General Intuition's $2.3B Bet on Video Game Data for AI

I’ve followed General Intuition since their initial $134M raise last year — the startup trains AI agents on video game clips from Medal’s 10 million monthly users. The pitch: use gameplay data to teach AI how to move through the physical world. My first reaction was the same as everyone else’s. Does watching someone play Fortnite actually help a robot cross an office?

Turns out yes, and the numbers back it up. General Intuition’s latest round pulled in $320 million at a $2.3 billion valuation, led by Khosla Ventures with Jeff Bezos, Eric Schmidt, and General Catalyst participating, bringing total disclosed funding to $454 million. In a TechCrunch demo, the same model that played Fortnite for 100 hours straight powered a quadrupedal robot that learned a brand-new office in eight minutes of real-world data. That last number is the whole story: traditional robotics needs thousands of hours of physical data collection. This needed eight minutes.

If you follow AI agents at all, you know the bottleneck isn’t intelligence. It’s spatial understanding. LLMs write code and answer questions, but they can’t work a door handle.

Why doesn’t regular video work for training spatial AI?

Regular video shows what happened, never why. Most companies building spatial AI agents train on ordinary footage — people walking, driving, handling objects — and hope the model infers how the world works.

It doesn’t, because video hides the inputs. When you watch someone walk through a door, you don’t see the weight shift, the hand position, the decision to push versus pull. You see the result without the action. General Intuition’s CEO Pim de Witte describes it as the difference between “watching someone drive and actually feeling the steering wheel.” Video teaches pattern recognition; it can’t teach causality, the link between an action and its consequence.

That gap is why most spatial models fall apart in novel environments. They memorized patterns from training data without learning the underlying logic, so a different room layout breaks them.

What makes gameplay data different?

Every Medal clip comes with action labels attached. When a player records gameplay, the capture logs each button press and joystick movement, timestamped alongside the footage.

So instead of seeing “the character walked through a doorway,” the model sees “the player held W for 1.2 seconds, moved the mouse 15 degrees right, then pressed E on the door handle.” That’s a fundamentally different training signal — closer to following a recipe with exact measurements than watching a cooking show. General Intuition trains on hundreds of millions of hours of these annotated clips, which teaches the model what de Witte calls the difference between “self” and “environment”: how its own actions change the world around it.

That’s the mechanism behind the transfer. Because the model learned cause and effect rather than surface patterns, the spatial reasoning that worked in Fortnite also worked on a robot. During a TechCrunch visit to their R&D floor, one model did both.

How did a Fortnite model end up on a robot?

Here’s the demo that sold me. General Intuition showed TechCrunch reporters an agent that had played Fortnite continuously for 100 hours. They loaded that same model onto a quadrupedal robot with a single camera. The robot walked the office, steered around chairs, and fine-tuned itself to the space with just eight minutes of real-world data.

Eight minutes deserves a pause. Conventional robotics means thousands of hours of physical data collection — expensive, slow, and sometimes dangerous, since you can’t let a robot trial-and-error its way through a hospital. Gameplay works as a pre-training environment, then a sliver of real data handles the fine-tuning.

The robot wasn’t graceful. It bumped into chairs like a toddler. But it was moving through a space it had never seen, using one camera. No LIDAR, no depth sensors, no pre-built map — only spatial understanding picked up from virtual worlds. This is what spatial AI agents look like in practice: models that understand space, time, and cause-and-effect well enough to control a physical body.

How does the Nerve platform fit in?

Nerve is a jobs marketplace built on Medal’s user base. Gamers earn money using hardware they already own, starting with data labeling — annotating gameplay, validating AI outputs — and graduating to robot teleoperation and higher-skilled work.

The logic is sharper than most coverage admits. Medal’s users belong to the generation most exposed to AI-driven job displacement, and Nerve gives them a stake instead of just extracting their clips. As models get more capable, they’ll need humans for edge cases, safety validation, and real-world data collection. It’s also a smart business move: a workforce with thousands of hours of spatial reasoning experience, even if it was earned in-game, is hard for competitors to hire into existence.

Who’s funding this, and what happens next?

The $320 million round at a $2.3 billion valuation came from Khosla Ventures, General Catalyst, Jeff Bezos, Eric Schmidt, Nico Rosberg, and researchers from Google DeepMind and MIT, bringing total disclosed funding to $454 million for a company less than a year old. The investor list matters more than the total — when the people who backed OpenAI and Anthropic put money into a company that learns from Fortnite, they’re betting the next breakthrough lives outside bigger language models.

Most of the new capital goes to compute through a CoreWeave deal, funding pre-training on the next model version. A slice funds an API launching by the end of summer, so developers outside the company can build on their spatial reasoning model.

What does this mean for solo builders?

Three things, and all of them are practical.

The API arrives this summer, which means you’ll be able to build spatial reasoning into your own applications — a robot vacuum that understands your furniture layout, or a warehouse drone that works without pre-mapped routes. Logistics, delivery, and automation are the obvious small-business use cases.

The method itself is replicable even without their dataset. The insight — action-labeled data beats raw video for spatial training — applies anywhere you can capture observations and actions together: game development, simulation platforms, even VR applications.

And the timeline compresses. If eight minutes of real-world data is enough to adapt a model to a new environment, you don’t need a data collection operation. You need good pre-training data and a small amount of domain-specific footage.

Who else is doing this?

General Intuition isn’t alone, but it’s ahead. Cambridge-based Worldmodeldata raised £7 million to turn licensed video-game data into training sets for AI world models. San Francisco’s Origin Lab raised $8 million in May for a similar approach, licensing gameplay from studios and selling it to AI labs.

The advantage General Intuition holds is Medal itself: 10 million monthly users uploading clips create a built-in data pipeline, the same kind of network effect that made platforms like YouTube so hard to displace. Competitors have to negotiate licenses. General Intuition just turns the faucet on.

FAQs

Why is video game data better than regular video for training AI? Gameplay recordings capture both the video and the inputs behind it: every button press, joystick movement, and click, timestamped alongside the footage. Regular video only shows outcomes, so models learn pattern recognition without causality. Action labels let an AI learn how decisions produce consequences, which transfers to controlling robots in physical space.

How much real-world data did General Intuition’s robot need? Eight minutes. After their model played Fortnite for 100 hours, the team loaded it onto a quadrupedal robot with a single camera, and the robot adapted to an unfamiliar office in eight minutes of real-world fine-tuning. Traditional robotics typically requires thousands of hours of physical data collection to reach the same point.

What is the Nerve platform? Nerve is General Intuition’s jobs marketplace for gamers. Users earn money with their existing hardware, starting with data labeling tasks like annotating gameplay and validating AI outputs, then progressing to robot teleoperation and higher-skilled work. It gives Medal’s 10 million monthly users a pathway into the AI economy rather than just supplying its raw material.

How much funding has General Intuition raised? Total disclosed funding is $454 million. That includes an initial $134 million raise and a later round of $320 million at a $2.3 billion valuation, led by Khosla Ventures with participation from General Catalyst, Jeff Bezos, Eric Schmidt, Nico Rosberg, and researchers from Google DeepMind and MIT.

When will General Intuition’s API be available? The company plans to make its API available by the end of summer, funded in part by the new round. Developers will be able to build applications on top of their spatial reasoning model, with most of the remaining capital going toward compute via a CoreWeave deal for pre-training the next model version.