Research report AMI Labs

AMI Labs

Advanced Machine Intelligence

World Models as the Post-LLM Architecture Bet

August 2026

Important notice: This document has been prepared for informational and educational purposes only. It does not constitute investment advice, a research recommendation, or an offer or solicitation to buy or sell any security. AMI Labs is a private company; there is no public market for its shares, and the seed round described below was not, and this note is not, an offer to any retail investor. All figures are drawn from public reporting as of early August 2026 and may be incomplete, imprecise, or superseded by subsequent disclosure. Readers should treat every valuation figure as a reported data point, not a verified fact, and should consult a qualified financial and legal adviser before making any investment decision.

1. Executive Summary

AMI Labs (Advanced Machine Intelligence) is a Paris-headquartered artificial intelligence research company founded by Yann LeCun following his departure from Meta in late 2025, where he had spent twelve years as chief AI scientist. The company's founding thesis, stated plainly by LeCun and reiterated by its executive team, is that large language models (LLMs) are structurally incapable of reaching human-level intelligence, however much they are scaled, and that the next architectural generation of AI must be built instead on “world models” — systems that learn how physical reality behaves and use that understanding to plan and reason, rather than predicting the next word in a sequence.

The company closed a $1.03 billion (€890 million) seed round in March 2026 at a $3.5 billion pre-money valuation, reported to be the largest seed round in European technology history and one of the largest of any kind globally. The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions, with a strategic and individual investor base spanning Nvidia, Toyota, Samsung, Groupe Industriel Marcel Dassault, Eric Schmidt, Mark Cuban, and others.

AMI Labs is, by its own leadership's description, a research organisation first and a product company second. Chief executive Alexandre LeBrun — who transitioned from running medical-transcription AI company Nabla — has stated publicly that AMI does not expect to ship a commercial product for years, distinguishing it from a typical applied AI startup that reaches revenue within months. This report sets out what is known about the company's technology, team, capitalisation, market context, and principal risk factors, to support an informed view of the opportunity it represents.

1. Company Overview & History

Reports of LeCun's departure from Meta and a prospective new venture surfaced in mid-December 2025, when the Financial Times reported that LeCun was in preliminary talks to raise approximately €500 million for a venture provisionally named Advanced Machine Intelligence Labs, targeting a €3 billion valuation. LeCun formally confirmed the venture in a LinkedIn post that same month, naming Alexandre LeBrun — then CEO of Nabla — as AMI's chief executive, with LeCun himself taking the role of Executive Chairman rather than CEO.

The company remained largely in stealth through January and February 2026, during which press coverage focused on LeCun's public criticism of the AI industry's direction — most notably a January 2026 New York Times profile in which he characterised Silicon Valley's LLM-scaling strategy as a herd mentality heading toward a dead end. AMI Labs formally unveiled itself and its funding round in March 2026, describing its mission as building AI systems that understand the world, hold persistent memory, can reason and plan, and remain controllable and safe.

Since the funding announcement, the company has continued to expand its research team and has begun publishing technical work. In late May 2026, researchers associated with LeCun's group posted two related preprints: one offering a formal proof of the conditions under which the JEPA architecture can recover a faithful model of the world, and a companion benchmark paper finding that current implementations remain brittle under minor visual perturbations. Taken together, these represent the company's most substantive public research output since its founding and the clearest available evidence of technical progress against its stated thesis.

é. Management & Team

AMI Labs has assembled a senior team that combines LeCun's research standing with operational leadership drawn from adjacent ventures, and has recruited a number of high-profile researchers into founding technical roles.

Yann LeCun

Founder & Executive Chairman

2018 Turing Award; 12 years as Chief AI Scientist, Meta; originator of JEPA architecture

Alexandre LeBrun

Chief Executive Officer

Co-founder & former CEO, Nabla (medical AI); former Facebook AI research engineer

Laurent Solly

Chief Operating Officer

Former Meta VP for Europe

Saining Xie

Chief Science Officer / Co-founder

Prominent computer vision researcher

Pascale Fung

Chief Research & Innovation Officer / Co-founder

AI researcher, prior academic leadership roles

Michael Rabbat

VP, World Models

Research leadership in self-supervised learning

This is a notably research-heavy leadership bench relative to typical seed-stage companies, reflecting the company's stated identity as a research organisation. LeBrun's dual position — stepping away from an operating CEO role at a revenue-generating company to run a pre-product research venture — is itself a signal of how the company frames its near-term priorities: scientific validation ahead of commercialisation. Nabla, the company LeBrun previously led, has become AMI's first disclosed commercial-adjacent relationship (see Section 7), though no formal equity or licensing agreement between the two has been disclosed as of this writing.

