Paris-based Arlequin AI has raised €28 million in a Series A funding to develop proprietary AI models based on ‘topological neural networks’, or TNNs, and expand deployments with governments and large organisations.
The round was co-led by redalpine and OTB Ventures, with participation from Bpifrance’s Defence Innovation Fund. Existing investors Vsquared Ventures and 10x Founders increased their stakes, while French billionaire Xavier Niel also joined the round.
Founded in 2024 by Hugo Micheron, a terrorism and geopolitical instability researcher, and former CNRS research engineer Antoine Jardin, Arlequin is taking a different route from the transformer-based large language models currently dominating AI investment.
“AI is of course dominated by LLMs that are so impressive at generating text, summarising and increasingly doing complex tasks,” Micheron told the Pathfounders podcast in an exclusive interview. “But they are built on semantics. And there are some critical aspects they are not designed for, such as massive, highly complex datasets where you need to prove and provide the evidence.”
Arlequin’s approach is to use the ‘topological neural network’, an architecture designed to learn from the connections and structures within data rather than primarily predicting sequences of words.
Jardin said the aim isn’t to ‘kill off’ LLMs, but to add another tool to the AI stack.
“TNNs are not here to replace LLMs, but to complement the toolbox that we have in AI,” he said. “They provide the capability to do things that LLMs cannot, even by scaling larger.”
The company says it has built a platform capable of ingesting information including documents, transactions, video and operational data and identifying connections between people, events and other entities. It says the technology is already being used by governments and large organisations across four European countries.
One obvious application is intelligence analysis.
Micheron gives the example of money laundering: individual transactions may look legitimate, but viewed as a network they can reveal somebody effectively moving money back to themselves through multiple proxies.
“All transactions might be legal,” he said. “But the scheme, which is one person actually sending to himself through a network of other proxies, [is] money laundering. The connectivity is the product.”
That approach is potentially applicable across areas such as counterterrorism, defence, criminal investigations and fraud. These are areas where LLMs can sometimes too confidently generate a plausible answer, but can’t show where the answer came from, but a TNN approach could.
Jardin said TNNs allow users to follow the reasoning behind an output back to the underlying data.
“You have the capability to check how each of the results is produced and to trace back to the raw data in a way that you can formally describe and formally verify,” he said. “You can control what is provided as an output to the user.”
That focus grew partly out of Micheron’s previous career. Before starting Arlequin, he researched terrorism and spent extensive periods in the Middle East, including covering the Syrian war and interviewing convicted terrorists.
“It’s because of that experience that I actually realised that we needed a very powerful tool that I could not find on the market,” he said, particularly one capable of finding evidence and tracing findings back to their original sources.
The company also argues that TNNs could perform some complex analytical tasks using substantially less compute. Micheron said Arlequin’s approach is partly about being able to “do a lot with less”.
“I think [that] is key in the world of AI at the moment: how to be efficient with the computing power you have,” he said.
That could give the architecture particular relevance in Europe, where access to compute, advanced semiconductors and energy has fallen short of what is possible in the US.
The field is also an early one.
“The world of TNN is, to be honest, pretty much what the world of LLMs was 10 years ago,” Micheron said. “You have a few worldwide specialists.”
Arlequin says it has already built its first models and believes the architecture can scale. Jardin argues that some of the underlying ideas were largely pushed aside during the transformer boom and have yet to be properly commercialised.
“Many of these have been overlooked over the last five to ten years because of the birth of language models that sucked the air out of the room,” he said. “Some of these ideas are ready to be implemented, but have never been built into a product and never scaled properly.”
The company employs around 50 people, including 35 engineers and 15 PhDs, researchers and data scientists. Alongside its Paris headquarters, it has opened offices in London and Berlin, and plans to establish an AI research lab in Silicon Valley by the end of 2026.