AMI Labs Raised $1.03B to Kill the LLM Era

Estimated reading time: 10 minutes

Key Takeaways

• Then AMI Labs raised $1.03 billion in seed funding, and suddenly the whole AI industry had to pay attention.
• It is built around one core idea: the path to true machine intelligence runs through world models, not the text-predicting systems that dominate AI right now.
• This is not a small bet. It is the largest seed round ever recorded in Europe, and one of the largest pre-launch AI financings anywhere on the planet.
• This article breaks down exactly what AMI Labs is, what JEPA architecture actually does, why serious investors backed it at a $3.5 billion valuation before it had a product, and what this means for the future of artificial intelligence.

Yann LeCun spent years telling anyone who would listen that large language models are a dead end. Most people ignored him. Then AMI Labs raised $1.03 billion in seed funding, and suddenly the whole AI industry had to pay attention.

AMI Labs — short for Advanced Machine Intelligence — is the Paris-based startup LeCun co-founded after leaving Meta. It is built around one core idea: the path to true machine intelligence runs through world models, not the text-predicting systems that dominate AI right now. This is not a small bet. It is the largest seed round ever recorded in Europe, and one of the largest pre-launch AI financings anywhere on the planet.

This article breaks down exactly what AMI Labs is, what JEPA architecture actually does, why serious investors backed it at a $3.5 billion valuation before it had a product, and what this means for the future of artificial intelligence.

What Is AMI Labs and Why Did Yann LeCun Leave Meta

Yann LeCun’s world models vision takes shape at AMI Labs

LeCun is not a newcomer trying to get rich off AI hype. He is one of the three researchers who won the 2018 Turing Award — widely considered the Nobel Prize of computer science — for foundational work on deep learning. He spent years as Chief Scientist of Meta’s Fundamental AI Research (FAIR) lab, one of the most respected AI research groups in the world.

But Meta’s direction and LeCun’s vision kept drifting apart.

Meta struggled to turn FAIR’s research into practical technology. The company’s open-source model, Llama, faced its own headaches. There were internal tensions, including controversy around Meta’s acquisition of ScaleAI. And as the rest of the industry doubled down on larger and larger language models, LeCun grew increasingly vocal about why that path was wrong.

So he left. And he built AMI Labs.

The company is headquartered in Paris, with plans for offices in North America and potentially Asia. Alex LeBrun serves as CEO, while LeCun holds the role of Executive Chairman. At the time of its landmark funding round in March 2026, the company had roughly a dozen employees. It was less than three months old.

The fact that investors handed a 12-person company $1.03 billion says everything about how seriously the market is taking LeCun’s argument.

(Source: nytimes.com, siliconrepublic.com)

The $1.03 Billion Seed Round — By the Numbers

AI seed funding 2026 hits a new record with AMI Labs

Let’s be clear about the scale of this funding event, because the numbers are genuinely unusual.

• Round size: $1.03 billion
• Pre-money valuation: $3.5 billion
• Company age at close: Under 3 months
• Employee count: Approximately 12
• Classification: Europe’s largest seed round on record

The round closed in March 2026. It was co-led by five venture capital firms: Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions. But the investor list goes much deeper than that.

Strategic corporate backers include:

• NVIDIA — the dominant force in AI chip infrastructure
• Samsung — a global hardware and semiconductor giant
• Toyota Ventures — signalling interest in embodied AI and autonomous systems
• Temasek — Singapore’s sovereign wealth fund
• SBVA and New Legacy Ventures

And then there are the angel investors: Tim Berners-Lee (inventor of the World Wide Web), Eric Schmidt (former Google CEO), Mark Cuban, Xavier Niel, and Jim Breyer.

This is not a group of people who chase trends blindly. When NVIDIA and Toyota both write checks into the same seed round, it is because they see a concrete technical roadmap that could change what AI hardware and autonomous systems look like in five years.

Analyst coverage from Futurum Group frames this as the largest institutional bet on record that large language models alone cannot reach general intelligence. The investor mix — spanning traditional VCs, corporate strategics, sovereign capital, and high-profile angels — reflects a genuine industry hedge against LLM-centric approaches.

(Source: futurumgroup.com, reuters.com, startuphub.ai)

Why LeCun Says LLMs Are Not Enough

Yann LeCun’s world models argument targets LLM limitations

To understand why this funding matters, you need to understand the core argument LeCun has been making for years — and why he believes it is backed up by the architecture AMI Labs is building.

The problem with large language models:

LLMs like GPT-4, Claude, and Gemini are trained on vast amounts of text. They are extraordinarily good at predicting the next word or token in a sequence. That single capability, scaled up enormously, produces systems that can write essays, answer questions, generate code, and hold conversations.

