Buy compute to train and scale, even while the pitch is that the architecture should need less of it than a large language model.
• The map: Push commercial deployment across Europe and abroad, copying the public-private split that already exists in four countries.
• The lab: Open a Silicon Valley research site in the coming months, alongside the Paris, London and Berlin offices.

The governments already on the platform will learn first whether the blackboard holds when the data is ugly. The defence fund is paying for the chance that it does, and that the next investigation starts from the links rather than from a generated paragraph.