FORGIS

Out of 14,000 startups at VivaTech last week, only one was awarded for bringing AI to manufacturing.   And when the award comes from Orange, France's largest telco. €40 billion in revenue and 290 million customers worldwide. You know it's serious.   Forgis won the Orange Startup Challenge.  The award is good. The partnership behind it is the actual prize.   Building a purpose-built world model for manufacturing needs data the internet does not have, talent the AI market does not train, and years of patient compute. Then comes the part that breaks most startups: getting it into thousands of European plants with their own buying processes, security policies, integration constraints. Most industrial AI startups die before their second deployment because of this layer, not the model layer.   When the telco that runs the networks inside those plants picks the partner it brings to its industrial customers, the signal is not marketing. It is the kind of partnership European industrial AI has been missing: a channel into the factories instead of a pitch at them. Thank you to Nelly Rousseau, François Jézéquel, Emmanuel Routier, Sebastien DUDREUILH, Vincent Imbert, Benjamin SPUND and the team at Orange Business for the trust and the conviction to bet early on a European company building this.   Industrial AI is built in factories, not at conferences. France is next.

Today, every factory still runs on bespoke solutions. A different model for every robot, every machine, every line. What if one architecture could predict all of them: a bearing failure in a stamping press, a quality drift in a chemical reactor, a robot's next move? Over the past year, working alongside our customers, we kept seeing the same three patterns: → Engineering takes too long. Commissioning control for a single robot cell takes months. A full line takes years. → The data is unreadable. Decades of signals sit in historians that no model, and soon, no human, can interpret. → AI that needs failure to predict failure. Today's models can't generalize. They have to break something before they learn anything. The result is the same everywhere: fragmented systems, downtime, scrap, and engineering cycles measured in quarters, not days. We're a tech-first company, but we lived these problems firsthand. And that's where we started, with one uncomfortably big question: Can we build AI robust enough to scale to any machine and any robot on earth, understanding any signal, in any context, and cut deployment from months to days? That led us to a single direction: a world model for the factory. Instead of one narrow model per asset, a world model learns the underlying dynamics of how machines, signals, robots and processes behave, so it can reason about a press it has never seen the same way it reasons about a reactor or a robot arm. And as always, we moved at FORGIS speed. Four building blocks, now published across 5 ICML 2026 workshops - not four separate ideas, but one pipeline: 🟠 FactoryNet: the data. A large-scale dataset of industrial sensor data supporting the pretraining of the entire stack. (FMSD + AI4Physics) 🟠 HEPA: the world model architecture. A foundation model for event prediction in time series, running on the edge. (FMSD - Spotlight ⭐) 🟠 RASA: the factory graph. Shows transformers can reason over the factory as a graph, where topology, not relation weights, drives multi-hop reasoning. (GFM) 🟠 TEMPO: the language. Reads raw sensor streams and explains, in plain words, what a machine is doing. (FMSD) Sensors → HEPA → RASA → TEMPO → a plain-English answer for the operator: "inner-race fault → replace bearing, 48 h." Imagine a single model you can plug into any factory and ask to predict any event - on any machine, reasoning over the whole plant as a connected graph, collapsing deployment cost for every machine and robot inside it. Want to put it to the test on your own machines? Let's talk 🚀

We're sitting at a table with companies that have raised around $15M on average. And this table? It's HQ'd at the Google DeepMind office. 🤯 After the hackathon we organized together with Google DeepMind, we decided to take our relationship further. For us, that means setting the foundations to collaborate with one of the most prestigious research centers in the world, and scaling our foundation model research toward a new level of intelligent, physical AI for manufacturing. Yesterday, we left Zurich to attend the kick-off. And this isn't just any accelerator, it's a gathering of the highest-potential startups out there, like Generative Bionics ($70M seed), Robeauté ($28M Series A), and Automated Architecture (AUAR) ($8M seed), all backed by DeepMind's technical expertise. Over the next 3 months as part of the DeepMind accelerator, we'll advance our industrial world model, a system that understands the physical world and reasons over machine behavior. By bringing this industrial intelligence to manufacturers, FORGIS will scale its commercial deployment across Fortune 500 companies in Europe, enabling a new level of efficiency on the shop floor. And we couldn't be more excited to have Gianmaria Sbetta as Technical Success Manager. This is just the beginning. 🚀 Acumino Adapta Robotics Bubble Robotics Danu Robotics Embodied AI Extend Robotics Qualia Staer 3D-Components AS Touchlab Limited #PhysicalAI #Manufacturing #WorldModels #GoogleDeepMind #FORGIS #AcceleratedWithGoogle

