Legal name: Advanced Machine Intelligence (AMI Labs)
AMI Labs: A USD 1.03B Seed Bet on World Models, LeCun's Non-Generative Path and Its Blank Scorecard
A Paris world-model company co-founded by Turing Award laureate Yann LeCun, pursuing a non-generative JEPA path; it closed a USD 1.03B seed round in March 2026 and released the open-source H-JEPA results in October 2026.
Executive Summary
AMI Labs, Advanced Machine Intelligence Labs, is the most watched world model company of 2026. Its legal entity is PengyouCo SAS, incorporated in Paris in December 2025 with 10,000 euros of share capital and registered at 51 rue de Seine in the sixth arrondissement, rebranded as AMI Labs in January 2026. It was co-founded by Turing Award laureate Yann LeCun, executive chairman, who left Meta at the end of 2025, and Alexandre LeBrun, CEO and founder of Wit.ai and Nabla, with a team including Laurent Solly as COO, formerly Meta vice president for Europe, Saining Xie as chief science officer, Pascale Fung as chief research and innovation officer and Michael Rabbat as vice president of world models. Headquarters are in Paris with offices in New York, Montreal and Singapore.
On March 10, 2026 the company announced a 1.03 billion dollar seed round, about 890 million euros, at a 3.5 billion dollar pre-money valuation and roughly 4.53 billion post-money. It is regarded as the largest seed round in European history and exceeded the company's initial 500 million euro target. It was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions, with strategic backing from NVIDIA, Samsung, Toyota Ventures, Sea, Temasek, SBVA, New Legacy Ventures, Alpha Intelligence Capital, the Mulliez family association and the Marcel Dassault industrial group, and angels including Eric Schmidt, Xavier Niel, Publicis, Jim Breyer, Bpifrance Digital Venture and Tim and Rosemary Berners-Lee.
The technical thesis is JEPA, the Joint Embedding Predictive Architecture proposed by LeCun in 2022. Its core idea is to predict missing or future parts in an abstract representation space rather than generating pixels or token sequences. The opposing camp is the prevailing route of autoregressive language models and diffusion or generative video models.
One framing point matters. This is a company with extreme conviction, extreme capital and a very distant product. It raised a billion dollars at seed yet acknowledges that the first year is fundamental research, commercial partner discussions begin in six to twelve months and commercially viable products may be years away. It has no commercial product, no revenue and no paying customer, with Nabla holding privileged access as a research partner. What changed in October 2026 sits on the evidence and ecosystem side rather than the commercial side: the first public research output, H-JEPA, was released with open code and checkpoints, and a non-binding memorandum of understanding with G42 and its portfolio company Analog was signed in Abu Dhabi. Three commercial questions carry the weight here, while whether JEPA is correct remains an academic matter without consensus: how much time a billion dollars buys, which of the first scenarios in industrial process control, robotics, wearables and healthcare pays first, and how much voice outside shareholders have under the dual-class structure.
Company Overview
From PengyouCo to AMI Labs
The origin is an unremarkable French simplified joint stock company incorporated in December 2025 with 10,000 euros of capital in the sixth arrondissement. The same month's articles split one million shares into ordinary shares and SP preference shares carrying ten times voting rights, an arrangement preserving founder control, especially LeCun's, through dilution.
In January 2026 the brand changed from PengyouCo to Advanced Machine Intelligence with the stated mission of building AI systems that go beyond language and are grounded in the physical world. In February LeCun joined as executive chairman and on March 10 the seed round was announced.
Team Composition
The team reveals positioning as neither a pure research lab nor a pure product company but a research company with commercial operators.
LeCun supplies academic direction and convening power as a Turing laureate, founder of Meta FAIR and pioneer of convolutional networks and energy-based models. LeBrun supplies commercialization experience through Wit.ai's acquisition by Facebook and founding Nabla. Solly brings large-company operations from Meta Europe. Saining Xie brings execution as a co-author of DiT and ConvNeXt. Pascale Fung brings research organization from Meta AI and HKUST. Rabbat owns the concrete world model technical route from Meta FAIR.
The combination covers academia, commerce and operations, and the risk is fragmentation, since LeCun remains a professor at NYU and LeBrun remains at Nabla, so time allocation by the two central figures affects execution speed.
On headcount, earlier public material was generally read as about a hundred people. In an AIM 2026 interview with La Tribune, LeCun put the research team at roughly 50 to 60 people across Paris, New York, Montreal and Singapore, and described 2026 as a research year with no product release planned in the near term. This report uses the interview figure and labels it as a media interview rather than a formal company disclosure.
The Choice of Paris
Headquarters in Paris rather than Silicon Valley is deliberate. LeCun has long argued Europe needs its own AI capability, and the French state through Bpifrance plus the Paris ecosystem of Mistral, Kyutai and Inria provide soil. Commercially that eases access to European public funding and policy support and to European research talent, at the cost of distance from Silicon Valley capital and customers, partly offset by offices in New York and Singapore.
Technology
What JEPA Solves
Understanding the thesis requires understanding LeCun's critique of current AI, which has three layers.
