Invest1 publisher3 min readPublished
Xavier Niel and Bpifrance's defence fund bet 28 million euros on topological neural networks
A 50-person Paris company with more than 30 clients has convinced Bpifrance's defence fund and Xavier Niel that the way to compete with American compute is to change the mathematics rather than match the budget.
The Investor · Invest desk
What happened
- Paris-based Arlequin AI has raised a 28 million euro Series A to scale an architecture that does not use large language models to do its analysis.
- Redalpine and OTB Ventures co-led, Bpifrance's Defence Innovation Fund took part, Vsquared Ventures and 10x Founders increased their holdings, and Xavier Niel came in as a new investor.
- The company says its platform is used by more than 30 clients across four countries in both Western and Eastern Europe.
- Arlequin employs about 50 people, of whom 35 or more are engineers and 15 are PhDs and postdocs.
- Chief executive Hugo Micheron estimates that only about 20 researchers worldwide work on topological neural networks, and says Arlequin is the first to sell a commercial product rather than a prototype.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- decision European defence buyers now have a second architecture to evaluate on the merits. Money spent proving topology, though, is money not spent closing Mistral's scale gap.
- exposure At roughly 930,000 euros of fresh capital per client already signed, the round's backing rests mainly on customers Arlequin has not won yet, not on revenue already booked.
- constraint With the analysis running on unsupervised models and the language model confined to phrasing questions, a procurement officer ends up buying a trial rather than comparing scores across vendors.
- contradiction Either Arlequin employs three quarters of the world's topological researchers, or the 20-person figure describes something narrower than its own PhD bench. Which one it is decides whether the moat is real.
Arlequin's cumulative total is now 32.4 million euros [3], and taking out the 28 million just announced leaves roughly 4.4 million raised across the first two years of the company's life [4], which makes this Series A about six and a half times all the capital that preceded it [5]. Step-ups that steep normally follow a product that started selling, and the disclosure here stops at client count and geography: no revenue, no valuation, no contract values, so the multiple measures investor conviction rather than traction.
What the conviction buys is an argument about the cost of being wrong rather than the rate of being right. Micheron's version is that 80 per cent confidence is acceptable for drafting an email and not for counterterrorism, where, as he puts it, "you end up being wrong... you have an attack" [11], and that a 90-per-cent-confidence mistake at an energy company can end in an ecological crisis [12]. The architecture follows the complaint. Unsupervised models take in raw video, audio, text, images and seized devices, and the connections they find can be traced and audited [10]; the topological neural network organises relationships in the data rather than patterns in language [13]. Antoine Jardin, a former research engineer at CNRS, brought the dimensionality reduction that sits at the centre of the models, and the two founded the company two years ago [17]; Micheron brought the fieldwork, hundreds of interviews with convicted terrorists and a Princeton lectureship from 2020 to 2023 during which he decided the available tools were "full of biases" [18].
Micheron's own numbers carry a risk worth weighing. A global research population of roughly 20, with Arlequin first to a commercial product [14], is a moat if the count is right and a warning if it is: there is no second implementation to price against, and the people who could build a rival are the same handful Arlequin needs to hire. He also places the research where large language models stood a decade ago [15], and a decade is longer than any Series A runway. Then there is his own framing of the race, with the United States ahead on compute and data, China holding values Europe will not adopt, and Mistral advancing but short of scale [16]: if language-model vendors narrow the traceability gap, the argument falls back to accuracy, which is the ground compute and data decide.
At 28 million euros against about 50 people, or some 560,000 euros a head [9], the round funds a few years of the current shape, and the German expansion [19] is where it gets spent. Renewals settle the thesis. The existing clients bought a first look, and a second European architecture only exists in procurement if they sign again.
What to watch
- Whether the German push produces named agency customers or only headcount, since only one of those tests demand.
- Whether the next round is led from outside or by Vsquared Ventures and 10x Founders, who have already increased their holdings.
- Whether Bpifrance's Defence Innovation Fund follows with a larger cheque or a procurement-linked contract, and whether terms are disclosed.