On a crisp March morning in Paris, Advanced Machine Intelligence—better known as AMI Labs—did something few European startups have managed: it closed a seed round that would make even Silicon Valley blink. The figure: $1.03 billion, at a $3.5 billion pre-money valuation. The date: March 9, 2026, according to company announcements. The ambition, perhaps more audacious than the check itself: to prove that the future of artificial intelligence lies not in parsing ever-more text, but in understanding the physical world through sensor data.
It is, by any measure, Europe's largest seed round on record. And it arrives at a moment when some investors—flush with capital but wary of another wave of chatbot startups—are hunting for what comes after large language models.
The syndicate reads like a who's who of global tech finance. Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions co-led the round. Strategic players piled in: Nvidia, Temasek, Toyota Ventures, Sea, Publicis Groupe, Samsung, Bpifrance Digital Venture. The angel roster includes Eric Schmidt, Jim Breyer, Mark Cuban, Xavier Niel, and—notably—Tim and Rosemary Berners-Lee. SBVA committed €30 million.
For a company starting with roughly 10 employees, it's an extraordinary mandate.
When Text Isn't Enough
AMI's thesis is straightforward, if technically daunting. Language models learn from text—billions upon billions of words scraped from the internet. World models, the term AMI uses, learn from something closer to raw experience: video feeds, sensor streams, the continuous flow of data that constitutes how robots, wearables, or autonomous systems perceive their environments.
The distinction matters. Text is discrete, structured, relatively clean. Sensor data is messy, high-dimensional, relentless. Training models on it requires massive computational resources and, AMI's founders argue, fundamentally different architectures.
Guiding the technical vision is Yann LeCun, the company's Executive Chair. LeCun, who shared the 2018 Turing Award for pioneering work in deep learning and spent years as Meta's Chief AI Scientist, has long championed approaches that go beyond language. His presence lends AMI both credibility and a certain intellectual heft—though whether his theories translate into deployable products remains an open question.
CEO Alexandre LeBrun is under no illusions about the timeline. In an interview with TechCrunch, he conceded that "world models" has become something of a buzzword, one that risks overpromising. Real-world evaluation, he stressed, takes years, not months.
LeBrun's track record suggests patience born of experience. He co-founded Nabla, a healthcare AI company, and before that sold Wit.ai to Facebook in 2015 for what insiders described as a tidy sum. He's seen hype cycles before.
Building a Research Machine
AMI is launching with a senior team drawn from academia and industry. Laurent Solly, formerly Meta's VP for Europe, serves as COO—a choice that signals operational ambition beyond pure research. Saining Xie, a respected figure in computer vision, holds the Chief Science Officer title. Pascale Fung is Chief Research & Innovation Officer, while Michael Rabbat, who led research efforts at Meta, oversees world models development as VP.
The company's first model—referred to as "AMI Video," according to Sifted—targets robotics, manufacturing, wearables, and healthcare. AMI has already inked its first partnership: Nabla, LeBrun's former venture, secured an exclusive deal in December 2025 for early access to AMI's healthcare-focused world model technology. The arrangement raises questions about conflicts of interest, though both companies insist the terms are arms-length.
AMI's geographic footprint spans Paris, New York, Montreal, and Singapore. The company plans to grow from 10 to somewhere between 30 and 50 employees within six months—a measured pace, especially given the capital at hand. LeBrun has committed to publishing research papers and releasing significant portions of AMI's codebase as open source. It's a posture that positions the company as a research lab first, a product company second.
Whether that approach wins over enterprise customers remains to be seen.
The Compute Arms Race
Where is the $1.03 billion (roughly €890 million) going? Mostly into two buckets: compute infrastructure and talent. World models demand enormous processing power. Unlike language models, which work with discrete tokens, world models process continuous streams of sensory input—video frames, audio waveforms, spatial coordinates. The computational overhead is staggering.
AMI's emphasis on compute also reflects a broader industry reality: as AI models grow more ambitious, the cost of training them has become a primary competitive moat. Access to Nvidia chips, in particular, has become a strategic asset. Nvidia's participation as an investor likely includes preferential access to hardware—though neither party would confirm specifics on the record.
The funding puts AMI in direct competition with World Labs, the startup founded by AI luminary Fei-Fei Li, which raised $1 billion in February 2026, including a $200 million strategic investment from Autodesk. The back-to-back mega-rounds suggest something like a sector rotation among AI investors: a pivot from language models toward systems designed for physical reasoning and spatial understanding.
Perhaps the LLM boom that defined 2023 through 2025 is giving way to something else. Or perhaps this is simply the next wave of capital chasing the next grand promise.
Europe's Moment?
For a European startup, AMI's valuation and investor roster are uncommon. The continent has produced plenty of successful tech companies, but few that command Silicon Valley-style valuations at the seed stage. AMI's Paris headquarters—and its commitment to research openness—may appeal to researchers wary of the concentrated power of a handful of AI labs in California.
Then again, talent flows where the resources are. And for all its ambition, AMI remains unproven. The company has published no papers yet, shipped no products, demonstrated no working prototypes to the public. What it has is capital, pedigree, and a bet that the next frontier in AI requires machines that don't just read the world—they experience it.
Whether that bet pays off will depend on execution, not just vision. And in AI, as LeBrun himself acknowledged, the timeline stretches out in years. The investors have placed their chips. Now comes the hard part.