First, autoregressive language models learn next-token probabilities rather than how the world works. Text is a compressed representation of the world, and compression discards physical information such as mass, spatial relations and causal structure. A model can write the steps for pouring water without knowing inertia, center of mass or grip force.
Second, generative prediction of every pixel or token is wasteful. The world contains unpredictable detail such as leaf positions and cloud shapes, and forcing prediction of that detail spends capacity on meaningless randomness. He argues for prediction in abstract representation space, covering only what affects decisions.
Third, genuine intelligence requires a world model plus planning. An agent must simulate internally what happens if it acts a certain way and then choose the best action, which requires a model predicting state transitions rather than one continuing text.
JEPA targets all three. Encoders map observations such as images, video and sensor data into representation space, predictors forecast masked or future parts in that space, and regularization against representation collapse keeps the learned representations informative.
Technical Difference From Generative Models
The three routes differ fundamentally in what they predict. Autoregressive language models predict the next token, diffusion and video models predict the denoising process ending in pixel reconstruction, and JEPA predicts target embeddings in representation space.
The engineering consequences are that JEPA needs no pixel generation, so inference is cheap and real-time prediction is feasible, and it is not distracted by unpredictable detail, so representations are more abstract and better for downstream decisions. It also lacks generation, since it cannot directly output images or video without an additional decoder or policy head.
The trade-off suits robotics where the need is predicting action consequences rather than producing attractive video, and it disadvantages content creation where creators want visual output. That explains why target scenarios are industrial control, robotics, wearables and healthcare rather than content generation.
Progression From I-JEPA to V-JEPA 2
LeCun's team published the key work while at Meta: I-JEPA in 2023 for images, V-JEPA in 2024 for video and V-JEPA 2 in 2025 for video plus robot planning. V-JEPA 2 matters because it connects the world model to action planning, pretraining on video to learn world representations and then learning policies from limited robot interaction data to achieve zero-shot pick and place.
AMI Labs starts from that line but must solve what the Meta versions did not: how to move from lab demonstration to industrial usability. Three engineering problems arise, namely data diversity since industrial video and sensor distributions differ greatly from internet video, real-time response since industrial control demands millisecond latency, and safety and controllability since wrong actions on industrial equipment are costly.
The First Public Output: H-JEPA
On 5 October 2026 the team published H-JEPA, End-to-End Learning of Hierarchical World Models for Visual Planning, as arXiv:2610.06805, authored by Wancong Zhang, Basile Terver, Michael Rabbat, Yann LeCun and Randall Balestriero. The code is released under the MIT licence and 93 checkpoints are published on Hugging Face as jepa-world-models/h-jepa, covering AntMaze, FourRoom, Cube, Push-T and DROID, of which 84 are simulation checkpoints and 9 are DROID robot data checkpoints.
The reported figures are means with standard errors over three seeds. On Visual AntMaze the three-level model reaches 73.3 percent success against 18.0 percent for the flat LeWM baseline, and on FourRoom 96.0 percent against 40.7 percent. The gain is not uniform across tasks. On Push-T the three-level model reaches 17.3 percent, below the two-level model at 45.3 percent and the baseline at 40.0 percent, because deeper hierarchies train on far fewer clips. On DROID the model needs an inverse dynamics loss, without which performance collapses to zero, and with that loss the two-level H-JEPA improves on the LeWM plus inverse dynamics baseline.
What this changes is verifiability rather than readiness. A first public result with open code and weights makes the JEPA line inspectable by outside parties and gives the company a measurable artefact for its 2026 research year. It remains a benchmark result on simulation and robot data, with no industrial deployment, no product and no pricing attached.
Technical Thresholds by Scenario
The four scenarios differ greatly in threshold.
Industrial process control needs models that understand causal structure among temperature, pressure and flow and predict before anomalies occur, with difficulty in scarce and expensively labeled industrial data and in reliability where both false positives and misses are costly.
Robotics needs understanding of 3D space, object properties and action consequences, with difficulty in the sim-to-real gap and long-horizon planning.
Wearables need models running in real time on low-power chips understanding user behavior and intent, with difficulty in compute constraints and privacy since data cannot leave the device.
Healthcare needs understanding of physiological processes and clinical decisions, with difficulty in regulation such as FDA certification and in liability for wrong advice.
All four need physical understanding with completely different engineering constraints, and pursuing all four divides resources, which is why the company says the first year is research only.
Technical Shortfalls
First, JEPA generality is unproven. I-JEPA and V-JEPA perform well on visual representation learning, but whether they support industrial-grade world models with long horizons, multimodality and high precision is unevidenced.
Second, no generation capability. Being unable to generate images or video excludes content creation and simulation visualization, limiting market size.
Third, data and compute demands. World models need large video and sensor datasets and large-scale compute, and a billion dollars is enough to start but not to sustain.
Fourth, no evaluation standard. There is no accepted metric for world model quality, making progress hard to verify externally.
Product and Commercial Model
Current State: No Commercial Product
As of November 2026 there is no commercial product, no pricing and no paying customer, with Nabla holding privileged access as a research partner. What exists instead is an open research output, the H-JEPA code and checkpoints, and a research collaboration framework, the non-binding memorandum of understanding with G42 and Analog. The company states the first year is fundamental research, partner discussions start in six to twelve months and commercial products may take years.