3. Technology: World Models and JEPA

AMI Labs' technical thesis rests on the Joint Embedding Predictive Architecture, or JEPA, a framework LeCun first proposed in a 2022 position paper, “A Path Towards Autonomous Machine Intelligence.” Its central design choice is to predict the future in an abstract latent representation space rather than by reconstructing raw pixels or generating tokens. LeCun's argument is that most of what happens in a video — the precise flutter of leaves, the exact texture of noise, unpredictable microscopic detail — is inherently unpredictable, and that forcing a model to reproduce that detail wastes capacity and degrades the quality of what it learns. A representation-space predictor, by contrast, is trained to discard the unpredictable and retain only what matters for understanding cause and effect.

This positions JEPA as a direct architectural alternative to both autoregressive LLMs and generative video models such as Sora-style systems, rather than an incremental improvement on either. Public research under the JEPA umbrella — image-based I-JEPA, video-based V-JEPA and V-JEPA 2, and more recent extensions — originated largely during LeCun's tenure at Meta and continues to be developed by researchers now at AMI Labs. Reported results include V-JEPA 2 achieving materially higher zero-shot robot control success rates than competing approaches after training on internet-scale video supplemented with a comparatively small amount of robot trajectory data, evidence LeCun's camp cites as validation that representation-space prediction generalises efficiently.

The unresolved technical debate

The architecture is not uncontested within the research community. At a March 2026 academic forum, LeCun's abstraction-first approach was directly challenged by researchers who argue that predicting in latent space without any reconstruction mechanism risks losing contact with reality — in effect, that a system can become internally consistent without remaining accurate. The more considered technical commentary in the field treats this as a genuine, unresolved question rather than a settled advantage for either camp, with a growing view that hybrid architectures combining abstraction and reconstruction may ultimately be required.

The company's own research output partially substantiates this open-question framing rather than resolving it: the May 2026 preprints from LeCun's group establish the theoretical conditions under which JEPA can recover a faithful world model, while the companion benchmark finds that today's implementations remain brittle under minor visual shifts — in other words, the theory and the current state of the art are still meaningfully apart.

4. Funding History & Capitalisation

AMI Labs' fundraising trajectory shows a rapid escalation from initial ambition to closed outcome, compressed into roughly three months.

Dec 2025

Preliminary talks reported (Financial Times)

Target: ~€500m raise at ~€3bn valuation

Dec 19, 2025

LeCun confirms venture publicly

Reported target: $5bn+ valuation

Mar 9-10, 2026

Seed round closes; company unveiled

$1.03bn raised (~€890m); $3.5bn pre-money valuation

The final round materially exceeded the initial funding target disclosed in December, both in absolute size and in the pool of investors it drew. According to CEO LeBrun, strong investor interest allowed the company to select backers for strategic and philosophical alignment rather than purely on price. The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions, with broader participation including Nvidia, Toyota, Samsung, Groupe Industriel Marcel Dassault, and Eric Schmidt, alongside individual backers such as Tim and Rosemary Berners-Lee, Jim Breyer, Mark Cuban, Mark Leslie, and Xavier Niel.

Two aspects of the capitalisation are worth noting for anyone assessing the company's position. First, the round was structured and priced as a single seed financing rather than a staged Series A/B progression, which is unusual and reflects the scale of capital the founding team's reputation was able to attract at inception. Second, the company disclosed no revenue, and none is implied by the round's structure; the valuation reflects a bet on the team and the underlying research thesis rather than on any demonstrated commercial traction.

5. Market Opportunity

AMI Labs does not compete in an established, well-defined market category; it is pre-positioning itself at the center of a nascent one, generally described by third-party research as “physical AI” or “embodied AI” — AI systems that perceive, reason about, and act within physical or simulated environments, spanning robotics, autonomous systems, industrial automation, healthcare, and immersive media.

Third-party market-sizing estimates for this category vary enormously depending on scope and methodology, which is itself a relevant data point about how immature and loosely defined the category still is. Commercial research providers place the current embodied/physical AI market anywhere from under $1 billion to over $80 billion depending on definition, with forecasts for the next decade ranging from roughly $15 billion to nearly $1 trillion by the early-to-mid 2030s, and compound annual growth rates quoted between 18% and 47%. Such a wide dispersion of published estimates should be read as a sign that the category lacks a stable, agreed definition rather than as evidence of any specific number.

Analyst note: given this dispersion, any single total-addressable-market figure applied to AMI Labs should be treated as illustrative rather than a reliable basis for valuation. The more defensible framing is qualitative: AMI is positioning for a category that multiple well-capitalised research groups and market forecasters agree is directionally large and growing quickly, without yet agreeing on its size, boundaries, or timeline to materiality.

