AMI Labs and JEPA: The $1.03 Billion Architecture Bet That Language Models Are a Dead End

On March 10, 2026, Yann LeCun announced that Advanced Machine Intelligence Labs raised $1.03 billion in seed funding at a $3.5 billion pre-money valuation, making it the largest seed round in European startup history. Every major outlet covered the money. Almost none explained the architecture. AMI is not building a better language model. It is building a fundamentally different type of AI system based on LeCun’s Joint Embedding Predictive Architecture (JEPA), and the technical differences between JEPA and autoregressive language models determine whether this billion-dollar bet pays off or evaporates.

What LLMs Actually Do (and Why LeCun Says It Is Wrong)

Large language models predict the next token in a sequence. Given “The cat sat on the,” GPT-5.4 calculates probability distributions over its vocabulary and selects “mat” or “couch” or “floor.” This autoregressive prediction operates in discrete token space, generating output one subword at a time, left to right.

LeCun has argued for years that this approach has structural limits. Token prediction optimizes for plausible text, not for understanding the world that text describes. When an LLM writes a paragraph about physics, it is selecting statistically likely word sequences, not reasoning about physical systems. The hallucination problem is, in LeCun’s framing, a direct consequence: a system trained to produce plausible text will sometimes produce plausible text that happens to be false, and it has no mechanism to tell the difference.

This is a contested claim. GPT-5.4 scored 83% on GDPval across 44 professional occupations. Claude Opus 4.6 leads agentic coding benchmarks. These are real capabilities produced by token prediction. LeCun’s position is not that LLMs are useless. It is that they will never produce genuine understanding of the physical world, and that genuine understanding requires a different architecture.

How JEPA Works

JEPA operates in a continuous embedding space rather than discrete token space. Instead of predicting “the next word,” JEPA predicts abstract representations of what comes next. The distinction matters at the mathematical level.

In an autoregressive LLM, the model outputs a probability distribution over all possible tokens. In JEPA, the model outputs a vector in a learned embedding space that represents predicted features of future input. The prediction target is not the raw data itself (pixels, words, sensor readings) but an abstract encoding of that data. This is what LeCun means by “predicting in representation space rather than pixel space.”

The architecture uses two networks. An encoder processes the current input into an embedding. A predictor takes that embedding and produces a predicted embedding for what comes next. A separate target encoder processes the actual next input into its own embedding. The system trains by minimizing the distance between the predicted embedding and the target embedding. There is no decoder that reconstructs raw data. The system never tries to generate pixels or words. It only tries to match abstract representations.

The hardest engineering problem in this design is representation collapse. If the system can minimize its loss by mapping every input to the same embedding vector, it will. Earlier self-supervised methods like SimCLR and BYOL fought collapse using contrastive learning: explicitly pushing apart representations of different inputs. JEPA avoids contrastive pairs entirely. Instead, the target encoder updates its weights as an exponential moving average of the main encoder, creating a slowly shifting prediction target that the main encoder must continuously chase. Getting this balance right is where the engineering difficulty lives, and it has not been validated at production scale.

AMI claims this design prevents hallucination in the LLM sense. A generative model producing tokens can produce plausible but false output. A model predicting only abstract features does not generate human-readable output at all. JEPA-based systems need additional components to translate embeddings into actions or descriptions, and those downstream components can be constrained in ways raw text generation cannot.

What AMI Is Actually Building

AMI’s stated goal is AI systems for robotics, healthcare, and industrial applications where physical world understanding matters. The first disclosed partnership is with Nabla, a French clinical AI company where CEO Alexandre LeBrun previously worked. Key hires include Saining Xie (formerly Google DeepMind), Mike Rabbat (formerly Meta FAIR research director), and Pascale Fung (formerly Meta senior director of AI research). LeCun serves as executive chairman while remaining a professor at NYU.

The company will operate across Paris, New York, Montreal, and Singapore. LeBrun stated publicly that the first year will focus entirely on research, with product timelines measured in years, not quarters. AMI plans to publish papers and release code as open source, continuing the open research philosophy LeCun championed at FAIR. The open-source commitment differentiates AMI from OpenAI’s closed approach and aligns with LeCun’s long-standing public criticism of proprietary AI development.

What Could Go Wrong

JEPA has never been validated at the scale AMI is proposing. Meta released V-JEPA for video understanding and I-JEPA for image understanding, with promising results on specific benchmarks. But no JEPA-based system has been deployed at production scale. The gap between “interesting research direction” and “system that works in a hospital” is measured in years of engineering, not months of scaling compute.

The company has no product, no revenue, and no near-term prospect of either. At current compute costs, $1.03 billion buys roughly 18 to 24 months of serious research before AMI needs either results or another raise. Investors are betting on LeCun’s conviction that the entire LLM approach will hit a ceiling. If LLMs continue improving at their current pace (and GPT-5.4’s benchmark numbers suggest they might), the window for an alternative architecture narrows. Every quarter that autoregressive models post gains on professional-work benchmarks is a quarter where AMI’s thesis looks harder to prove.

The team quality is not in question. LeCun shared the 2018 Turing Award. Xie and Rabbat are established researchers. The risk is structural: a research-first startup with a multi-year timeline, zero revenue path, and a thesis that contradicts the demonstrated capabilities of the industry’s dominant approach.

AMI also enters a crowded “world model” space. Fei-Fei Li’s World Labs raised over $1 billion for spatial intelligence. SpAItial secured $13 million in European seed funding for 3D world models. Meta’s FAIR lab continues internal JEPA research. None have shipped a production system, which makes this the most expensive unvalidated thesis in machine learning. The question of who owns the core JEPA intellectual property, given Meta funded the original research, LeCun published it as open science, and AMI now builds on it commercially, remains unaddressed.

Why It Matters Either Way

AMI represents the most well-funded test of a specific hypothesis: that AI grounded in physical world understanding will outperform text prediction for real-world tasks. The competition between architectures is intensifying, and whether LLMs are sufficient or merely impressive will determine which companies dominate the next decade.

If LeCun is right, the current LLM approach is a local maximum and AMI is building the path to the next one. If he is wrong, AMI is the most expensive academic research lab in Europe. Either way, the architectural question is real, the talent concentration is unusual, and the bet is now large enough that the outcome will be visible.

Sources: JEPA framework (2022). TechCrunch. Crunchbase. TNW. OpenAI.