The timeline is honest and implies a long payback. Against World Labs, which launched Marble commercially in November 2025, AMI Labs trails productization by about a year.
Research Collaboration Framework: G42 and Analog
On 20 October 2026, during the UAE and France business dialogue in Abu Dhabi, G42 and its portfolio company Analog signed a non-binding memorandum of understanding with AMI Labs. The scope covers frontier AI research including world models, advanced compute infrastructure, sovereign enabled cloud and a potential integration of Analog Physical Intelligence for industrial applications.
Two features matter for how this is read. The document is explicitly non-binding, and G42 stated that scope, commercial terms and implementation are left to subsequent definitive agreements. No value, duration or milestone is disclosed, so the agreement is a research and infrastructure framework rather than a contract, a purchase order or a revenue event. Its significance is optionality: it places AMI Labs in front of a sovereign compute provider and an industrial AI integrator, the kind of relationship that tends to precede a first industrial deployment.
Nabla as First Partner
Nabla is LeBrun's medical AI company and gains privileged access to AMI world model technology to build healthcare software pursuing FDA certification. The logic is that healthcare needs understanding of physiology and clinical decisions, matching world model capability, and LeBrun's familiarity reduces partnership risk.
It is also a related-party transaction, with the CEO's own company as the first partner. Common in early companies validating technology in familiar settings, it nonetheless requires disclosure and review in governance.
Relationship With Meta
After LeCun's departure Meta holds no stake but signed a research partnership giving access to AMI research outputs, and LeCun says early discussions are underway about applying the technology to Ray-Ban Meta smart glasses.
The arrangement benefits AMI through Meta's scenarios and data feedback while raising a question about ownership and exclusivity, since priority access for Meta would dilute value for NVIDIA, Samsung and Toyota.
Possible Business Models
Three models are inferable: technology licensing of world models to industrial and robotics companies, a platform offering APIs and tooling, and joint development of vertical applications with large customers.
Margins and scalability differ, with licensing lightest, platforms most scalable and joint development heaviest with the highest value per account. Given capital and team, the early path is likely joint development plus licensing, moving to a platform later.
Competitive Landscape
Competitors in World Models
The direct rival is World Labs under Fei-Fei Li, acquired by AMD in September 2026 for about 8.2 billion dollars. Both build world models with different routes: World Labs pursues 3D-native generation with Marble and Atlas producing interactive 3D scenes, while AMI Labs pursues JEPA representation prediction without generating pixels.
Other related companies include Physical Intelligence for robot foundation models, Skild AI for general robot brains, 1X for humanoids and internal projects at Google DeepMind with Genie and at OpenAI.
Relationship With the LLM Route
LeCun's position is that LLMs have a ceiling and world models are next, a judgment still debated, with supporters believing scaling laws continue and critics arguing language models lack physical grounding.
Commercially the routes are not fully opposed, since LLMs can serve as the interface layer interpreting instructions and explaining results while world models serve the physical layer predicting state transitions. The rational architecture combines both, while AMI's narrative emphasizes going beyond language.
Europe's Strategic Position
AMI Labs is positioned as a European alternative to American and Chinese AI giants, which carries policy value through EU AI sovereignty narratives and French national champion strategy as well as commercial risk since European customers pay less and adopt more cautiously than American ones.
Against Mistral, Europe's LLM champion valued above six billion euros, the differentiation is not building LLMs but world models, which avoids direct European competition with Chinese vendors while requiring the company to open a new market itself.
Talent Competition
Talent in world models is extremely scarce. World Labs has Li and a NeRF inventor, AMI Labs has LeCun and the V-JEPA team, Physical Intelligence has Sergey Levine and Skild has a CMU robotics team. A billion dollars mainly buys higher compensation and greater research freedom.
Financing and Capital Structure
Size and Structure of the Seed
The 1.03 billion dollar seed at 3.5 billion pre-money implies about 4.53 billion post-money, with five co-leads in Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions, strategic investors including NVIDIA, Samsung, Toyota Ventures, Sea and Temasek and angels including Eric Schmidt and the Berners-Lees.
The structure makes it not a venture-led company but a coalition of industrial capital, celebrity angels and European policy capital, with NVIDIA and Samsung seeking compute and hardware synergy, Bezos and Schmidt acting on conviction and Bpifrance representing French state support.
Dual-Class and Governance
The articles split shares into ordinary and SP preference shares with ten times voting rights, preserving founder control through dilution.
Dual-class structures are increasingly common in AI companies, from OpenAI's special structure to Anthropic's long-term benefit trust and Mistral's founder control, protecting long-term research orientation from short-term capital pressure at the cost of limited outside shareholder voice.
What the Capital Buys
A billion dollars for a company of roughly fifty to sixty people buys a top research team at 500,000 to one million dollars per researcher, a large training cluster of thousands of GPUs and several years of research. The smaller headcount implied by the AIM 2026 interview figure stretches that runway and lowers spend per employee, while it also concentrates value in a very small number of people. It does not buy indefinite exploration, and without verifiable technical or commercial progress within three years the next round becomes difficult.
