Free tools Windows power users keep installed
One-click scans. No signup required.
Yann LeCun’s AMI Labs has announced a $1.03 billion seed round at a $3.5 billion pre-money valuation. The company says it will use the financing for fundamental research into “world models”—AI systems intended to learn from reality and reason about physical environments, not only generate text.
AMI is led by chairman Yann LeCun and CEO Alexandre LeBrun. It has not announced a product launch date; the company describes useful commercial applications as a multiyear prospect.
The financing: amount, valuation and backers
TechCrunch reported the financing on March 9, 2026, as a $1.03 billion seed round at a $3.5 billion pre-money valuation. The publication also gave an approximate euro equivalent of €890 million. Singapore’s Economic Development Board (EDB) separately described the financing as S$1.31 billion, equivalent to US$1.03 billion, in its March 10, 2026 account.
| Detail | Reported information |
|---|---|
| Round | Seed financing |
| Amount | $1.03 billion; Singapore EDB reports S$1.31 billion (US$1.03 billion) |
| Valuation | $3.5 billion pre-money |
| Approximate euro equivalent | €890 million, as reported by TechCrunch |
| SBVA commitment | €30 million |
| Announcement window | March 9–11, 2026, across the cited company, investor and government accounts |
Who invested
The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions. Reported participants also include Nvidia, Samsung, Sea, Temasek and Toyota Ventures, along with other investors.
Recommended Free Tools
#1 Best Overall
AMI’s leadership and locations
LeCun is chairman and LeBrun is chief executive officer. AMI has planned locations in Paris, New York, Montreal and Singapore. The Singapore presence is part of the country’s stated role in the company’s expansion and industrial collaboration plans.
What AMI means by “world models”
AMI uses “world models” for systems that learn from observations of reality, build internal representations of environments and seek to reason about how those environments behave. The emphasis is physical-world understanding: objects, space, actions, cause and effect, and the consequences of decisions.
That goal differs from an AI system trained primarily to predict the next token in text. A language model can describe a robot or a factory, but text descriptions alone do not give it a reliable, continuously updated model of what happens when a robot moves, an object is grasped or a machine state changes.
Rank #2
LeBrun told TechCrunch, “We are developing world models that seek to understand the world, and you can’t do that locked up in a lab. At some point, we need to put the model in a real-world situation with real data and real evaluations.”
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →JEPA and self-supervised learning
AMI’s technical direction uses self-supervised learning and Joint Embedding Predictive Architectures, commonly called JEPAs. In a JEPA-style approach, a model learns to predict a useful representation of missing or future information rather than reproducing every detail pixel by pixel or selecting the next word in a sequence.
Self-supervised training can draw learning signals from data itself, reducing reliance on manually labeled examples. For a world model, the intended result is a representation that captures meaningful structure and dynamics in an environment. AMI has not published benchmark scores or accuracy figures establishing how well its systems currently perform.
World models versus large language models
The distinction is best understood as a difference in learning objective and grounding, not as a claim that one category makes the other obsolete. A future system could combine language and world-model components.
| Comparison | World-model direction described by AMI | Typical large language model |
|---|---|---|
| Learning objective | Learn predictive representations of environments, including aspects of physical dynamics, using approaches such as JEPA | Predict the next token from patterns in language data |
| Grounding | Observations and other real-world data intended to represent physical settings | Primarily text interaction, even when text contains descriptions of the physical world |
| Evaluation | Planned tests with real data, partners and evaluations in practical settings | Often assessed first with language and other offline benchmarks |
| Potential deployment | AMI has indicated that paid APIs or downloadable and adaptable models could eventually follow the research phase; no commercial release is confirmed | Commercial APIs and downloadable models already exist across the industry |
| Time to market | AMI characterizes the work as fundamental research whose useful applications may take years | Many language-model products are already available, although capabilities and reliability vary by use case |
How AMI plans to use the money
Advance fundamental research
The primary use is research into world-model architectures, self-supervised learning and JEPA-based methods. AMI is positioning the company as a research organization first, rather than announcing a near-term software subscription.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchTrain and evaluate models with real data
AMI intends to move beyond laboratory demonstrations by testing models with real-world data and practical evaluations. Nabla is the first disclosed partner. The partnership signals an effort to expose models to operational conditions, although no public performance results or product specifications have been released.
Open-source part of the work
LeBrun told TechCrunch, “We will also make a lot of code open source.” That is an intention rather than a published release schedule, so the scope, licenses and timing of future code releases remain unspecified.
Develop industrial proofs of concept in Asia
SBVA said its €30 million investment will support proof-of-concept initiatives with robotics and manufacturing companies in Asia. These projects could provide controlled environments for testing perception, prediction and decision-making against physical processes, but the companies, project specifications and results have not been publicly detailed.
When will AMI’s technology be available?
There is no confirmed launch date for an AMI commercial product. The company says useful applications may take years because it is pursuing fundamental research and plans to validate systems in real environments before broad deployment.
Best Value
AMI’s stated path is therefore staged:
- Research: develop and train world-model systems using self-supervised and JEPA-based methods.
- Partner evaluation: test models with real data and disclosed collaborators such as Nabla.
- Industrial proofs of concept: explore robotics and manufacturing applications, including initiatives supported by SBVA in Asia.
- Possible products: consider paid APIs or downloadable, adaptable models if the technology reaches a useful level of reliability.
Those stages should not be read as a product roadmap with fixed dates. AMI has not announced pricing, access requirements, public benchmarks or a general-availability release.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What is known—and still unknown
- Known: the financing size, pre-money valuation, major backers, leadership, planned locations and broad research direction.
- Known: AMI plans real-world evaluations, has identified Nabla as its first disclosed partner and intends to open-source substantial amounts of code.
- Known: SBVA plans to support robotics and manufacturing proofs of concept in Asia.
- Not established publicly: model accuracy, benchmark performance, revenue, customer count, headcount, compute scale or production deployments.
- Not announced: a confirmed commercial launch date, API pricing, downloadable model license or list of available products.
Why this round matters
The financing gives AMI the resources to pursue a research problem that requires more than text corpora: learning representations that remain useful when an AI system must predict and act in a changing physical environment. The investor group combines technology, semiconductor, automotive, manufacturing and regional investment interests, while the planned sites span Europe, North America and Asia.
The important near-term test is not the size of the valuation but whether AMI can demonstrate that its models transfer from training data to reliable behavior in real settings. Its own evaluation plans, the Nabla collaboration and the SBVA-backed industrial proofs of concept are the clearest milestones currently disclosed.
LeBrun has predicted that “world models” will become the next AI buzzword. For AMI, the term will need to translate into measurable performance in physical environments before it becomes a mature commercial category.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
