Spatial Intelligence: The AI Layer Behind Robotics, Games and VR, and Why Acquirers Are Paying for It

Most people met modern AI through chatbots. Those systems are trained to predict the next word, and they are very good at language. They are not built to understand where a chair is in a room, how far a robot arm must reach to pick up a cup, or what the back of a building looks like when you have only seen the front.

Spatial intelligence is the branch of AI that fills that gap. A spatially intelligent model can perceive a scene, rebuild it in 3D, generate new scenes that hold together from every angle, and let a person or an agent move and act inside them. The models that do this are usually called world models.

World Labs (opens in a new tab), the San Francisco lab founded by Dr. Fei-Fei Li, puts it simply: spatial intelligence is the next frontier in AI, and world models should reconstruct, generate and simulate 3D worlds that both humans and agents can interact with.

The field has moved from research demos to products in about a year. Two approaches lead, and both are useful to buyers for different reasons.

• Explicit 3D worldsWorld Labs' Marble, generally available since 12 November 2025, turns text, images, video or a rough 3D layout into a full 3D world. Worlds can be edited, expanded and combined, then exported as Gaussian splats, triangle meshes (including low-fidelity collider meshes for physics) or video. That output drops straight into existing engines and pipelines.
• Real-time generated worldsGoogle DeepMind's Genie 3 (opens in a new tab) generates interactive environments frame by frame from a text prompt, at 24 frames per second and 720p, staying consistent for several minutes. Users can trigger 'promptable world events' such as a change in weather, and DeepMind has used Genie 3 worlds to train its SIMA agent.
• Physical-AI platformsNVIDIA's Cosmos (opens in a new tab) platform packages world foundation models and tooling for robotics and autonomous-vehicle developers, who use them to generate and test synthetic training data at scale.

The common thread is that each of these systems produces space, not sentences. That is what makes them commercially valuable in industries that have always paid heavily to build, capture or simulate the physical world.

1. Autonomous robotics

Robots learn by trying things, and trying things in the real world is slow, expensive and occasionally dangerous. World models let developers generate unlimited training environments, from cluttered warehouses to volcanic terrain, and test 'what if' situations before a robot ever touches a real object. Genie 3's documentation explicitly frames this as a path to training and evaluating robots and autonomous systems.

2. Game worlds and physics

Building a high-quality 3D level can take a studio team months. Generative world models compress the first draft of that work dramatically: a designer can block out a scene with simple shapes, describe the style, and get a detailed, explorable world back. Marble's export of collider meshes alongside visual meshes is a direct bridge into the physics systems game engines already use. Studios, engine-tooling businesses and asset marketplaces all sit close to this shift. See our gaming M&A coverage for where consolidation is already heading.

3. Virtual and mixed reality

VR has always been constrained by content: headsets are only as compelling as the worlds available to step into. Spatial models that can lift a few photos or a short video into a navigable space make it possible to create environments for training, property, retail and entertainment on demand, rather than commissioning each one by hand.

On 18 February 2026, World Labs announced $1 billion in new funding (opens in a new tab) from investors including AMD, Autodesk, Emerson Collective, Fidelity, NVIDIA and Sea. The investor list itself tells the story: chipmakers, a design-software leader and a games-and-internet group all wanted a seat at the table.

In July 2026 World Labs made its own acquisition, buying SceniX (opens in a new tab). Then on 28 September 2026, AMD announced an agreement to acquire World Labs in an all-stock transaction valued at approximately $8.2 billion, with closing expected by the end of 2026. On reported valuations for the February round, the price is roughly 1.5 to 1.6 times the value set seven months earlier.

$1bn round closes

18 Feb 2026

World Labs acquires SceniX

21 Jul 2026

AMD deal announced

28 Sep 2026

Step-up on reported round valuation

~1.5x

The logic for AMD is straightforward: owning a frontier world-model team gives it first-hand insight into what spatial and physical-AI workloads will demand of its silicon. Our full breakdown of the terms, the VWAP share mechanics and the road to close is in AMD to Acquire World Labs for $8.2 Billion.

World models depend on three scarce inputs: rich 3D and spatial data, specialist research and engineering talent, and access to compute. When an input is scarce and strategic buyers are competing for it, the businesses that hold it gain leverage.

• Data ownersLibraries of 3D scans, game assets, mapped interiors, motion capture or robot interaction logs are training fuel. Clean rights and documented provenance turn them into a valuation driver.
• Tooling and pipelinesBusinesses that convert, render or deploy 3D content, such as splat renderers, mesh tooling and engine plug-ins, become the bridge between models and production use.
• Applied vertical playersSimulation, digital-twin, XR training and robotics software companies that can plug world models into a paying customer base are natural targets for both strategics and growth investors.

For acquirers, the opportunity is to secure spatial capability early, while many of the best teams and datasets still sit inside smaller, founder-led digital businesses. Acquiry runs buy-side mandates across AI, software, gaming and digital infrastructure. Discuss a mandate (opens in a new tab).