A Paris-based startup that argues the future of artificial intelligence lies in topology rather than scale has secured €28 million in Series A funding. Arlequin AI, founded in 2024, is building what it calls topological neural networks, architectures designed to learn not only from individual data points but from the relationships between them. The round was co-led by the Swiss-German venture firm redalpine and Poland's OTB Ventures, with the Bpifrance Defence Innovation Fund taking part. Existing backers Vsquared Ventures and 10x Founders doubled down, and the telecoms billionaire Xavier Niel added his name to the cap table.

An unusual founding pair

The company's origins are not the typical computer-science spinout. Hugo Micheron, the chief executive, built a career studying terrorism and geopolitical instability. Antoine Jardin, the chief technology officer, spent years as a research engineer at the French national research centre CNRS, specialising in data science and human behaviour. Their collaboration began from a shared frustration: organisations making high-stakes decisions, whether in security, fraud detection or cyber defence, were drowning in fragmented data but lacked tools that could map the connections across documents, transactions, video feeds and operational logs.

That problem set the direction. Arlequin's platform is already deployed with governments and large organisations across Western and Eastern Europe, where it is used to trace links in criminal investigations, money-laundering networks and information-integrity operations. The new capital is earmarked for turning the platform's underlying approach into proprietary models that can be trained and sold more broadly.

What topological neural networks actually do

Conventional deep learning treats data as a flat collection of features. A transformer model, for instance, learns statistical associations between tokens but has no native representation of the structure that binds them. Topological methods, by contrast, borrow from algebraic topology, the branch of mathematics that studies properties preserved under continuous deformation, to represent data as shapes and connections. In Arlequin's implementation, the network learns a topological signature of the dataset: which clusters are linked, where holes appear, how components persist across scales. The claim is that this lets the model capture multi-way interactions, three, four or more entities acting in concert, without the combinatorial explosion that forces standard models to grow exponentially.

Jardin put it bluntly: the company believes further AI advances will require different architectures rather than simply larger models trained on more data and computing power. If the claim holds, the implications for the semiconductor supply chain and energy demand are significant. Training a frontier large language model today can consume tens of thousands of GPUs and megawatts of electricity. A topology-first approach that reaches comparable performance on structured, relational tasks with a fraction of that hardware would upend the economics of AI deployment.

The compute question

Arlequin has not yet published benchmarks comparing its models head-to-head with transformer baselines on standard tasks. The company says its proprietary models are now in development, with the Series A funding the expanded team needed to train and evaluate them. Sceptics will note that topological data analysis has promised computational efficiency for years without displacing gradient-based deep learning in production. The difference this time, the founders argue, is the integration of topological layers directly into the neural architecture, making the topology learnable end-to-end rather than a pre-processing step.

The Bpifrance Defence Innovation Fund's participation signals where the first paying customers are likely to come from. Defence and intelligence agencies sit on vast, heterogeneous datasets, satellite imagery, signals intercepts, open-source reports, human intelligence, where the relationships between entities matter more than any single data point. A system that can surface a hidden network of shell companies, or trace the evolution of a disinformation campaign across platforms, without requiring a dedicated data centre, has obvious appeal.

European footprint, American ambition

The company has already opened offices in London and Berlin, positioning itself on both sides of the Channel and at the heart of the EU's largest economy. A Silicon Valley AI lab is slated for the coming months, a move that reflects the reality that US defence and intelligence budgets still set the pace for advanced analytics procurement. Niel's involvement adds a French tech heavyweight to the investor base; his Iliad group has been building its own GPU cloud through the subsidiary Scaleway, and he has publicly backed Mistral AI's push for European sovereign models.

The scaling debate in context

The funding arrives at a moment when the dominant paradigm, bigger models, more data, more compute, is facing scrutiny on multiple fronts. The European Commission's AI Act, which entered force in August 2026, imposes transparency and risk-management obligations that scale with model capability. Energy costs in Germany and France remain elevated after the gas crisis, making power-hungry training runs a board-level concern. And the US CHIPS Act restrictions on advanced semiconductor exports to China have reminded every non-American buyer that GPU supply is political. Arlequin's pitch is that a different mathematical foundation sidesteps several of these constraints at once.

Micheron framed the moment in sweeping terms: "Today, another revolution is taking shape: the development of new AI systems capable of understanding highly complex dynamics hidden within millions of data points." Whether that revolution materialises in a product that outperforms a fine-tuned Llama or GPT variant on a fraud-detection benchmark remains to be proven. The Series A gives them roughly 18 to 24 months of runway to show it.

People mentioned

  • Hugo Micheron

    Chief executive and co-founder, Arlequin AI

  • Antoine Jardin

    Chief technology officer and co-founder, Arlequin AI

  • Xavier Niel

    Investor, Iliad

Organisations

Arlequin AI · redalpine · OTB Ventures · Bpifrance Defence Innovation Fund · CNRS · Vsquared Ventures