But LeCun argues this approach has a hard ceiling. Here is why:

1. LLMs only know text. They have never seen the physical world. They have no internal model of how objects move, how forces work, or how actions lead to consequences.
2. They cannot truly reason or plan. When an LLM appears to reason, it is doing sophisticated pattern matching on text. It does not have a simulation of reality it can run forward to test possible outcomes.
3. They hallucinate because they lack grounding. Without a world model — an internal representation of how reality works — LLMs fill gaps with statistically plausible text, not verified facts.
4. They struggle with genuinely novel problems. Tasks that require physical intuition, multi-step planning in real environments, or understanding causal chains push current LLMs to their limits.

LeCun’s argument is not that LLMs are useless. It is that they are the wrong foundation for building AI that can act reliably in the real world. For that, you need a system that understands the world the way a child eventually does — by building an internal model of how things work.

(Source: taskade.com, startuphub.ai)

What Is JEPA and How Does It Actually Work

JEPA architecture is AMI Labs’ technical foundation

JEPA stands for Joint Embedding Predictive Architecture. It is the technical framework at the heart of AMI Labs, and it represents a fundamentally different way of training an AI model to understand the world.

The core idea:

Instead of predicting raw output — like the next word, or the next pixel in an image — JEPA trains an AI to predict in abstract representation space. Think of it this way: when you watch someone throw a ball, your brain does not try to predict the exact color of every pixel the ball will pass through. It builds an abstract internal model: “the ball will arc upward, reach a peak, and come down over there.” JEPA tries to make machines do something similar.

More technically:

• JEPA learns latent embeddings — compressed, abstract representations of states in an environment
• It then makes predictions in that latent space, rather than predicting raw sensory data
• This makes the learning process more efficient and the representations more meaningful for reasoning and planning

V-JEPA 2 and recent advances:

Newer variants of the architecture, such as V-JEPA 2, add components like 3D Rotary Position Embeddings (3D-RoPE) to stabilize training of billion-parameter models over temporal sequences. This is critical for robotics and embodied AI, where the system needs to understand how a scene evolves over time in three dimensions.

The formal mathematical proof:

In May 2026, AMI Labs published a preprint on arXiv delivering a landmark result: the first formal mathematical proof that JEPA can recover the hidden structure of an environment under specific statistical conditions — specifically, Gaussian latent variables with stationary additive-noise transitions.

This matters because it moves JEPA from a promising intuition to a provably grounded theory. It also honestly defines where current implementations still fall short, setting a clear technical roadmap for the work ahead.

Model Predictive Control (MPC):

One of the key applications JEPA world models are designed to enable is model predictive control with receding horizon planning. This is a control paradigm where a system uses its world model to simulate multiple future trajectories, evaluates them, and picks the best action — then repeats the process at every step. This is how you build an AI that can plan reliably in a changing environment, rather than one that just reacts.

(Source: startuphub.ai, taskade.com, startuphub.ai)

Where World Models Get Applied

Yann LeCun world models target robotics, autonomous driving, and agentic AI

The reason corporate giants like Toyota and NVIDIA joined this round is not abstract. World models built on JEPA architecture have specific, high-value applications that existing LLM-based systems handle poorly.

Robotics and embodied AI:

A robot navigating a factory floor needs to predict the consequences of its movements in real time. An LLM cannot do that. A world model trained with JEPA can build an internal simulation of the robot’s environment, predict what happens if it takes any given action, and plan accordingly. V-JEPA 2’s temporal and 3D capabilities are built specifically for this.

Autonomous driving:

Self-driving systems need to model how other vehicles, pedestrians, and road conditions will evolve moment to moment. World models that understand physical dynamics — not just statistical patterns in text — are far better suited for this than LLMs.

Smart glasses and wearable AI:

LeCun has spoken specifically about AI-powered smart glasses as an early consumer application. A system that can interpret your physical environment, understand spatial context, and give you relevant information in real time needs a world model, not a text predictor.

Industrial processes:

World models can simulate industrial systems, predict failures before they happen, and optimize complex multi-step processes. This is the kind of deep operational intelligence that automation-heavy industries have wanted for years.

Agentic AI platforms:

For AI agents — systems that take a sequence of actions to achieve a goal — reliable prediction of action consequences is essential. Current LLM-based agents frequently fail because they lack this grounding. World models give agents the ability to simulate outcomes before acting, making them far more reliable.

(Source: taskade.com, futurumgroup.com)

Stop Wasting Hours on Work That Machines Should Be Doing

While AI architectures battle it out in research labs, the businesses winning right now are the ones using automation to eliminate manual busywork today — not waiting for the next frontier model.

