Frontier Rounds
AMI Labs logoFunding timeline

AMI Labs

Advanced Machine Intelligence Labs is Yann LeCun's world-model company, founded after he left Meta's chief AI scientist role in late 2025. It builds on his JEPA framework instead of language-first models, on scientific credibility alone, with no product shipped.

AMI Labs logo
Founded
Nov 2025
Headquarters
Paris (NY, Montreal, Singapore)
Founder / Chairman
Yann LeCun (Turing Award 2018; ex-Meta Chief AI Scientist)
CEO / Co-founder
Alexandre LeBrun (ex-CEO of Nabla; Wit.ai founder; former FAIR engineering lead)
Sector
World models / JEPA / embodied & physical AI research
Latest valuation
US$3.5B pre-money
Team
~12 at announcement
Rounds recorded
1
Total raised
$1.03B
amilabs.xyz

Funding Timeline

Every disclosed equity round for AMI Labs, newest to oldest, with lead and participating investors where reported. Data is stored oldest-first; the timeline is displayed latest-first.

  1. Mar 9, 2026

    Seed

    $1.03BLatest
    Led by Bezos ExpeditionsLed by Cathay InnovationLed by GreycroftLed by Hiro CapitalLed by HV CapitalNVIDIASamsungToyota VenturesTemasekSeaSBVANew Legacy VenturesBpifrance Digital VentureAlpha Intelligence CapitalAssociation Familiale MulliezGroupe Industriel Marcel DassaultPublicis GroupeAglaé Lab (Kering)Artémis (Pinault)ZEBOX VenturesJeff BezosEric SchmidtTim Berners-LeeJim BreyerMark CubanMark LeslieXavier Niel

    Europe's largest seed financing at the time (LeCun: 'probably the largest for a European company'), announced weeks after Ineffable Intelligence's record and before it. Oversubscribed well beyond the €500M target in the leaked late-2025 deck; LeBrun said AMI could select investors for alignment, not just capital. Pure equity, no disclosed debt component. Priced at US$3.5B pre-money (official); post-money ≈US$4.5B is an inference, no official figure.

    Source ↗

The Story Behind the Numbers

LeCun and the JEPA Bet

After 12 years at Meta — five as founding FAIR director, seven as chief AI scientist — LeCun left to industrialize his 'Advanced Machine Intelligence' research program. The core wager is JEPA: instead of predicting the next token or pixel, systems learn abstract representations of real-world sensor data and predict in representation space, ignoring the noise generative models are forced to reproduce. AMI wants AI that 'understands the real world, has persistent memory, can reason and plan, and stays controllable and safe' — a direct rebuke to what LeCun calls the industry's language-first monoculture.

A Scientific-credibility Valuation

At US$1.03B on a US$3.5B pre-money, investors paid primarily for scientific credibility and long-term option value: a company of roughly a dozen people, no product, and a research agenda measured in years. Structurally AMI differs from Thinking Machines Lab's US$2B single-lead seed: it assembled five co-leads spanning US and European VC plus a coalition of strategics and star angels — reading more like a coalition of patrons than a conventional round. The strategics were arguably the message: NVIDIA signals infrastructure, Toyota Ventures and Samsung point to industrial and device applications, Temasek mirrors the Singapore hub.

Team and Multi-hub Structure

AMI operated from day one across Paris (HQ), New York, Montreal and Singapore — deliberately built around where its leaders live: LeCun teaches at NYU, CSO Saining Xie came from Google DeepMind, and Singapore serves both talent and future Asian industrial clients. The founding roster is a Meta/FAIR diaspora: CEO Alexandre LeBrun (Nabla, Wit.ai), COO Laurent Solly (Meta VP Europe), CRIO Pascale Fung. LeCun remains a NYU professor, and Paris was welcomed by President Macron — positioning AMI within Europe's sovereign-AI narrative as an explicitly 'neither American nor Chinese' frontier lab.

Roadmap, Proof, and Risk

LeBrun is candid that AMI is 'not your typical applied AI startup' — commercial applications could take years, with early industrial targets in robotics, manufacturing and process control (partners scouted in South Korea mid-2026). The first rigorous public output came in May 2026, when LeCun and co-authors posted a formal proof that LeJEPA recovers a world's true latent variables (verified in Lean 4). Independent benchmarks of early JEPA preprints still show fragility under visual distractors — underscoring that the framework must close a large gap between theory and reliable real-world deployment before the next financing round tests the thesis.

More in the Atlas

AMI Labs's lead investors are profiled in the Investor Atlas when their firms join the directory. Track who backs every frontier AI round on the Funding Leaderboard.

Frequently asked questions

Why did AMI Labs split its seed across five co-leads?

Because LeBrun said the firm could choose investors for alignment rather than chase a single cheque: five co-leads — Bezos Expeditions, Cathay Innovation, Greycroft, Hiro, HV — bracket US and European VC. The oversubscribed close ran to US$1.03B, more than double the leaked deck's €500M target, and held Europe's seed record for about seven weeks before Ineffable's US$1.1B edged past it.

What justifies a US$3.5B pre-money for a lab of roughly twelve people?

Credibility priced as the whole asset: no product, a research agenda measured in years, and a Meta/FAIR diaspora led by a Turing Award winner. The cross-checks bracket it — Ineffable, also pre-product, cleared a ~US$5.1B post on David Silver's record, while World Labs, with products already shipped, sat near a reported ~US$5B. AMI's number only gets tested when JEPA theory meets deployment.

Why is a LeCun lab headquartered in Paris with hubs in three other cities?

Because it was built around where its leaders live: LeCun teaches at NYU (hence New York), and Singapore serves Asian industrial clients and talent. The map is also the message — welcomed by President Macron and framed as 'neither American nor Chinese,' AMI shares Paris with Mistral's sovereign stack, making the city the only European capital hosting two frontier-flagship bets.

Both AMI Labs and World Labs say 'world models' — what actually differs?

Method and maturity. AMI's JEPA predicts in abstract representation space and refuses to reproduce pixel-level noise; its first rigorous output was a Lean 4–verified proof (May 2026) that LeJEPA recovers true latent variables. World Labs generates persistent, downloadable 3D environments — Marble, Atlas — already in users' hands. A theory-first wager against a product-first company.

Why are Samsung, Toyota and Temasek in a pre-product research lab?

As mapped options: NVIDIA signals infrastructure, Toyota Ventures and Samsung point at robotics, manufacturing and devices, Temasek mirrors the Singapore hub — the roster doubles as a map of where JEPA would first be sold. Higgsfield's Series B runs the same play in video (Intel Capital, NTT DOCOMO, Mirae Asset), corporate tiers pre-positioning around AI categories before revenue exists.