DISCREPANCIES
1. Two dates for the seed announcement. The official announcement and LeCun's post are dated March 10, 2026, while some media record March 9 through time zone differences. This report uses March 10. Source tier: OFFICIAL (official announcement and founder post).
2. Pre-money and post-money valuation framings. The official figure is 3.5 billion dollars pre-money, some media cite 4.53 billion post-money, and others write 3.5 billion without distinguishing. Both are correct with different meanings. This report labels both. Source tier: OFFICIAL (pre-money) and inference (post-money).
3. Target versus actual raise. The initial target was 500 million euros and the final raise was 890 million euros, or 1.03 billion dollars. Some coverage cites only the billion without noting the oversubscription. This report presents both. Source tier: OFFICIAL (company statement) and MEDIA.
4. Legal entity versus brand. The legal entity is PengyouCo SAS incorporated in December 2025, rebranded as AMI Labs in January 2026, while some material states AMI Labs was founded in 2026 and omits the 2025 incorporation. This report separates incorporation from brand launch. Source tier: OFFICIAL (registry) and MEDIA.
5. Strength of the largest European seed claim. The company and media call it the largest seed in European history with qualifiers such as believed and probably. The first depends on how seed is defined, whether strategic investment and convertible instruments count, and cannot be independently verified. This report retains the claim with qualifiers. Source tier: MEDIA with qualifiers.
6. Composition of founders. Some sources name LeCun and LeBrun as founders, others list Saining Xie and Pascale Fung as co-founders. The official framing is co-founding by LeCun and LeBrun with the others as core team members. This report uses the official framing. Source tier: OFFICIAL.
7. Optimistic and conservative readings of the timeline. The company says the first year is research, commercial discussions begin in six to twelve months and products may take years. Some read this as pragmatic and others as distant commercialization. Both readings rest on the same statement, and this report presents both. Source tier: OFFICIAL (company statement).
8. Nature of the Nabla partnership. Nabla holds privileged access, and whether payment or exclusivity applies is undisclosed. Some interpretations treat Nabla as the first customer. This report labels it a partnership rather than a paying customer. Source tier: OFFICIAL (partnership announcement) distinguished from inference.
9. Team size. Earlier material put the research organisation at about a hundred people, while an AIM 2026 interview with LeCun published by La Tribune describes roughly 50 to 60 people across Paris, New York, Montreal and Singapore. No formal company headcount disclosure exists. This report uses the interview figure and marks it as an interview statement. Source tier: MEDIA (interview).
10. Statistical form and depth dependence of the H-JEPA results. The paper reports means with standard errors over three seeds, and the plus or minus values are frequently read as standard deviations. The gains are also not monotonic in depth, since on Push-T the three-level model reaches 17.3 percent against 45.3 percent for two levels. This report uses the standard error reading as stated in the paper and reports the Push-T inversion explicitly. Source tier: OFFICIAL (paper tables) distinguished from MEDIA paraphrase.
11. Legal and commercial weight of the G42 and Analog memorandum. The document of 20 October 2026 is explicitly non-binding, G42 stated that scope, commercial terms and implementation are left to subsequent definitive agreements, and no value, duration or milestone was disclosed. It is therefore a research and infrastructure framework rather than a contract, an order or a revenue event. Source tier: OFFICIAL (announcement).
Risks and Diligence Gaps
First, productization time. The company itself says the first year is research and products may take years, implying no revenue for three to five years, and whether a billion dollars lasts depends on burn rate.
Second, JEPA validation. JEPA has results in visual representation learning, but industrial-grade world model feasibility is unproven, and failure would force a pivot toward generative methods and damage the narrative.
Third, related-party transactions and governance. Nabla is the CEO's company with privileged access, Meta has a research partnership and dual-class shares give founders board control, all reasonable early but scrutinized in later rounds and listings.
Fourth, talent dependence. Value concentrates in LeCun and Saining Xie among others, and departures would materially affect valuation.
Fifth, scenario dispersion. The four target scenarios differ greatly in engineering constraints, and pursuing them simultaneously divides resources.
Sixth, competition. World Labs has been absorbed by AMD with more resources, Physical Intelligence leads in robot foundation models and Google and OpenAI also pursue world models.
Seventh, European commercialization capability. European industrial customers pay well, especially in German and French manufacturing, but adopt new technology cautiously, making the path from research to industrial deployment potentially longer than in the United States.
Eighth, the open source path. Releasing H-JEPA under the MIT licence with 93 checkpoints makes the work reproducible and easy to build on, which helps adoption and talent attraction, and it leaves the company with no direct revenue from the artefact itself. The G42 and Analog memorandum adds reach without adding commitment, since neither carries a price, a schedule or an obligation. If the first product arrives later than the ecosystem expects, an open research programme is what third parties will keep using.
