World models become AI's new race, from Niantic to Wayve

What world models are

A language model (LLM) takes a sentence and predicts the next word. A world model does something similar, but with images: it takes a visual state of the environment and predicts the next visual state, including physics, cause and effect. That's the definition given by Oliver Cameron, CEO of Odyssey, an AI lab dedicated to the topic, in an interview with Wired.

So with a language model, you have a sentence, and it will predict the next word in the sentence. With a world model, you give it a visual state of the world, an image, and it will then predict the next visual state of the world.

Oliver Cameron, CEO of Odyssey

The output of these systems isn't text or a label: it's a navigable 3D environment, generated from photos or video, that behaves as if it were real. Cameron sums it up this way: "they're simulations, but they look and feel real." For developers, this changes the kind of training data that enters the pipeline: instead of text corpora, video, sensor data and spatial trajectories.

From Pokémon hunting to home robots

Niantic, the creator of Pokémon Go, sold the game last year and redirected the business toward world models, naming the new company Niantic Spatial. The bet is to reuse the archive of real-world environment scans collected by the game's players to train robots to recognize obstacles and interact with objects.

We're very much about reconstructing the real world.

Brian McClendon, CTO of Niantic Spatial

According to McClendon, the proposed workflow is relatively straightforward: someone photographs the environment where a robot will operate, the software generates a 3D model of that space, and from it, "millions of simulations" are run to train the robot on that specific location's parameters, without needing real physical movement for much of the training cycle. For robotics teams, this reduces iteration cost: testing layout, lighting and obstacle variations in simulation is cheaper than repeating physical tests with every code change.

Cars drive through simulations of dangerous roads

Self-driving cars rely on the same principle, but with greater urgency: without a human driver, the software's reaction in fractions of a second is critical. Jamie Shotton, chief scientist at Wayve (a British startup that builds world models for vehicles), explains the core technical requirement.

Importantly, a world model must be able to react to the actions that we take in the world, and so it must understand physics, cause and effect, and how other objects and people change their actions in response to our own.

Jamie Shotton, chief scientist at Wayve

Wayve's generative model is fed, in part, by dashcam footage. The company uses these recordings to recreate dangerous roads and unsafe conditions that would be impractical to test in real life with an actual car. It's a pattern that repeats across the sector: the raw data comes from the physical world captured on video, and the model learns to reconstruct and extrapolate variations of that scenario.

Heavy money moving into the sector

Commercial interest in world models has grown fast. Yann LeCun, Meta's former chief AI scientist, raised more than a billion dollars to fund his new startup, Advanced Machines Intelligence, focused on the topic. Fei-Fei Li, a computer scientist known for her work in computer vision at Stanford, founded the startup World Labs in 2024; the company was acquired by AMD this year for more than 8 billion dollars.

Google keeps announcing improvements to its own world models, and Tencent is also investing in the technology applied to games, since physics and 3D rendering are the core of any video game graphics engine. The investment pattern signals that world models have stopped being niche research and have become a product line with dedicated infrastructure budgets.

What changes for those who build software

In short: if world models move out of the lab and into production in home robotics and self-driving cars, the type of infrastructure that supports applied AI changes profile. Some concrete points that follow from what the sources describe:

• Data pipelines shift to video and sensor, not text. Teams that today organize text datasets for fine-tuning will increasingly deal with raw video, dashcam footage and 3D scanning, with very different storage and preprocessing requirements.
• Physical simulation enters robotics' CI/CD cycle. Running "millions of simulations" before any physical deployment, as McClendon describes, requires asynchronous testing pipelines designed for 3D environments, not just unit tests for code.
• GPU demand changes shape. Training and running inference on models that generate continuous visual states tends to be more computationally heavy than equivalent text LLMs, which puts even more pressure on the already tight data center capacity supply for those operating AI infrastructure in Brazil.
• Billion-dollar acquisitions concentrate proprietary technology. AMD's purchase of World Labs for more than 8 billion dollars suggests that a significant part of this technical stack will be born closed, inside big techs, rather than as an open tool available to any engineering team.

The job market that's already changing

For a world model to exist, it needs real people recording the real world. The Wired report recounts that the author himself, earlier this year, strapped an iPhone to his forehead and recorded hours of video in his apartment, folding laundry and washing dishes, selling the data through an app with "pretty modest" pay. It's a sign that, as more labs try to build better world models, more gig work capturing everyday data is likely to emerge, a layer of cheap, mostly invisible labor sustaining physical AI infrastructure.

What's still open

Wired points to 2027 as a possible year of significant expansion for home robotics and self-driving cars, both powered by advances in world models, but that's still a prediction, not an accomplished fact. Questions such as the provenance of the video data used in training, fair pay for those who record that material, and benchmark standardization among competing labs (Niantic Spatial, Wayve, Odyssey, World Labs under AMD) remain publicly unanswered. For those deciding today which AI stack to invest engineering time in, the clearest signal is the volume of capital flowing into the area, not yet a set of mature, open tools ready for immediate adoption.