Within that broad category, AMI's specific angle — general-purpose world models based on representation-space prediction, rather than a single vertical application — places it in direct thematic competition with a small number of similarly ambitious, similarly well-funded research efforts rather than with an established incumbent industry (see Section 7).

6. Competitive Landscape

AMI Labs' most directly comparable competitor is World Labs, founded in 2023 by Stanford professor and ImageNet co-creator Fei-Fei Li, which pursues a related but distinct thesis under the label “spatial intelligence.” World Labs emerged from stealth in September 2024 with $230 million at a $1 billion valuation, and by February 2026 had raised an additional $1 billion — including a $200 million strategic investment from Autodesk — from backers including Nvidia, AMD, Fidelity, and Emerson Collective, bringing total disclosed funding to roughly $1.23 billion. Reports in early 2026 indicated the company was in talks for a valuation of approximately $5 billion.

The comparison is instructive on both similarity and difference. Both companies argue that language models lack a native understanding of physical structure — geometry, persistence, occlusion, causality — and both have raised unusually large sums relative to disclosed commercial traction. World Labs, however, has already shipped a commercial product, Marble, a platform for generating persistent, editable 3D environments from text, images, or video, alongside a research preview of a real-time world engine; AMI Labs, by contrast, has shipped no product and has stated it does not expect to for years. World Labs' near-term go-to-market also concentrates on creative and design tooling (gaming, visual effects, architecture, via its Autodesk partnership) rather than the more general, safety- and reasoning-oriented framing AMI has used.

Beyond these, AMI's own leadership acknowledges other, smaller efforts in the same conceptual space, citing France-based SpAItial's $13 million seed round as an example of continued investor appetite for world-model startups at a much smaller scale. The competitive set is therefore bifurcated: a handful of extremely well-capitalised efforts (AMI, World Labs, and the world-model research arms of Google DeepMind and Meta itself) pursuing the thesis at frontier-lab scale, and a longer tail of smaller, more narrowly scoped startups.

LeCun's departure from Meta is itself part of the competitive story. Meta's AI strategy is now directed under Alexandr Wang, following Meta's investment in and partnership with Scale AI, a shift in leadership direction that press reporting has characterised as the proximate cause of LeCun's exit. Meta retains the V-JEPA research lineage LeCun originated and continues to invest in it, meaning AMI Labs' most direct technical lineage is, in effect, competing against its own former home.

7. Business Model & Path to Revenue

AMI Labs has disclosed no formal business model, pricing strategy, or product roadmap with commercial dates attached. CEO LeBrun has been explicit that the company should not be judged against typical applied-AI startup benchmarks — shipping a product within three months, generating revenue within six, or reaching $10 million in annual recurring revenue within a year — describing AMI's work instead as starting from fundamental research, with a multi-year path to commercial application.

The one disclosed commercial-adjacent relationship is with Nabla, the medical-AI documentation company LeBrun co-founded and previously led. Nabla has stated it will have early access to AMI's technology as it develops, and LeCun is a Nabla investor, but no formal equity stake or licensing agreement between the two companies has been disclosed. Nabla's own CEO transition — LeBrun's move to AMI left Nabla under co-founder and COO Delphine Groll on an interim basis — adds a layer of organisational complexity to what is otherwise the clearest available signal of where AMI's technology might first reach a commercial setting: healthcare documentation and decision support, an application area where hallucination-prone LLMs carry acknowledged safety risk, and where a more grounded, world-model-based approach could plausibly command a premium.

Beyond healthcare, the world-model thesis points logically toward robotics, autonomous systems, and industrial automation as the eventual commercial destination, consistent with how third-party market researchers frame the broader physical/embodied AI category (Section 6). None of these have been confirmed as AMI product lines as of this writing.

8. Risk Factors

Prospective investors or partners evaluating AMI Labs should weigh the following factors, drawn from public disclosures and independent reporting:

Technical/thesis risk. JEPA remains an actively contested architecture within the research community. AMI's own May 2026 research shows a meaningful gap between the theoretical conditions under which the approach should work and the brittleness of current implementations. The core scientific bet has not yet been validated at the scale the company's valuation implies.

Timeline and monetisation risk. Company leadership has explicitly set expectations for a multi-year path to any commercial product, with no disclosed revenue or near-term monetisation plan. This is a materially longer and less certain horizon than is typical even for well-funded applied AI startups.

Valuation risk. The $3.5 billion pre-money valuation was set at seed stage, prior to any disclosed product, revenue, or independently verified technical milestone, on the basis of team reputation and thesis conviction. This is a valuation methodology inherently more exposed to sentiment and narrative than one grounded in operating metrics.

