AMI Labs raised a $1.03B seed round announced on March 10, 2026, at a post-money valuation of approximately $4.5B.

The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions. Additional participants included Toyota Ventures, Temasek, SBVA, NVIDIA, Samsung, Sea, Alpha Intelligence Capital, Bpifrance Digital Venture, Publicis Groupe, Groupe industriel Marcel Dassault, Association Familiale Mulliez, Aglaé Lab, ZEBOX Ventures, Artémis, Xavier Niel, Eric Schmidt, Jim Breyer, Mark Cuban, Mark Leslie, and Tim and Rosemary Berners-Lee.

This is AMI's first disclosed funding round. Total capital raised stands at $1.03B.

AMI is building what it calls a world model, an AI system that learns to understand the physical environment from raw sensor data, primarily video and camera streams, but also other sensor modalities, rather than from text.

The core idea is that most real-world AI problems are not language problems. A robot arm, a factory machine, a medical monitor, or an autonomous vehicle all generate continuous streams of noisy, high-dimensional sensory data. The challenge is not predicting the next word. It is predicting what will happen next in the environment, and what the consequences of a given action will be.

AMI's architecture is built around the JEPA family of models, most closely associated with the V-JEPA 2 research lineage. Rather than trying to reconstruct every pixel of a future video frame, which is computationally expensive and filled with irrelevant detail, the system learns to predict abstract latent representations of future states. It is designed to ignore unpredictable noise and focus on the stable, decision-relevant structure underneath.

In practical terms, this means the model can learn concepts like object permanence, cause and effect, motion dynamics, and action consequences from watching large amounts of real-world video. A model trained this way should infer that a cup near a table edge may fall, that a robot arm moving in a certain direction will collide with an object, or that a shift in a biomedical sensor signal corresponds to a meaningful physical change rather than random noise.

The intended workflow for an enterprise customer looks like this: ingest sensor streams from a robot, machine, or clinical device; train or adapt the world model on that environment's specific dynamics; use the model to simulate what would happen under different candidate actions; and then select a safe, planned course of action. The system is explicitly designed for prediction, reasoning, planning, and safe control, not for open-ended conversation.

AMI's target applications span industrial process control, factory automation, robotics, wearable devices, and healthcare. The Nabla partnership provides the clearest healthcare example: AMI's world models would sit beneath clinical AI agents that need deterministic, safety-conscious decision support rather than probabilistic text generation.

The eventual product is planned to be available in two forms: a hosted paid API for organizations that want managed access, and a downloadable, adaptable version that enterprises can run within their own infrastructure. The latter option is particularly relevant for customers in regulated industries that need on-premises deployment, data locality, or air-gapped environments.

As of March 2026, AMI is still in a fundamental research phase. The company plans to continue training its models on video data before moving into concrete application testing with industrial partners.

AMI operates as a research-first frontier lab, funded entirely by equity capital, with commercialization framed as a multi-year objective rather than a near-term priority.

The cost structure is straightforward: the two primary expenditures are compute and talent. Training world models on large-scale video and sensor data is computationally intensive, and AMI is building a small team, roughly 10 people at launch, targeting 30 to 50 within six months, distributed across Paris, New York, Montreal, and Singapore. That lean headcount relative to the capital raised indicates a research-lab operating model rather than a sales-and-marketing-led growth model.

The planned monetization architecture is dual-track B2B. The first track is a hosted API, where enterprise customers pay for managed access to AMI's world models. The second track is a deployable, self-hosted version that customers can adapt and run on their own infrastructure, potentially open source. This mirrors an open-core logic: publish research and base models to build ecosystem adoption, then monetize enterprise-grade deployment, domain adaptation, safety layers, and support.

The go-to-market motion is partnership-led rather than direct sales-led. AMI's relationship with Nabla is the template: a strategic partner gains early or exclusive access to AMI's technology in exchange for providing a real deployment environment, domain-specific data, and a channel into a high-value vertical. The investor roster, which includes Toyota Ventures, NVIDIA, Samsung, and industrial family offices with ties to Dassault and Mulliez, suggests that several future partnerships could emerge from within the cap table itself.

The long-run margin logic is similar to other frontier model platforms: heavy upfront R&D and compute costs, with the expectation that reusable base models can be deployed across multiple industries without proportional cost increases. The most relevant commercial analogs are not OpenAI or Anthropic at consumer scale, but something closer to Helsing's software licensing and integration model for defense-adjacent AI, or the recurring software layer that companies like 1X Technologies are building on top of physical-world deployments.