Composite Assessment
techScore 8.0. JEPA is a theoretically deep route, I-JEPA, V-JEPA and V-JEPA 2 demonstrate feasibility, and LeCun with Saining Xie provides execution. The October 2026 H-JEPA release adds the first checkable public evidence for the line, with a three-level model at 73.3 percent on Visual AntMaze against 18.0 percent for the flat baseline. Deductions for unproven industrial feasibility, limited scenarios from lacking generation and missing evaluation standards.
bizScore 4.5. Target scenarios in industrial control, robotics, wearables and healthcare are all high value, Nabla provides a first validation setting, and the G42 and Analog memorandum opens a research and infrastructure framework with a sovereign compute provider. Deductions for no commercial product, no revenue, no paying customer, a non-binding agreement, productization measured in years and a CEO-related party transaction.
fundScore 8.0. A 1.03 billion dollar seed, 4.5 billion post-money valuation and an investor base of industrial and policy capital provide ample funding, and a research team of roughly fifty to sixty people stretches that further. Deductions for a fast burn rate, next-round dependence on technical breakthroughs and a belief-based valuation.
productScore 4.5. There is still no commercial product, pricing or customer case, and the company now holds open research output in the form of H-JEPA code and 93 checkpoints, plus a signed research collaboration framework with G42 and Analog. Deductions for the absence of any paying customer and of a commercial offering.
marketScore 7.5. World models are seen as a key next-generation direction, industrial and robotics demand for physical understanding is real and the European AI sovereignty narrative provides policy support, reinforced by the G42 sovereign cloud channel. Deductions for an early market, long adoption cycles and intense competition.
Overall 6.5. This is a bet that buys time with a billion dollars: LeCun uses his academic standing and the JEPA route to assemble Europe's largest seed round and prove that language is not all of intelligence. October 2026 moved the bet from assertion to evidence, since H-JEPA gives the outside world code, weights and numbers it can check, and the G42 memorandum gives it a counterpart in sovereign compute. Upside is enormous if JEPA becomes the mainstream architecture, and downside is concrete if no commercial product appears within three years. The judgment on physical AI is that the question JEPA poses, how to predict the physical world in representation space, is profound, and the first public results make the route credible while leaving it industrially unvalidated, so AMI Labs' value lies more in exploring that possibility than in capabilities already built.
References
- AMI Labs official announcement, March 10, 2026: 1.03 billion dollar seed, about 890 million euros, at a 3.5 billion dollar pre-money valuation; co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions; strategic investors including NVIDIA, Samsung, Toyota Ventures, Sea, Temasek, SBVA, New Legacy Ventures, Alpha Intelligence Capital, the Mulliez family association and the Marcel Dassault industrial group; angels including Eric Schmidt, Xavier Niel, Publicis, Jim Breyer, Bpifrance Digital Venture and Tim and Rosemary Berners-Lee
- Governance and entity information from public sources: the legal entity PengyouCo SAS incorporated in Paris in December 2025 with 10,000 euros of capital at 51 rue de Seine in the sixth arrondissement; articles updated the same month splitting one million shares into ordinary and SP preference shares with ten times voting rights; rebranding to Advanced Machine Intelligence in January 2026
- Team information from public sources: Yann LeCun as executive chairman, 2018 Turing Award laureate, founder of Meta FAIR and departing Meta at the end of 2025 while remaining an NYU professor; Alexandre LeBrun as CEO and founder of Wit.ai and Nabla; Laurent Solly as COO and former Meta vice president for Europe; Saining Xie as chief science officer; Pascale Fung as chief research and innovation officer; Michael Rabbat as vice president of world models; headquarters in Paris with offices in New York, Montreal and Singapore
- Technical route from public sources: JEPA proposed by LeCun in 2022, predicting missing or future parts in abstract representation space rather than generating pixels or tokens, with related work including I-JEPA in 2023 for images, V-JEPA in 2024 for video and V-JEPA 2 in 2025 for video plus robot planning
- LeCun's technical critique from public reporting: autoregressive language models learn next-token probabilities rather than how the world works; text is a compressed representation discarding physical information; generative prediction wastes capacity on unpredictable detail; genuine intelligence requires a world model plus planning
- Target applications from company statements: industrial process control, automation, robotics, wearable devices and healthcare, with the mission statement that real intelligence does not start in language but in the world