At Flowgrammer, we build custom business automation systems that remove repetitive tasks from your team’s plate, cut operational costs, and free up time for the work that actually moves the needle. No PhD required. No massive tech budget needed.

👉 See How Flowgrammer Automates Your Business

AMI’s Open-Source Philosophy and the Geopolitical Angle

AMI Labs’ open-source AI strategy challenges closed-model dominance

AMI Labs is not just making a technical bet. It is making a political and strategic one.

LeCun has been a vocal critic of the industry’s drift toward closed AI models. OpenAI started with an open-source philosophy and gradually closed its systems. Anthropic was built as a closed lab from the start. These companies argue that powerful AI must be restricted to prevent misuse. LeCun argues the opposite: that open-source AI makes the ecosystem safer, more diverse, and more resistant to any single entity having too much control.

AMI’s position:

• AMI is committed to open-source AI development
• LeCun views AI as a platform — like the internet itself — that must be open to be trustworthy and widely beneficial
• He argues that a diverse AI ecosystem, where models can be fine-tuned by different groups with different languages and value systems, is essential for global adoption

The geopolitical dimension:

LeCun specifically frames AMI as a European alternative to the US-China AI binary. Right now, the dominant AI narrative is a two-horse race between American labs (OpenAI, Google, Anthropic, Meta) and Chinese ones (DeepSeek, Baidu, Alibaba). AMI, headquartered in Paris, is positioning itself as a third option — one that is open-source, globally accessible, and not beholden to either superpower’s strategic interests.

This matters for governments, universities, and companies around the world that are uncomfortable depending on AI infrastructure controlled by either Washington or Beijing.

The open-source investor case:

Investors backing AMI have explicitly noted the strategic value of its open-source approach for startups that rely on open-source models. A thriving open-source world model ecosystem would benefit a wide range of downstream developers, much as Linux and Android created enormous value without being proprietary products.

(Source: bloomberg.com, startuphub.ai)

What the Critics Say — and Where AMI Still Has to Prove Itself

Not everyone is convinced that AMI Labs’ bet will pay off. The honest picture includes real challenges.

The skeptical case:

• JEPA’s formal proof relies on Gaussian latent variables and stationary additive-noise transitions — relatively clean mathematical assumptions that may not hold in the messy, non-stationary complexity of the real world. AMI’s own preprint acknowledges this gap.
• LLMs have kept improving faster than many critics expected. Systems that seemed to hit a wall two years ago found new performance jumps. Declaring them a dead end has been premature before.
• Training world models at scale is computationally intensive in different ways than LLM training, and the engineering challenges are not trivial.
• AMI has no product yet. $1.03 billion and a compelling theory is the starting line, not the finish line.

Where AMI has to deliver:

The company needs to demonstrate that JEPA-based world models can scale beyond the current proof’s assumptions, work reliably in real robotic and embodied AI applications, and attract the engineering talent to build production systems — not just research prototypes.

LeCun acknowledges that recruiting people who believe in the world models future is a central challenge. The talent pool that understands this architecture deeply is still small compared to the armies of engineers trained on transformer-based LLM development.

(Source: futurumgroup.com, startuphub.ai)

What This Means for the Broader AI Landscape

AMI Labs signals a maturing market that is seriously hedging beyond LLMs

AMI Labs’ $1.03 billion raise is significant not just for what it tells us about one company, but for what it reveals about where the broader AI industry is heading.

The LLM era is not over — but it is no longer uncontested.

The investor syndicate behind AMI includes some of the most sophisticated money in tech. When NVIDIA backs a company that is explicitly building an alternative to the LLM paradigm, that is a signal worth paying close attention to. NVIDIA makes money selling hardware for AI training regardless of which architecture wins. Their bet on AMI suggests they genuinely believe world models could become a major training workload.

Europe is asserting itself.

AMI being based in Paris and billing itself as an alternative to the US-China AI binary reflects a real and growing push for European AI sovereignty. With regulatory frameworks like the EU AI Act already reshaping how AI is deployed in Europe, having a major open-source AI research company on the continent has significant strategic value.

The robotics and embodied AI markets are ready.

Toyota, NVIDIA, and Samsung are not backing AMI because they are interested in making chatbots. They are backing it because they have enormous commercial exposure to robotics, autonomous systems, and hardware that needs to interact intelligently with the physical world. They are betting that world models will be the architecture that makes their next generation of products possible.

Open-source AI is getting serious capital.

The AMI round is one of the clearest signals yet that open-source AI development can attract institutional-scale funding. This challenges the assumption that only closed, proprietary labs can raise the capital needed to compete at the frontier.

(Source: futurumgroup.com, reuters.com, startuphub.ai)

Frequently Asked Questions