The structural constraint is timing. AMI is explicitly not optimizing for near-term revenue, which means the business depends entirely on the $1.03B seed runway lasting long enough for the research to mature into deployable products that enterprise customers will pay for.

AMI enters a market converging around physical AI, systems that can perceive, predict, plan, and act in the real world. The competitive field spans frontier research labs, robotics foundation model companies, vertically integrated hardware players, and domain-specific autonomy platforms.

Google DeepMind is the largest large-scale rival. It has already shipped a family of embodied AI models under the Gemini Robotics umbrella and has established partnerships with robot OEMs including Boston Dynamics, Agility Robotics, and Apptronik. DeepMind's structural advantage is that it can treat physical action as an additional output modality layered onto an existing multimodal model stack, reusing Gemini's safety infrastructure, cloud distribution, and partner network. If enterprise buyers conclude that embodied reasoning can be served by extending a text-led foundation model into robotics, DeepMind becomes a compression force on AMI's differentiation.

Meta is simultaneously AMI's intellectual ancestor and a direct competitive threat. Meta has already released V-JEPA 2, a world model trained on over a million hours of internet video, capable of zero-shot robotic planning, and continues to advance the JEPA research lineage that AMI is commercializing. Meta's open research releases can commoditize parts of the world-model stack AMI wants to monetize, and Meta can recruit from the same talent pool with a larger balance sheet. The divergence is urgency and business model: Meta is a platform company for whom world models are one research thread among many, while AMI is entirely organized around turning this architecture into enterprise infrastructure.

World Labs is AMI's closest startup-category analog. Its Marble product is already publicly available and can reconstruct, generate, and simulate explorable 3D worlds from images, video, and coarse spatial inputs, with exports into formats useful for gaming, VFX, design, and robotics. World Labs is further along in showing a concrete developer-facing product surface, which gives it an early advantage in category definition and ecosystem building. The two companies are not identical in target: World Labs is currently stronger in spatial generation and world creation, while AMI is focused on prediction, planning, controllability, and safety for operational enterprise systems. But World Labs can move downstream into enterprise use cases faster than AMI can move upstream into productization.

Skild AI competes on the vision of a general-purpose intelligence layer for physical systems, but with a more deployment-driven go-to-market. Its Skild Brain robotics foundation model is already generating live revenue across security, construction, delivery, and warehouse deployments, and its flywheel is data-compounding: more robots in the field generate more training data, which improves the base model, which enables more deployments. AMI's research-first posture could yield deeper science but slower data accumulation relative to a company already in production.

Figure represents the thesis that the winning physical AI companies will bundle model, data collection, robot hardware, and deployment environment into a single integrated system. Its Helix AI stack is optimized around the Figure robot body, sensors, and logistics workflows simultaneously. For AMI, this creates a disintermediation risk: if customers in manufacturing and logistics prefer an integrated system provider over a standalone intelligence layer, vertically integrated players like Figure, alongside Anduril and Forterra in defense and industrial autonomy, can capture the value even if AMI contributes underlying conceptual advances.

NVIDIA occupies an ambiguous position. It is both an investor in AMI and a platform competitor through its Cosmos world foundation models, which offer prediction, controllable world generation, and physical AI tooling to the same robotics and automation customers AMI is targeting. NVIDIA's Cosmos ecosystem already counts Agility Robotics, Figure, and Skild as early adopters. NVIDIA benefits from a broad and competitive market of world-model companies built on its compute stack, it is not economically aligned to have AMI singularly win.

Domain-specific autonomy players like Wayve and Waabi show how world-model ideas become commercially legible when attached to a single high-value use case. Wayve's GAIA-3 is a generative world model for validating autonomous driving systems; Waabi World is a closed-loop simulator for training and testing the Waabi Driver. Both companies use world modeling as a practical lever for safety and validation in one vertical. The risk for AMI is that domain-specific players lock up the most valuable datasets, regulatory relationships, and customer trust inside verticals before AMI enters with a more horizontal platform.

Physical Intelligence adds further pressure at the robotics foundation model layer, having open-sourced its π0 model and weights, a move that accelerates commoditization of generalized embodied control and narrows the defensibility of any company whose moat rests primarily on having a world model rather than on proprietary data, deployment integration, or vertical-specific performance.