- Commercialization timeline from company statements: fundamental research in the first year, industry partner discussions in six to twelve months and commercially viable products possibly years away; LeBrun warning that world models will be the next buzzword and every company will call itself a world model within six months
- First partnership from public reporting: Nabla, LeBrun's medical AI company, gains privileged access to AMI world model technology to build healthcare software pursuing FDA certification
- Meta relationship from public reporting: Meta holds no stake but signed a research partnership with access to research outputs, and LeCun says early discussions are underway about Ray-Ban Meta smart glasses
- Competitive landscape from public sources: World Labs under Fei-Fei Li pursuing 3D-native generation and acquired by AMD in September 2026 for about 8.2 billion dollars; related companies including Physical Intelligence for robot foundation models, Skild AI for general robot brains and 1X for humanoids; corporate projects including Google DeepMind Genie and OpenAI world model research
- Industry background from public sources: world models learn state transitions rather than data distributions; the fundamental difference among JEPA, autoregressive language models and diffusion or video models lies in what is predicted, representation space versus tokens versus pixels; there is no consensus evaluation standard for world models
- European AI background from public sources: the EU promotes AI sovereignty narratives and France supports domestic AI companies through Bpifrance and a national champion strategy; Mistral represents European LLMs at a valuation above six billion euros; Paris hosts an AI startup ecosystem including Mistral, Kyutai and Inria
- H-JEPA preprint and open release, October 2026: H-JEPA, End-to-End Learning of Hierarchical World Models for Visual Planning, arXiv:2610.06805, submitted 5 October 2026, by Wancong Zhang, Basile Terver, Michael Rabbat, Yann LeCun and Randall Balestriero; code released under the MIT licence; 93 checkpoints published on Hugging Face as jepa-world-models/h-jepa across AntMaze, FourRoom, Cube, Push-T and DROID, 84 simulation and 9 DROID; results reported as mean with standard error over 3 seeds, Visual AntMaze 73.3 percent against 18.0 percent for the flat LeWM baseline, FourRoom 96.0 against 40.7, Push-T 17.3 against 45.3 for two levels, and DROID requiring an inverse dynamics loss
- G42 official announcement, 20 October 2026: G42, its portfolio company Analog and AMI Labs signed a non-binding memorandum of understanding during the UAE and France business dialogue in Abu Dhabi, covering frontier AI research including world models, advanced compute infrastructure, sovereign enabled cloud and a potential integration of Analog Physical Intelligence for industrial applications, with scope, commercial terms and implementation left to subsequent definitive agreements and no value, duration or milestone disclosed
- La Tribune interview with Yann LeCun during AIM 2026, reported in October 2026: research team of roughly 50 to 60 people across Paris, New York, Montreal and Singapore, with 2026 framed as a research year and no product release planned in the near term
This report is compiled from public sources for informational purposes only and is not investment advice. The company is private with no commercial product or revenue, funding and valuation data come from official announcements and media, and technical capability and commercialization timelines are company statements without third-party verification. Data is current as of 2 November 2026.
Related papers
- HOTICE: Whole-Body Humanoid Object Transportation in Cluttered Environments
- JAMB: Joint Action-Motion Diffusion for Bimanual Manipulation
- History-Conditioned Flow Matching for Probabilistic Dynamics of Tendon-Driven Continuum Robots
- Koopman-Accelerated Model-Based Diffusion for Real-Time Robot Control
- Towards Adaptive Interaction Strategies for Human Companion Robot via Deep Reinforcement Learning
- GINIO: A Geometric SO(3)-Equivariant Interface for Neural Inertial Odometry
- Transolver: A Transformer Solver for Large-Scale Industrial Geometries
Related tools
Related cases
Related reports66
- 51World Deep Dive: A Listed Digital Twin Company Where 51Sim Becomes the First Growth Engine While 51Earth Nearly Vanishes, and Two Irreconcilable Financing Naming Conventions Must Both Be Kept
- Anshi Pacific (PERA) Deep Dive: A Beijing Domestic-CAE Leader Founded in 2003 Whose PERA SIM General Simulation, AIROM Reduced-Order Models, AgentX Multi-Agent and Jingzhi iGPT/Whale Span from Industrial Simulation to Physics AI
- AspenTech Deep Dive: The Process-Simulation Incumbent Acquired by Emerson and Its Hybrid Industrial-AI Route
- Autodesk Deep Dive: Neural CAD Foundation Models and a $200M World Labs Bet on Physical-World AI, with NAVASTO Adding CFD Surrogate Capability
- Bentley Systems Deep Dive: iTwin Digital Twin and Generative AI Empower Full-Lifecycle Infrastructure Engineering
- BeyondMath Deep Dive: A Foundation Physics AI on the Neural-Operator Route, a $19M Consortium Budget, and Three Misread Facts
