World models: who is building and buying them

“World Models Will Soon Be Inescapable,” a special report Wired published on November 8 (by senior writer Reece Rogers), walks through the field of world models: systems that take visual input instead of text and generate digital 3D environments that obey physics — show one a visual state of the world and it predicts the next. The report’s working picture is that three industries are now funding this work in earnest: robot training, self-driving and video games. Yann LeCun, Meta’s former chief AI scientist, has raised more than $1 billion for his startup Advanced Machines Intelligence, per the article; Niantic Spatial and Wayve appear with on-the-record voices, and Tencent is mentioned in passing. On where the robot money lands, our earlier piece on Tesla renting Optimus out instead of selling it covers the demand side — the closer humanoid robots get to shipping, the better the economics of simulated training look.

Key points

The one-sentence definition, from Odyssey CEO Oliver Cameron: with a language model you predict the next word; “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.” His summary: “They’re simulations, but they look and feel real.”
Three named buyers: Niantic Spatial wants to use Pokémon Go-era environment scans to “reconstruct the real world” and help robots navigate; Wayve builds its generative world model partly from dashcam footage to recreate hazardous roads; and Tencent, Wired says, is investing because physics and 3D rendering are core to games.
The money: Wired says LeCun’s Advanced Machines Intelligence recently raised over $1 billion, and writes that World Labs was acquired by AMD “this year for over $8 billion” — a deal we independently verified at announcement in September.

Background

Predicting the world visually is not a November idea; this site has tracked the lane for weeks. Google DeepMind shipped Genie 4 on October 27, a real-time world model with native audio and mid-session rewriting. On October 12, 杨立昆 presented the H-JEPA paper in Seoul, lifting Visual AntMaze success from 18% to 73% — the first reproducible evidence for his “world models, not bigger LLMs” position. His company itself raised earlier and bigger: Wired reported on March 10 that AMI pulled in more than $1 billion at a $3.5 billion valuation, co-led by investors including Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions, with Mark Cuban among the other backers. And the AMD–World Labs deal is not gossip to relay: both companies announced it on September 28 — roughly $8.2 billion in an all-stock transaction, with Fei-Fei Li joining AMD as EVP and chief scientist. That is a confirmed fact with two sources, not a rumor, even though the new Wired piece compresses it into one clause.

The facts

The Wired report carries no benchmarks; its checkable material is people, money and use cases:

Player	Wired’s account	Independent cross-check
LeCun’s AMI	“over a billion dollars” raised recently	March: $1B+ at $3.5B valuation (Wired)
World Labs	AMD “acquired this year for over $8 billion”	Announced Sept 28: ~$8.2B all-stock
Niantic Spatial	Pokémon Go scans to help robots navigate; sold game in 2025	No independent cross-check, Wired only
Wayve	Generative world model powered partly by dashcam footage	No independent cross-check, Wired only
Tencent	Investing in its own world models; physics and 3D are core to games	No independent cross-check, Wired only

The quotes, one by one: Niantic Spatial CTO Brian McClendon says “We’re very much about reconstructing the real world.” Oliver Cameron runs the language-model comparison quoted above. Wayve chief scientist Jamie Shotton gives the strictest version: “a world model must be able to react to the actions that we take in the world” — it must understand physics, cause and effect, and how others respond to our actions. Tencent gets one sentence, no direct quote: since physics and 3D rendering are core to games, “major players like Tencent are investing in their own world models.” The report also expects 2027 to be a major year for home robotics and driverless cars, both bolstered by world-model advances.

What others say

Read the report closely and a sampling bias shows: all four on-the-record voices are builders explaining what world models will do; nobody in the piece pushes back. The cold water lives elsewhere, so we went and got it. In the Hacker News thread on Genie 4, practitioners picked glitches apart frame by frame and priced out input latency, warning that game studios haven’t even digested plain LLMs yet. LeCun himself supplies the sharpest crossfire — calling the idea that scaling LLMs reaches human-level intelligence “complete nonsense” — but his camp is equally unable to date when world models ship at scale. Wired’s distinct contribution is downstream: the author strapped an iPhone to his forehead, recorded hours of chores and sold the footage through an app for what he calls a “fairly measly” payout. The world-model data economy has grown a gig-work layer.

Our take

The report’s value is not exclusive facts — we covered both money events when they happened — but the map it draws: data collection, models, compute and buyers now have names attached across robotics, self-driving and games. For developers the practical read is that scanned real-world scenes are about to appear on simulation-data shopping lists; Niantic Spatial and World Labs (Marble, now inside AMD) are both selling that. Two honest caveats. First, AMD’s acquisition of World Labs is a confirmed announcement, but Wired compresses it to one clause — the deal terms (all-stock, regulatory approval pending, expected to close by year-end) are in our September entry. Second, the robot-training vision in the piece — photograph a space, generate a 3D model, run “millions of simulations” — has no published numbers behind it, and none of the four sources offers a timeline. Whether 2027 is that major year is checkable in twelve months.