ARLEQUIN AI raises €28 million to build a European alternative to Palantir

Arlequin AI has raised €28 million from Redalpine, OTB Ventures and Bpifrance’s Defence Innovation Fund, just over a year after securing a €4.4 million seed round. Founded by political science researcher Hugo Micheron and former CNRS research engineer Antoine Jardin, the Paris-based startup is developing topological neural networks designed to uncover hidden relationships across large volumes of data. Behind the promise of more traceable AI with lower compute requirements, the company must now demonstrate that its proprietary architecture can outperform the graph analytics and decision intelligence platforms already deployed by governments and large organisations.

Before seeking to analyse millions of data points, Hugo Micheron began with eighty interviews.

Between 2014 and 2019, the political scientist met in French prisons with members of the Islamic State or al-Qaeda who had returned from Syria. He also interviewed their families, relatives, residents of their neighbourhoods, community representatives, Kurdish fighters and refugees. His research took him from France and Belgium to Turkey, Lebanon and Iraq. His task was not simply to collect testimony, but to reconstruct trajectories, connect territories, place individual commitments within a broader history and understand how dispersed groups eventually formed a system.

The work resulted in a PhD completed at ENS-PSL in 2019, followed by the publication of Le Jihadisme français. Quartiers, Syrie, prisons (French Jihadism: Neighbourhoods, Syria, Prisons). After working in Princeton’s Department of Near Eastern Studies between 2020 and 2022, Hugo Micheron continued his research at Sciences Po, where he leads a seminar on artificial intelligence, democracy and the information environment. He was also called to testify at the trial over the November 2015 Paris attacks.

In 2024, the researcher changed both scale and tools. Together with Antoine Jardin, then a CNRS research engineer specialising in data science and human behaviour, he founded Arlequin AI. Their premise is that the difficulty encountered when investigating jihadist networks appears in many other environments: the data exists, but it is scattered across documents, transactions, images, communications, video and operational systems that do not always speak the same language, either literally or computationally.

The startup wants to automate part of this relationship-mapping work while preserving the analyst’s ability to return to the evidence supporting each result. Two years later, it has raised €28 million to try to turn that method into an AI architecture.

A Series A to move up the technology stack

The round was co-led by Swiss fund Redalpine and Poland’s OTB Ventures. Bpifrance’s Defence Innovation Fund also joined the company’s shareholders, while existing investors Vsquared Ventures and 10x Founders increased their stakes. Xavier Niel participated in the round as well.

The syndicate reflects the project’s different dimensions. Redalpine invests at the intersection of software and science. OTB Ventures targets AI and novel computing architectures, among other fields. The Defence Innovation Fund places Arlequin among the dual-use technologies likely to interest security and military organisations.

Arlequin raised €4.4 million in June 2025 from Vsquared Ventures, 10x Founders, Kima Ventures, Better Angle and several individual investors. That initial round was intended to industrialise HuDEx, its data analysis platform, strengthen on-premise deployments and double a still-small research team. At the time, the company employed around twenty people and said it was working with government ministries, Radio France, BNP Paribas and several European organisations.

The Series A is more than six times larger, bringing the total amount announced since the company was founded to at least €32.4 million. More importantly, it is no longer financing exactly the same project.

In 2025, Arlequin described HuDEx as a modular platform combining several specialised components, including open-source technologies, with unsupervised learning methods. The company now intends to develop its own models based on topological neural networks. It is therefore moving from assembling and industrialising an analytics platform to building a proprietary model layer.

The shift is strategic. Owning the interface gives the company control over the user experience. Owning the architecture that analyses the data may create more defensible intellectual property, but it also requires longer research cycles, credible benchmarks and scarce scientific talent.

Arlequin now reports a team of around thirty, including eight people in R&D, according to its job postings. The new investment will be used to expand that team, train the models and support commercial deployments. The company has opened offices in London and Berlin and plans to establish a laboratory in Silicon Valley.

From pairwise connections to complex systems

The language of topology can quickly make a company presentation sound like a mathematics seminar. The underlying idea is nevertheless accessible.

A large language model primarily learns patterns within a sequence of words, images or other numerical units. A graph neural network operates on entities connected by edges: a person owns an account, an account makes a transaction, one company shares an address with another. Topological models aim to go further by representing simultaneous relationships between multiple elements through groups, cycles, hierarchies and other so-called higher-order interactions.

In a money-laundering investigation, the objective is no longer simply to establish that account A transferred funds to account B. The system may seek to identify a configuration involving several accounts, a company, a device, an address, an intermediary, a sequence of transactions and a particular chronology. In a supply chain, it may connect a supplier, a port, a vessel, a geopolitical incident and several delays to reveal a dependency that remains invisible when each event is examined separately.

This is precisely the kind of problem that gives coherence to Hugo Micheron’s background. His work did not examine a collection of individuals in isolation, but the relationships between militants, neighbourhoods, prisons, conflict zones and mechanisms for spreading ideology. Arlequin is clearly not turning a political science thesis into software. Its objective is to understand a phenomenon through the relationships that structure it.

Antoine Jardin must turn that approach into a computable architecture. The challenge begins with transforming raw data into relevant topological structures. Which elements should be connected? At what threshold does a relationship become significant? How can a real group be distinguished from accidental proximity? Models necessarily inherit the choices made at this stage.