- Cadence Deep Dive: System-Level Multiphysics EDA and the Millennium M2000 Supercomputer Anchor Physical AI, with AuraStack Agentic AI and Fidelity GPU CFD as Key Levers
- COMSOL Deep Dive: Four Decades of Multiphysics Solvers, Entering Physical AI via DNN Surrogate-Model Training and All-Physics GPU Acceleration
- Dassault Systèmes Deep Dive: 3DEXPERIENCE and SIMULIA Virtual Twin Physics Behavior Anchor Industrial AI, with LEO Virtual Companion and GPU Explicit Dynamics as Physical-AI Levers
- DP Technology: From the Gordon Bell Prize to National-Level Unicorn Status and the Door of a Hong Kong Listing, the AI-for-Science Leader's Technical Delivery and Commercial Blank
- Emmi AI: From Linz Spin-out to a EUR 330M Mistral Acquisition in 17 Months
- Fainite Deep Dive: A Zurich Physics-AI Simulation Startup Whose Caltech, ETH Zurich, Cambridge and Google Team Uses Physics-Aware Deep Learning to Cut Simulation from Tens of Hours to Seconds, With Only Venture Kick Grants and No Priced Round
- Fysics AI: The PhysX Founder's Second Act, a Domestic Differentiable Physics Engine and Its Ecosystem Play
- Frontop (301313) Deep Dive: From Digital Exhibition Halls to a Physical-AI Data Foundation - FTE Engine, FunSim Simulation and the Gap Between Positioning and Revenue
- Godela Deep Dive: A YC X25 Physics-AI Simulation Startup Whose Ex-Apple, Google and Intel Team Uses Physics-Constrained Surrogate Models to Cut Simulation from Days to Seconds, at Seed Stage with Pilot-Only Access
- HollyWave (688682) Deep Dive: RDSim AI Electromagnetic Simulation and Engineering Agents Close the Simulation-Measurement-Product Loop
- IAT (Alt) Deep Dive: China's First Listed Independent Automotive Design Company Pivots to AI-native R&D with the Yuanfang, Yufeng and Taiyi 3.0 Product System and an NVIDIA Omniverse Partnership, Reporting RMB 975M Revenue and a Widened RMB 250M Net Loss in 2025
- Inductiva Deep Dive: A Closed Case in Open-Solver Orchestration, Eight Premises That Must Be Corrected, and Where Co-Founder and CTO Hugo Penedones Went Next
- Intesim (INTESIM) Deep Dive: A Home-grown Chinese CAE Leader Founded in 2009 with Self-developed Solvers for Structure, Fluid, Electromagnetics, Thermal and Acoustics and Multiphysics Coupling, a ~RMB 100M B Round from CNII and Shenzhen Capital, and 2025 Revenue and Bookings Both above RMB 100M
- iSoftStone (301236) Deep Dive: A ChiNext-Listed IT Services Giant Whose TianShu Platform Builds a Three-Layer Physical-AI Stack with Self-Developed TwinHub Digital Twin, FSim Industrial Simulation, CAE Multiphysics and Phyxis Embodied Control
- Jua Deep Dive: The Earth Physics Transformer as a Large Physics Model for Energy Trading
- Keysight (with ESI Group) Deep Dive: Virtual Prototyping and Hybrid AI Bridge Measurement and Simulation, with Natural-Language Assistants in EDA
- Lightwheel Deep Dive: A Physical AI Data and Evaluation Infrastructure Startup Founded by a Former NVIDIA Simulation Lead, Seven Rounds Totalling Roughly CNY 2.65 Billion at a Valuation Above USD 2 Billion, Seats on Both the Newton and EgoVerse International Standards Committees, With Two Incompatible Accounts of Q1 Orders
- Luminary Cloud: The Physics AI Model Factory Backed by NVentures
- MathWorks Deep Dive: MATLAB/Simulink and Simscape Multiphysics Power Physical AI, with R2026a Copilots and an Agentic Toolkit for Edge Intelligence
- Maniformer (Mifeng) Deep Dive: An Agibot-Spun Physical-AI Data Infrastructure Firm Whose Bodyless MEgo Capture Hardware Is in Mass Production and Which Closed Three Rounds in Half a Year Led by China Telecom
- Mistral AI: How Physics AI Moved From Language Models to EUV Lithography and Nuclear Power
- Monolith AI: The British Engineering-AI Firm That Cut Physical Test Counts Instead of Accelerating Simulation, Now Inside CoreWeave
- Monumo Deep Dive: A Cambridge Deeptech Company Redesigning Electric Motors with the Anser AI Engine, Tens of Millions of Multiphysics Simulations per Day, GBP 10.5M Raised in Total, and a Series A Being Raised
- NAVASTO: Moving Solver Insight Onto the Designer's Desk, and How Autodesk Bought That Entry Point
- Navier: Handing CFD Workflows to Agents, a Seed-Stage Bet on AI for Hardware Engineering
- Neural Concept: An EPFL Spin-out Betting on Neural Surrogates, with $138M Raised
- NP Co. (NP Company) Deep Dive: A 2025-founded Paris Inria-born AI-native simulation startup that raised a EUR 6M pre-seed led by Partech with Mistral AI co-founders as angels, claiming a pre-trained physics foundation model at 1,000x speedup — most claims unverified
- NVIDIA Deep Dive: Building the Full-Stack Physical-AI Infrastructure with PhysicsNeMo, Warp, Cosmos World Foundation Models and Omniverse
- Physical Intelligence Deep Dive: A $2.1B Bet on the GPT Moment for Robotics, With pi0.7 Self-Reporting 60-80 Percent Success on Unseen Tasks and Neither Pricing nor Revenue Disclosed for the Pi API
- Physical Superintelligence Deep Dive: A $58M Seed Bet on Virtual Physicists, With Emmy Targeting AI Data Center Multiphysics Optimization as Its First Commercial Proof Point
- PhysicsX: Pretrained Geometry and Physics for Industrial Engineering at a $2.4B Mark