🚀 From Document AI to an AI-Native Document Standard 🚀 For decades, we've relied on formats that were built for humans: 📄 PDF was designed for printing. 📝 DOCX was designed for editing. But neither was designed for AI. Today, we are excited to share a major step forward. The creators of Docling at IBM have joined forces with industry leaders across the document, AI, and enterprise ecosystem (ABBYY, NVIDIA, HumanSignal, FORGIS, Red Hat) to launch DocLang (https://doclang.ai/) — an open effort to create an AI-native document format. Why does this matter? As AI becomes the primary consumer of information, documents need to be more than visually correct. They need to be: ✅ Machine-readable ✅ Semantically structured ✅ Multimodal by design ✅ Reliable for AI agents and foundation models DocLang aims to provide a universal representation for any form of unstructured content — documents, reports, presentations, tables, forms, images, and beyond. Even more exciting, this initiative is being developed as an open standard under the Joint Development Foundation (JDF) with the long-term goal of becoming a candidate for international standardization. This is not a vendor project. It's an open invitation to researchers, developers, enterprises, content creators, and standards experts to help define how documents will be represented in the age of AI. The future of AI won't just depend on better models. It will depend on better data formats. 👉 Learn more and get involved: https://doclang.ai/ 👉 github: https://lnkd.in/etBHgZWk 👉 press-release: https://lnkd.in/eV7t4niA many thanks to: - the LF AI & Data Foundation for facilitating the international standardization process in record time - our many great collaborators: Maxime Vermeir, Douglas O'Flaherty, Daniel Fatade, Kari Ann Briski, Lauren Sell, Micaela Kaplan, Riccardo Maggioni, Jehlum Vitasta Pandit - our amazing Docling team: Panos Vagenas, Maksym Lysak, Michele Dolfi, PhD, Dr. Christoph Auer, Ahmed Nassar, Matteo Omenetti, Cesar Berrospi, Kasper Dinkla, Nikos Livathinos, Said Gürbüz, Jovana Kondić, Tim Strohmeyer and many more! cc: Abdel Labbi, Sriram Raghavan, Ed Anuff, Jordan Youngblood, Emma Gauthier, Lauren McHugh Olende, Bill Higgins #AI #IBM #DocumentAI #DocLang #Docling #OpenSource #Standards #GenerativeAI #LLM #EnterpriseAI #KnowledgeManagement #DataEngineering #MachineLearning

Europe built 3 frontier AI models last year. The US built 40. That race is already over. Here's the one that isn't. I mapped every serious AI and physical-AI powerhouse on Earth, by country. The picture explains the whole race. The US owns the frontier: OpenAI, Anthropic, Google DeepMind, Meta, xAI on models. NVIDIA, Tesla, Figure, Physical Intelligence, Boston Dynamics, Skild, Waymo on machines. One country, and the logos barely fit. China closed the gap in 24 months: DeepSeek, Qwen, Zhipu, Tencent, ByteDance, Baidu. And in robots it isn't close. One Chinese firm, Unitree, shipped 5,500 humanoids last year. Tesla, Figure and Agility shipped roughly 150 each. Europe shipped almost none. Europe? Mistral and Kyutai in France. Prior, Max Planck and Neura in Germany. NXAI in Austria. Real, but thin. The twist: the models race rewarded raw compute, and Europe lost it. The physical-AI race rewards what Europe already owns. Factories, robots, and the real-world data they throw off. Intelligence became a commodity. Physical-world data is the scarce asset now. Europe could still win this one. The data already exists, generated every day on its factory floors. But it sits unlabelled, useless to a model as it is. Labels are the missing piece, and without them Europe's manufacturers stay stuck. That changes soon. Big news from Forgis in a few days: a new Swiss physical-AI lab staking its claim on the world map. Who's missing from the map? #ArtificialIntelligence #Robotics #DeepTech #Automation

A 119-year-old company runs a startup challenge to find what comes next before anyone else does. Hundreds apply to the SKF Group Industry of the Future Challenge 2026. 10 made the finals. FORGIS is one of them. SKF is the world's largest bearing manufacturer. 20% global market share. Factory robots, machine tools, conveyors, assembly lines: bearings sit in every one. Around 1 million of those bearings are already cloud-connected, streaming vibration, load, and temperature data in real time. The largest industrial time-series dataset on earth. We are building the AI that operates on signals of that kind. World models for industrial systems, trained on raw machine signals. When SKF Group picks 10 startups to build with, the signal is clear: AI is moving onto the factory floor. A special thank you to the SKF Industry of the Future judges Nicolas Picco, Antoine Julien, Souleymane Mbengue, Mohamed si Achamrah, Nicolas Merlette, PASCAL BARBETTE, Lakdar SADI-HADDAD, Alain Bilde, and François Darmon, to our mentors Joël Mouterde and Antoine Brisson. And to the FORGIS team that carried this work forward: Francesco Nobili and Atharva Dastenavar. June 25. Saint-Marcel. We pitch in person. #Manufacturing #IndustrialAI #Automation #Entrepreneurship