Imagine asking an AI system to redesign a factory, generate a walkable city or teach a robot to navigate an unfamiliar kitchen. The promise of world models is that such systems could understand not only what an image contains, but how spaces work, how objects behave and what might happen next. Yet the companies building this technology are raising money and forming partnerships while saying remarkably little about the products they plan to bring to market.
A technology searching for its first business
World models are designed to give artificial intelligence an internal representation of the physical or digital world. Instead of responding only to patterns in text or pixels, a world model could learn that a cup can be picked up, that a vehicle needs room to turn and that moving through a crowded room requires anticipating other people's actions.
That ability could support a wide range of industries. Robots might use world models to practice tasks in simulated environments before attempting them in warehouses or hospitals. Autonomous vehicles could predict how pedestrians, cyclists and other cars will move. Manufacturers could test factory layouts without rebuilding physical equipment. Game developers and filmmakers could create interactive settings that respond consistently to a user's actions.
The same breadth that makes the technology attractive also makes the market difficult to read. A company developing a general world model might eventually sell software to robotics companies, license simulation tools to manufacturers or provide creative systems for media studios. It could also become a supplier to larger artificial intelligence platforms.
For now, many of the companies involved are not saying which path they intend to take.
AMI Labs, co-founded by Meta's former chief AI scientist Yann LeCun, is among the most closely watched entrants. Michael Rabbat, another co-founder, told TechCrunch that the company remains in a research and building phase. AMI Labs is not discussing product plans or timelines, leaving investors and potential customers to infer its direction from the company's technical ambitions and the reputation of its founders.
World Labs, led by Stanford professor Fei-Fei Li, has offered a clearer glimpse of its work. Its Marble system can generate three-dimensional worlds from visual inputs, creating scenes that users can explore and adapt. Demonstrations have pointed toward media creation, game development and visual effects.
Those examples show how world models might become creative tools. A designer could describe a location and then move through it as if scouting a film set. A game studio could generate a landscape that remains coherent when players enter buildings, change objects or return hours later. A visual effects team could manipulate a scene from different camera angles without constructing every detail by hand.
But the demonstrations do not fully explain how World Labs intends to build a durable business, particularly in robotics. A model that produces visually convincing environments is not automatically capable of controlling a machine in the physical world. Robots need reliable information about weight, friction, balance, timing and failure. A simulated object may look correct while behaving incorrectly when grasped.
The suppliers cannot see the destination
The lack of clarity extends beyond the startups themselves. Companies supplying data and technical infrastructure to world model developers may know that their work is valuable without knowing what it will ultimately enable.
Alex de Vigan, CEO of Physicl, told TechCrunch that his company understands its data has been useful to customers but has limited visibility into the applications those customers are developing. That uncertainty is revealing. It suggests that the commercial ecosystem is forming before its participants understand the final shape of the products.
Data could become one of the most important resources in this competition. World models may require enormous collections of video, three-dimensional scans, simulated interactions and real-world sensor recordings. They could also need carefully labeled examples showing how objects move, how environments change and which actions produce useful results.
Access to such data may determine which companies can move from impressive demonstrations to dependable systems. A startup with a clever architecture but weak training material could struggle to match a rival with years of proprietary recordings from factories, vehicles or robots. At the same time, collecting physical-world data raises questions about privacy, copyright, safety and consent.
Why secrecy makes strategic sense
The companies' silence is not necessarily a sign that they have no commercial plans. It may reflect the unusual competitive landscape around world models.
A company that announces a narrow product too early could reveal its most valuable insight to potential rivals. If it says it is building software for warehouse robots, competitors can recruit talent, target the same customers and reproduce the product strategy. If it describes a platform for interactive video, larger firms such as OpenAI or Anthropic may have enough resources to enter the market quickly.
Keeping the destination hidden allows startups to build foundational technology while preserving flexibility. They can test several markets, gather feedback privately and wait until a system is reliable enough to support a stronger launch.
Investors appear willing to fund that approach because the potential markets are enormous. World models could become an infrastructure layer beneath robotics, simulation, autonomous systems and generative media. In that scenario, the eventual winners might not sell a familiar application at all. They could provide the underlying intelligence that other businesses use to build their own products.
That possibility has created what some observers describe as a dark forest dynamic. Companies raise substantial capital, limit public disclosures and work quietly until their technology is difficult to copy. Public attention focuses on funding announcements and demonstrations, while the more consequential work happens behind closed doors.
The missing evidence
The challenge for the public is separating long-term potential from current business traction. At present, there is limited information about deployment schedules, revenue models, operating costs and technical limitations across much of the sector.
A compelling generated environment does not prove that a system can guide a robot safely. A model that understands a short video may not maintain a stable representation of a complex factory over several hours. A system that works in simulation may still fail when lighting changes, surfaces become slippery or people behave unpredictably.
These gaps matter because world models are being discussed as if they could become a foundation for the next generation of artificial intelligence. That may happen, but the path will depend on practical details that remain largely undisclosed.
The first meaningful products may therefore arrive gradually rather than as a single dramatic breakthrough. Designers could use world models as creative assistants. Engineers might employ them to test layouts and machinery. Robot developers could use them for limited training tasks before trusting them with physical operations.
The future promised by world models is easy to picture: machines that understand spaces, anticipate consequences and collaborate with people in environments that once required direct human judgment. What remains difficult to picture is the business that will deliver it. Until these companies reveal more about their customers, economics and constraints, world models will remain one of the most promising and least transparent corners of the AI market.
This article was written with the assistance of an AI system and published automatically.