Competitive risk. AMI faces well-capitalised competitors pursuing adjacent or overlapping theses, including one (World Labs) that has already shipped a commercial product and secured strategic commercial partnerships, and two (Google DeepMind and Meta) with substantially greater compute and distribution resources, the latter also retaining the original research lineage AMI is built on.

Key person risk. The company's identity, fundraising narrative, and scientific credibility are closely tied to Yann LeCun personally, and its operating leadership to Alexandre LeBrun. Departure or reduced involvement of either would likely have an outsized effect on investor and partner confidence relative to a more conventionally diversified leadership structure.

Governance and disclosure risk. As a private company, AMI Labs is not subject to public-company disclosure obligations. Figures in this report, including funding amounts and valuations, are drawn from press reporting rather than audited or regulator-filed data, and are subject to revision.

Macro/AI-capital-cycle risk. AMI's fundraising success occurred within a broader environment of exceptionally large AI-related capital raises across multiple companies and architectures; a broader repricing of AI-sector risk appetite would likely affect AMI's ability to raise follow-on capital on similar terms, independent of its own technical progress.

9. Valuation Considerations

AMI Labs' $3.5 billion pre-money seed valuation should be read in the context of comparable transactions rather than in isolation. World Labs, pursuing an adjacent but distinct thesis, priced its initial 2024 round at a $1 billion valuation with a shipped-product-free company, then reportedly sought approximately $5 billion in valuation talks roughly sixteen months later, ultimately raising $1 billion in February 2026 without publicly confirming the resulting valuation. Viewed against that trajectory, AMI's $3.5 billion figure — achieved at first institutional financing, with no product at all — represents a materially faster and steeper valuation escalation than its closest comparable, which some commentary in the technology press has characterised as reflecting investor conviction in the founding team as much as in the technology itself.

A conventional valuation framework based on revenue multiples, discounted cash flow, or comparable operating metrics cannot meaningfully be applied to AMI Labs at this stage, since none of the standard inputs — revenue, margin, user metrics, or a demonstrated product — exist yet in disclosed form. The valuation that has been placed on the company is better understood as a venture bet on a specific team and a specific scientific thesis, priced at a scale usually reserved for companies with demonstrated product-market fit. Any future financing round, and any eventual monetisation event, will likely hinge on whether the company's research — such as the May 2026 identifiability and brittleness papers — continues to show credible progress narrowing the gap between JEPA's theoretical promise and its practical performance.

This section describes how AMI Labs' valuation compares with disclosed market data points; it is not a valuation opinion, price target, or recommendation, and should not be relied upon as such.

10. Summary Considerations

AMI Labs represents one of the largest and most closely watched wagers in the current AI landscape: a scientifically credible, well-capitalised, and clearly differentiated bet that the industry's dominant architecture is heading toward a structural ceiling, made by one of the researchers most responsible for the field's foundations. Its principal strengths are the depth and standing of its founding and research team, the scale and quality of its investor base, and a technical thesis (JEPA) with a multi-year publication record predating the company's founding.

Its principal open questions are, in order of materiality: whether the underlying architecture can be shown to work reliably at scale within a timeframe that satisfies its investor base; how the company will monetise a technology it has explicitly said will take years to commercialise; and how it will differentiate and compete against both a nimbler, already-shipping rival pursuing an adjacent thesis (World Labs) and the vastly larger compute and distribution resources of Google and Meta, the latter of which continues to develop the very research lineage AMI was built to extend beyond.

For a reader evaluating AMI Labs as a potential capital allocation, the honest framing is that this is, at the current stage, a research-conviction investment rather than a traditional growth-stage or product-metrics-based one, and should be sized and diligenced accordingly.

Selected Sources

TechCrunch (Dec 2025, Jan 2026, Mar 2026) — coverage of AMI Labs' founding, confirmation, and seed round; Bloomberg (Mar 2026) — AMI seed round reporting; Bloomberg (Jan 2026) — World Labs valuation talks; MIT Technology Review (Jan 2026) — exclusive interview with Yann LeCun; The New York Times (Jan 2026) — “An A.I. Pioneer Warns the Tech ‘Herd’ Is Marching Into a Dead End”; STAT News (Mar 2026) — AMI/Nabla relationship; Latent Space / AI News (Mar 2026) — team and launch detail aggregation; TechCrunch (Feb 2026) — World Labs/Autodesk financing; Turing Post and independent technical commentary (2026) — JEPA architecture explainer and critique; TechTimes (May 2026) — AMI research group identifiability and brittleness preprints; third-party market research aggregators (Grand View Research, MarketsandMarkets, Research and Markets, and others, 2026) — physical/embodied AI market sizing.