3 Startups That Are Building World Models, AI’s Next Frontier

Language models have long been the darlings of the AI industry. But an increasing number of researchers believe that merely predicting the next word will not result in pushing machine intelligence forward in new ways. 

Today, world models are emerging as an advanced way to achieve intelligence through AI. Three startups – Decart, AMI Labs, and World Labs – are at the forefront of this game.

World Models as the Next Leap in AI Development

World models are multimodal, and they make sense of physics, spatial relationships and consequences of actions, all of which LLMs can’t handle. While LLMs focus on predicting text patterns, any improvement in their performance will come through additional training data and computational power, often with decreasing marginal gains.

World models continuously learn from their surroundings, as people do, adjusting to new information and optimizing output accordingly. This is precisely what differentiates world models from LLMs as we move towards AGI. 

An LLM might be able to generate an image of a staircase, but a world model will know what happens when a ball bounces down a staircase, because it’s familiar with gravity, movement, and friction. Emerging use cases for world models range from robotics and healthcare to self-driving cars and gaming.

Here are three startups that have developed world models that push the boundaries of AI’s capabilities.

Decart: Real-Time World Generation at Scale

Founded by Israeli intelligence whiz kids Dr. Dean Leitersdorf and Moshe Shalev in 2023, Decart’s AI research laboratory came out of stealth in late 2024 with the release of Oasis. 

This model is essentially a playable AI-generated world that demonstrates unusually low latency and real-time responsiveness, signaling a leap toward immersive AI experiences for gaming, simulation, robotics and training applications. Within just three days of Oasis’s launch, its user base exceeded one million globally. Decart’s Lucy models apply style effects and can replace objects and backgrounds to and live video feeds continuously, while remaining faithful to spatial relationships and physics.

What sets Decart apart is its approach, which combines proprietary infrastructure optimization with its own generative models, allowing the company to reduce the massive compute costs typically associated with world model output. Their efficiency engine has attracted enterprise clients, generating ongoing revenue from day one. 

Decart has also partnered with Amazon to run models through AWS’s Trainium accelerators, and has launched a new distributed hyperlocal compute grid in partnership with Comcast and NVIDIA.

Investors are betting heavily on that advantage. Decart has raised more than $150 million from backers including Sequoia Capital, Benchmark and Zeev Ventures. In its most recent funding round, a $100 million Series B, Decart was valued at $3.1 billion.

AMI Labs: The Turing Award Bet Against LLMs

When Turing Award winner Yann LeCun left his post as Meta’s longtime chief of AI, the industry wondered what he’d do next. From one of the most influential AI labs in Big Tech, he went on to launch a startup, AMI Labs (Advanced Machine Intelligence). The venture is his direct bet that large language models will prove insufficient as a response for true machine intelligence. 

The Paris-based startup leverages the Joint Embedding Predictive Architecture (JEPA), which was created by LeCun when he was still at Meta. This architecture allows its algorithms to learn abstract models of reality rather than predict outcomes token by token as large language models do.

Unlike Generative AI, JEPA was designed for the uncertainties of the real world. Hence, it’s most applicable in healthcare and robotics situations, where there’s less leeway for LLMs’ mistakes. LeCun has been claiming since 2022 that LLMs will not result in AGI, and AMI is the manifestation of that idea.

In March 2026, AMI secured $1.03 billion in funding at a valuation of $3.5 billion, the biggest AI seed funding ever in Europe, with support from NVIDIA, Bezos Expeditions, Jeff Bezos, Mark Cuban, and Eric Schmidt. 

French med-tech firm Nabla has become the first corporate partner of AMI and received early access to its models for clinical applications. AMI plans to publish its research, which is unusual for any frontier lab but strategically important for JEPA to establish a global research community.

World Labs: Training AI to See and Create

World Labs was established in 2024 by Fei-Fei Li, known as the “godmother of AI” because of her computer vision research and invention of ImageNet, along with three other Stanford professors. 

From Li’s perspective, spatial intelligence is the dimension missing from AI. The inability to reason and visualize three-dimensional environments would continue to be an insurmountable barrier for any AI system, no matter how proficient its language skills were.

Marble, World Labs’ first commercial product, allows users to create persistent and editable three-dimensional worlds from text prompts, images, videos, and 3D designs. These can be exported as either a mesh file or a video format. The product is intended for applications in gaming, architectural design, visual effects, and robotics. 

What makes Marble unique compared to other products in the market is its persistence and ability to be edited.

Li’s legitimacy is not limited to the lab alone. She was selected to be one of Time Magazine’s “Architects of AI” as part of their 2025 Person of the Year selection. Moreover, she was appointed as a member of the UN Scientific Advisory Committee, where she influences how conversation lands are modeled in global policy.

In 2024, WorldLabs raised $230 million in funding at a unicorn valuation, followed by a second round of funding worth $1 billion in February 2026. This was supported by NVIDIA, AMD, Autodesk, and Andreessen Horowitz, with Autodesk leading the investment at $200 million.

The Race Has Just Begun

Decart, AMI Labs, and World Labs each exemplify their own approach to entering the world models paradigm, in consumer experience, research fundamentals, or enterprise spatial intelligence. What all three have in common is a clear understanding that the future of AI belongs to world interaction, rather than word prediction.