Arlequin AI Has Raised €28M to Replace LLMs for High-Stakes Decisions

Arlequin AI, a Paris based company, has raised 28 million euros in Series A funding to accelerate development of what it calls topological neural networks, an AI architecture built as a deliberate alternative to large language models for the kind of high‑stakes, consequential decisions where a confident but wrong answer can be catastrophic.

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, and French billionaire and Iliad founder Xavier Niel also joined the round. The raise follows a 4.4 million euro seed round Arlequin closed in June 2025, led by Vsquared Ventures with participation from 10x Founders, Kima Ventures, Better Angle, and a group of French business angels including former Meta executives Julien Codorniou and Julien Lesaicherre.

Arlequin was founded in 2024 by chief executive Hugo Micheron, a scholar of terrorism studies and geopolitical instability, alongside Antoine Jardin, a former CNRS research engineer specializing in data science and human behavior. Micheron's academic background sits directly behind the company's founding thesis. In a conversation with Tech Funding News, he explained why he considers large language models fundamentally unsuited to the kind of analysis Arlequin was built to perform: the problem is not merely a model's accuracy, but the consequences of it being wrong. An 80 percent confidence level might be perfectly acceptable when drafting an email, he argued, but in counterterrorism work, that same margin of error can mean the difference between prevention and an attack. He drew the same comparison to energy infrastructure, where a 90 percent confidence level in a flawed analysis could trigger an ecological crisis.

That reasoning shapes Arlequin's entire technical approach. The company's unsupervised models process raw data directly, including video, audio, text, images, and even seized devices, detecting connections between elements that can be audited and traced back to their source, rather than generating fluent but potentially fabricated text the way a large language model does. Arlequin does use a large language model within its platform, but strictly as a conversational interface, helping users formulate and refine their questions rather than performing the underlying analysis itself. "We comprehend patterns rather than generate text," the company states on its own site, describing itself as the only company in the world currently scaling unsupervised systems for large‑scale data analysis at this level.

The centerpiece of the new funding is further development of topological neural networks, or TNNs, an architecture Micheron says is designed to learn from both individual data points and the relationships between them simultaneously, allowing it to capture complex, multi‑way interactions that become increasingly difficult to interpret using conventional approaches as datasets grow larger and more interconnected. "Today, another revolution is taking shape: the development of new AI systems capable of understanding highly complex dynamics hidden within millions of data points," Micheron said.

Vsquared Ventures' Maria Juesas Portoles framed the investment as a deliberate bet against the prevailing architectural consensus dominating AI funding. "Most of the capital in AI right now is chasing the same architecture and hoping scale solves everything," she said. "Arlequin is pursuing a genuinely different architecture: developing topological neural networks as a distinct architectural approach, rather than relying solely on larger models and more compute."

Arlequin's technology is already deployed with government and institutional clients, with the company citing applications spanning security and defense, criminal investigations, fraud and money‑laundering detection, information integrity, cybersecurity, and AI safety and security. Every finding the platform surfaces is designed to trace back to the underlying data supporting it, and the company says it never stores client data or uses it to train its models, a data‑handling posture aimed directly at the sovereignty and confidentiality concerns of government and defense customers.

The market opportunity behind Bpifrance's participation specifically is substantial: Europe's AI and analytics market for defense purposes alone is projected to grow from 4.8 billion dollars this year to 19.27 billion dollars by 2035, a figure that does not even account for Arlequin's separate applications in banking fraud detection and due diligence. Micheron has claimed the potential benefits of topological neural networks could be "10s or 100 times more profound" than those of large language models, though that assessment remains his own and has not been independently verified by outside researchers.

Arlequin's positioning taps directly into Europe's broader push for sovereign, auditable AI infrastructure that does not depend on American or Chinese providers, particularly for government and defense applications where data control and explainability carry legal and security weight beyond raw model performance. Whether topological neural networks can genuinely deliver the kind of step‑change improvement in high‑stakes analytical accuracy that Micheron is promising, rather than simply offering a more interpretable but ultimately comparable alternative to existing approaches, will likely become clearer as Arlequin's government and financial‑sector deployments scale beyond their current early customer base.