Arlequin has begun to make this scientific work visible. Antoine Jardin and team member Rémi Devaux are among the authors of a paper published in August 2026 with the Geometric Intelligence Lab. The researchers present TopoExplorer, a tool designed to visualise how a dataset is transformed into a topological structure before a model is trained. Across twenty benchmark datasets, they found that some indicators available before training correlate with subsequent model performance.

The research is consistent with Arlequin’s promise of interpretability: before examining what the model produces, users should be able to inspect the map across which the information will travel. It does not, however, validate the performance of the proprietary models the startup now intends to develop.

The real competitor is not ChatGPT

Arlequin presents its approach as a break with the race to build ever-larger models. Antoine Jardin argues that the next advances in AI will come from new architectures rather than continuously increasing parameter counts, training data and computing power. Hugo Micheron, meanwhile, describes the emergence of systems capable of understanding complex dynamics hidden within millions of data points.

The distinction from LLMs helps position the project, but it does not define its market. Arlequin is not primarily seeking to replace ChatGPT, Claude or Le Chat. Its real competitors are the platforms that already connect fragmented data to support high-stakes decisions.

Palantir has built this category over the past two decades. Its architecture rests on an “ontology” connecting an organisation’s data to its real-world counterparts: equipment, products, orders, transactions, people and events. This representation then becomes an operational layer through which users can analyse a situation, apply rules or trigger an action.

Arlequin embraces the comparison. In its recruitment material, it describes itself as “the sovereign European alternative to Palantir”. The ambition is considerable. The US group does not sell an isolated model. It provides data integration, governance, access controls, interfaces, analytical tools and the support required to embed them in the day-to-day operations of a government agency or large company.

Other companies already occupy this market. Britain’s Quantexa combines data ingestion, entity resolution and graph analytics for fraud prevention, anti-money laundering, customer intelligence and supply-chain analysis. France’s Linkurious, acquired by Australia’s Nuix, enables investigators to visualise complex networks and was notably used in the Panama Papers investigation. Around them, companies including Primer, Babel Street and Dataminr address different parts of the chain, from open-source intelligence to emerging-signal detection.

Defence provides strategic validation, but not yet commercial proof

The arrival of the Defence Innovation Fund adds another dimension to the round. Created by France’s Ministry of the Armed Forces and managed by Bpifrance, the vehicle invests in technologies considered capable of strengthening French capabilities and technological sovereignty. Initially endowed with €200 million, it grew to €275 million in 2025 following commitments from Caisse des Dépôts, Allianz France and MBDA.

Security, intelligence and defence are natural markets for a system capable of connecting documents, transactions, video and operational data. The same engine can also be used to examine a criminal network, detect an information-manipulation campaign, investigate fraud or analyse a cyberattack. The technology’s dual-use nature broadens its market while bringing Arlequin closer to the most demanding requirements for confidentiality and control.

Bpifrance’s investment does not mean that the startup has won a contract from the Ministry of the Armed Forces. It constitutes neither a security certification nor operational validation. It does indicate that the technology is considered strategically relevant enough to justify a public equity investment.

Arlequin says its platform is being used by governments and large organisations across Western and Eastern Europe, but does not disclose its revenue or the number of paid contracts. The customer references made public in 2025, including Radio France and BNP Paribas, show that the product has moved beyond the laboratory. They do not yet reveal the depth of adoption.

These are precisely the kinds of organisations in which pilot programmes can last for a long time. Several years may separate a demonstration from a pilot, a first contract and the integration of a platform into operational decision-making. A €28 million Series A gives Arlequin more time to clear those stages.

Traceable does not mean infallible

Arlequin emphasises the ability to connect each result to the information supporting it. This traceability addresses a well-known weakness of generative models: an answer can be delivered with confidence while leaving the user unable to identify the evidence on which it rests.

For a bank, a public authority or a security service, the distinction is essential. An analyst must be able to open the document, retrieve the transaction, verify the date and understand the path taken by the system. Auditability is not an optional convenience. It determines whether the machine can be challenged.

It does not solve every problem. An unsupervised system is not necessarily free from bias. Bias can enter through the selection of sources, missing data, the definition of entities, chosen thresholds or the way relationships are constructed. A non-generative architecture may reduce the risk of inventing a sentence or reference. It may still identify a misleading configuration, place too much weight on a coincidence or overlook an event poorly represented in the data.

The same caution applies to causality. Retrieving the evidence supporting a relationship improves documentary traceability. It does not necessarily demonstrate that one event caused another. Two transactions occurring close together, two people visiting the same place or several pieces of content promoting the same narrative may be correlated without being causally connected.

In software advertising, the distinction may appear almost pedantic. In a criminal investigation or defence decision, it becomes central. A clear visualisation can help correct an error. It can also give a fragile hypothesis the appearance of evidence.

European sovereignty is also being organised from Silicon Valley

International expansion introduces one final tension. Arlequin is opening offices in London and Berlin to move closer to European customers, talent and security ecosystems. At the same time, it plans to establish a laboratory in Silicon Valley.

The move does not necessarily contradict its sovereignty narrative. By opening a Californian laboratory, Arlequin is seeking access to the world’s leading pool of AI researchers and partners. It must simultaneously demonstrate that it can keep in Europe the assets that make its technology strategic. Modern sovereignty is not autarky. It is control over the dependencies one chooses to accept.

The Series A gives Arlequin the means to push this ambition beyond a statement of intent. Its history, positioning and the participation of the Defence Innovation Fund place it at the intersection of several European priorities: building architectures distinct from LLMs, processing sensitive data on the continent and offering an alternative to US platforms in critical environments.