AMI's most immediate TAM expansion path is building upward from a base model into a full product stack. The planned dual-track architecture, hosted API plus downloadable, adaptable deployment, already implies multiple product tiers serving different customer profiles. Above that base layer, AMI can expand into simulation tooling, planning software, safety and evaluation frameworks, and control interfaces. Each of these layers addresses a distinct buyer need and a distinct budget line within industrial, robotics, and healthcare organizations.

The JEPA architecture supports this roadmap. Action-conditioned world models extend from representation learning into planning, which extends into safe control, which extends into workflow integration and monitoring. A company that starts by selling model access can progressively capture more of the value chain if it proves reliability in production environments.

Healthcare is AMI's most concrete near-term expansion opportunity because the Nabla partnership already provides a deployment channel. Nabla's presence across more than 150 health systems gives AMI a path into real clinical workflows without building its own distribution from scratch.

The healthcare wedge is attractive because it is a domain where the limitations of text-centric AI are most visible and most consequential. Clinical decision support, remote monitoring, wearable biosignal interpretation, and agentic care navigation all require systems that can model continuous multimodal data streams, simulate outcomes, and behave deterministically under safety constraints. Nabla's stated goal of building toward FDA-certifiable agentic AI systems points to the kind of reliability and controllability that AMI's architecture is designed to provide.

If the architecture proves out in healthcare, the same capability set extends into adjacent regulated workflows: medical device software, clinical operations, pharmaceutical process monitoring, and biomedical research automation.

AMI's stated target sectors, industrial process control, factory automation, robotics, aerospace, and automotive, represent a large underserved market for AI systems that can operate reliably in safety-critical, sensor-rich environments. These are the settings where text-based AI is a poor fit and where the procurement case for a world-model-native architecture is strongest.

The investor base creates optionality here. Toyota Ventures, Samsung, Dassault-linked capital, and Mulliez family interests all represent potential entry points into automotive, industrial manufacturing, aerospace, and consumer electronics supply chains. These are not passive financial relationships. They are potential design partners with proprietary sensor data, operational workflows, and high-value decisions that AMI's technology is built to improve.

The pattern established by companies like Forterra, a single autonomy platform sold across defense and commercial settings, and Skyfish, where sensor fusion and synchronized data collection become a durable moat, suggests that AMI's expansion path in industrial and defense markets is a verticalized stack rather than a generic API. One core world-model engine, adapted for specific environments, can address multiple high-value verticals without proportional R&D cost increases.

AMI is operating across Paris, New York, Montreal, and Singapore from day one, which is unusual for a company at seed stage. The Paris headquarters and geographically balanced investor base give AMI a credible position as a non-US, non-China alternative in frontier AI, a distinction that matters increasingly to European industrial customers, government procurement bodies, and regulated enterprises navigating data sovereignty and AI governance requirements.

The EU AI Act's full obligations for high-risk embedded-product applications come into effect in August 2027, which creates a tailwind for a company whose core pitch is reliability, controllability, and safety rather than raw capability. European industrial buyers in healthcare, manufacturing, and aerospace will face increasing compliance pressure to document AI behavior and demonstrate safe deployment, requirements that AMI's architecture is explicitly designed to satisfy. That regulatory environment expands AMI's addressable market in Europe at the same time that it raises barriers for competitors whose systems are harder to audit or constrain.

Proof risk: AMI's business rests on the claim that JEPA-style world models will outperform or meaningfully complement text-centric architectures in real-world prediction, planning, and control, but the company is still in a fundamental research phase and management has acknowledged that commercial validation may take several years. If the architecture proves effective in research settings but not measurably superior in production environments, AMI's commercialization timeline extends indefinitely while better-capitalized rivals continue shipping.

Data bottleneck: World models for physical environments are only as good as the sensor data they are trained and evaluated on, and AMI's target markets, industrial process control, healthcare, robotics, depend on proprietary, domain-specific data that AMI does not yet control. Value in these markets tends to accrue to whoever owns the best deployment loops and data pipelines, which could force AMI into partnership structures where it contributes core intelligence but captures only a fraction of the economics.

Vertical integration pressure: The most dangerous competitive dynamic for a horizontal intelligence-layer company is the rise of vertically integrated players who bundle model, hardware, data collection, and deployment into a single system. Figure in logistics, Anduril and Forterra in defense, and potentially Toyota-backed robotics platforms in manufacturing can all optimize across the full stack simultaneously, making it structurally difficult for AMI to win on system-level performance even if its underlying world models are technically superior.