SAN FRANCISCO — When Fei-Fei Li’s World Labs placed its Atlas model in early access on September 1, the company released a demonstration reel and a blog post—but no research paper explaining how Atlas was trained.
It also disclosed no pricing and named no commercial partners. Three weeks later, the AI industry is still waiting for all three.
There is a parallel with Yann LeCun’s Advanced Machine Intelligence Labs. AMI raised $1.03 billion in March in what was described as the largest seed round in European history, then spent the following six months pursuing foundational research without offering a product to external developers.
AMI chief executive Alexandre LeBrun said in June that commercial applications of the company’s JEPA architecture would take “years.” Investors who valued the company at $3.5 billion before the financing appear to have accepted that timeline.
Together, World Labs and AMI Labs have raised more than $2.3 billion. The combined number of products commercially available today, priced and ready to bill a customer: one, the World API that World Labs opened in January 2026 at undisclosed commercial terms, and a Marble subscription at $95 a month. Atlas, the company’s most technically ambitious release, is “early access,” which in practice means no published pricing and no stated timeline for general availability.
This is not credibility the industry can easily dismiss. Both LeCun and Li are among the most accomplished researchers in the history of the field. LeCun co-invented convolutional neural networks, won the Turing Award, and spent decades as Meta’s chief AI scientist before departing last year to launch AMI. Li built ImageNet, the dataset that triggered the deep learning era, and later served as director of Stanford’s AI lab. Their credentials are precisely why the funding materialized at the scale it did.
The bet they are asking investors to make is specific. The core claim behind world models is that large language models are approaching a ceiling their architecture cannot break through. LLMs are trained to predict tokens: they learn what words follow other words with extraordinary sophistication, but have no direct experience of space, physics, or time. A world model is trained to predict how a scene evolves, how objects move, how a room looks from a different angle. That grounding, the argument goes, is what separates useful general intelligence from a very sophisticated autocomplete.
Atlas demonstrates what that difference looks like in practice. The model generates video at up to 1440p, controlled not by a text prompt but by a geometric camera path: the user specifies where the camera is and where it points, and Atlas renders what the scene should look like from that angle with physically consistent geometry. World Labs says it is multimodal from scratch, pretrained on text, images, video, and 3D data simultaneously. Whether it constitutes a genuine world model or a video generator with strong spatial awareness is a question the absence of a research paper makes impossible to resolve from the outside.
The strategic logic behind the secrecy is clear even if the companies will not state it plainly. Once you disclose a specific commercial direction, you invite capital into the same space. The early investors in AMI Labs, whose backers include Nvidia, Temasek, and Bezos Expeditions, funded a thesis and a team: the expectation is that the path to revenue clarifies over the next few years rather than immediately. Anthropic, which has built commercial products throughout its research program, is preparing for a potential IPO at a $2 trillion valuation. The gap between that disclosed revenue trajectory and the world model companies’ complete opacity is one of the defining tensions in AI funding right now.
LeBrun has not backed away from his timeline. At a Paris conference in May, he said that building a world model capable of grounding physical knowledge at the level AMI intends requires solving problems not yet solved, and that “quickly” means inside a few years, not months. LeCun, who has argued for years that LLMs cannot reach general intelligence and that the field needs a fundamentally different architecture, has offered no revision to that position. The broader debate over the pace of frontier AI development has not interrupted the research.
Whether the funding window stays open long enough to reach that point is the question neither company has answered. Nvidia doubled its chip output forecast for 2026, which means the infrastructure that makes world model training possible is spreading to organizations that did not need to raise a billion dollars to access it. The current advantage is time and talent, not access to compute. That window closes.
What Atlas can do is genuinely impressive. A photograph of a living room, a specified camera path, and a video of that room rendered from a moving angle that has never been captured: that is not a trick, and it is not something a language model produces. Whether it is a product, or a demonstration of a direction, is the question $2.3 billion has yet to answer.