- Prometheus: A $41B Mark, $18.2B Raised, and Not Yet a Product
- PTC Deep Dive: Creo/Onshape and the ThingWorx Digital Thread Power Physical AI, with NVIDIA Isaac Sim Bridging Robotics Design to Simulation
- Qianjue Technology Deep Dive: How a Body-Agnostic Embodied Brain Pairing the Predictive World Model Polibrain with Brain-Inspired Partitioning and Polynomial Representation Reached a Hundred Thousand Cleaning and Service Robots
- Quanscient Deep Dive: Cloud-Native and Quantum-Powered Multiphysics Simulation as the Data Engine for AI-Native Hardware Engineering
- Rescale Deep Dive: An HPC Operating System and Desktop Engineering Workbench in the Cloud, a $115M Series D, and Five Misread Facts
- Schrödinger Deep Dive: Physics-Based FEP+ Simulation and Machine Learning Power Molecular Discovery, with the 2026 Bunsen Agentic Co-Scientist
- Secondmind Deep Dive: A Cambridge Engineering-AI Firm Whose Bayesian Active Learning Cuts CAE Simulation and Test-Bench Runs by up to 90%, with Roughly USD 73M Raised and Mazda Both Customer and Investor
- Shichuang Technology Deep Dive: SupreCAST Cloud-Native Material-Processing Simulation and Industrial Agents Advance the CAE+AI Strategy
- ShonCloud (Shunyun) Deep Dive: Meshless Particle-Method CFD for EV and Nuclear Fluid Multiphysics
- Shuzhi Maijin Deep Dive: A Shenzhen Fully Self-Developed SNN-plus-LBM CFD Startup Claiming Second-Level Results on Hundreds-of-Millions of Grids on Consumer Hardware, With a Tens-of-Millions-RMB Angel Round From a Ministry-of-Finance-Affiliated Fund and an FPGA Migration Plan
- Siemens (incl. Altair) Deep Dive: The 2025 Altair Acquisition Unites Simcenter and Altair, with PhysicsAI Geometric Deep Learning and PhysicsAI Generate as the Physical-AI Core
- Silurian Deep Dive: A Generative Forecasting Transformer Earth-System Foundation Model for Weather Risk on Energy and Infrastructure Assets
- SimScale Deep Dive: A Cloud CAE Platform Built on Core-Hour Subscriptions, EUR 52M of Equity Kept Separate from EUR 25M of Debt, and Five Misread Facts
- Skild AI Deep Dive: A General Robot Foundation Model Incubated at CMU's Robotics Institute, With a USD 1.4B Series C at a USD 14B Valuation, a Conflicting Series B Size, and an ARR Figure That Is a Run-Rate Rather Than Revenue
- Songying Deep Dive: A Huawei-Cloud-Alumni Startup Building a Domestic Physical AI Operating System, With ORCA OS Positioning Against NVIDIA Omniverse, Three Incompatible Financing Narratives, and a Verifiable Third-Party Anchor in a CNY 20 Million Minority Investment
- Sudo Tech Deep Dive: Zero-Real-Data Simulation Training, the Fully In-House Sudo R1 Stack, and an Atomic-Skill Commercialisation Strategy from a Startup Valued Above CNY 20 Billion in 17 Months
- Suochen Technology: The Pain of Becoming a Physics-AI Platform from a CAE Vendor
- Synopsys (Ansys) Deep Dive: The 2025 Ansys Acquisition Turns the EDA Leader into a Silicon-to-Systems Engineering Powerhouse, with Ansys SimAI and the NVIDIA Physical-AI Partnership as the Core Physics-AI Levers
- TenFong Deep Dive: An Academician-Led Chinese Industrial Simulation Software Vendor Whose Official Timeline and Database Records Use Incompatible Round Names, With an Official Financing PDF and an Investment Database Both on File
- Tianfu (Nanjing Tianfu Software) Deep Dive: A 2011-founded Chinese Intelligent-CAE Veteran That Closed a Several-Hundred-Million-RMB Series C, Pivoted to an Industrial World Model in 2024, With Unreconcilable Financing and Valuation Records and No Disclosed Revenue
- Tongyuan Soft Control Deep Dive: MWORKS (Syslab/Sysplorer) as a Domestic MATLAB/Simulink Alternative, with MWORKS AI and the MOGICK Engineering Agent Advancing Home-Grown Physical AI
- Vinci AI: Turning Physics Simulation from a Checkpoint into a Continuous Constraint
- Westlake Robotics Deep Dive: Westlake University's First Robotics Spin-out Builds a Dual-Pretrain Brain-and-Cerebellum Architecture with WR1 Whole-Body Foundation Model and GAE Behaviour Expert
- World Labs: Fei-Fei Li's Spatial-Intelligence Unicorn, Acquired by AMD for $8.2B
- XVERSE Deep Dive: A Simulation-World Route to Physical AI Built on the XVERSE Model Family, an Edge-Cloud 3D Engine and the Physics-Based MotionGen Model
- SimWE: From In-House Solvers to a Physical-AI Engine, the Domestic CAE Leader and What Comes Next
- Yunjing Zhifang Deep Dive: A Peking-University-Incubated Cloud-Native GPU CAE Company with the Simversus Platform and the YJ-DORGP Small-Sample Intelligent Solver, Reporting 275 Percent Order Growth, 160 Percent Revenue Growth and Breakeven in 2025
- Zhongke Diwiji (Five Ages) Deep Dive: A CAS-Institute-of-Automation and Tsinghua Spin-out Whose FAM Embodied Foundation Model Uses Action Optical Flow for Cross-Embodiment Generalisation, Reaching 97% Success from 3 to 5 Demos, with EraBrain and Jiling Robots Serving Sinopec, Leapmotor and Midea
- ZWSOFT (688083) Deep Dive: Overdrive Kernel and ZW3D Wukong 2027 Underpin an Integrated CAD/CAE/CAM Stack, with a CAx Agent Platform as the Practical Route to Engineering AI
This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of the publication date; please refer to the company’s official filings.