Two years ago, OpenAI offered $500 million to acquire the gameplay recording platform Medal, a deal rejected by founder Pim de Witte. In June 2026, his company, General Intuition, completed a $320 million Series A round, achieving a $2.3 billion valuation. Unlike traditional robot training that relies on simulators or real-world data collection, General Intuition leverages billions of hours of gameplay footage accumulated by Medal, each frame annotated with actual player inputs, enabling the training of spatial reasoning capabilities. The company was able to navigate unfamiliar indoor environments after just eight minutes of fine-tuning with real robot data. World models have become a new frontier in AI investment, with companies like World Labs and Decart also securing substantial funding; game data is emerging as a critical asset for training physical AI.

Author and source: GeekPark

Two years ago, OpenAI offered $500 million to acquire Medal, a gaming highlight sharing platform. Founder Pim de Witte declined the offer.

At the time, this decision may have seemed a bit impulsive.

General Intuition Founder Pim de Witte | Image source: TechCrunch

But today, in June 2026, de Witte’s AI company, General Intuition, completed a $320 million Series A round at a $2.3 billion valuation, led by Khosla Ventures, with participation from Jeff Bezos, Eric Schmidt, and researchers from Google DeepMind and MIT. Including the $134 million raised at its launch in October last year, their total funding has now exceeded $454 million.

At the time, OpenAI offered $500 million; today, General Intuition’s valuation is nearly five times that amount. The market is telling everyone with real money: this collection of game footage is worth far more than most people realize.

01 What's Hidden in the Game Replay

To understand what General Intuition is doing, you first need to clarify a prerequisite question: What training data are robots missing?

Traditional robot AI training relies on two approaches. One is collecting data in real-world environments, which is extremely costly and time-consuming; the other is training in simulators, which is faster but suffers from a persistent problem everyone struggles with: the "reality gap." Models trained in Unity or Unreal often behave like they’re lost when placed in front of real floors and walls.

The entry point for General Intuition is that game data is neither pure simulation nor real-world environment, but it may serve as a bridge connecting the two.

As a game replay platform, Medal has accumulated billions of hours of gameplay footage, but these are not ordinary videos. As Crypto Briefing’s analysis points out, General Intuition possesses data where each frame is paired with the player’s actual inputs—mouse movements, key presses, and strategic decisions.

Game footage used for training | Image source: General Intuition

This is something pure gameplay footage cannot achieve: you don’t just see the character moving—you also know what decisions the human player made at that moment.

From navigating buildings in first-person view in Fortnite to quickly identifying terrain and selecting landing spots in Apex Legends, these gaming scenarios contain vast amounts of spatial reasoning information. Human players daily train their brains to process three-dimensional space, predict object trajectories, and make real-time decisions in dynamic environments—skills that are precisely what robots need in the real world.

Witte said in an interview, highlighting a limitation of the LLM approach: “As humans, we create words to describe what happens in the world, but in doing so, we lose a great deal of information.” What they seek is more raw, perception-level data—not human-generated textual summaries created after the fact.

02 8-minute data

At the launch event in New York, General Intuition presented a demonstration worth examining in detail.

They pre-trained a spatial reasoning model using game data, then fine-tuned it with just eight minutes of real robot motion data. The robot, which collected this eight-minute data on the street, was then placed in an indoor office it had never visited before—and still navigated successfully.

8 minutes—this number is critical.

Game footage aggregates various virtual space data | Image source: General Intuition

Traditional robot training requires collecting vast amounts of data through repeated trials in the target environment or running countless simulations in highly accurate virtual environments. General Intuition’s approach, however, is based on the insight that game data has already provided the model with a strong foundation in spatial reasoning, significantly reducing the amount of real-world data needed for fine-tuning.

This logic sounds elegant enough. However, there's one aspect that's still unclear—the boundaries of generalization for this migration.

The journey from street to office is a demonstration, but whether complex real-world industrial scenarios, varying lighting conditions, and changes in ground materials can be bridged by just eight minutes of data remains unverified, as there are currently no publicly available benchmark tests to confirm this. AI CERTs News also highlighted this point: “The simulation-to-reality gap still exists and requires rigorous benchmarking.”

General Intuition promises to release the public evaluation results later this year.

03 Money Flow World Model

This round of funding by General Intuition is not an isolated event, but part of a wave of funding this year.

World Labs raised $1 billion in February; Decart secured $300 million in May; Odyssey closed a $310 million round in June, with Amazon and AMD participating. The common thread among these companies is "world models"—using vast amounts of real or simulated visual data to enable AI to understand the physical world.

Capital is uniformly betting on the assessment that the next phase of AI competition will not be about language, but about physics.

But in this competition, General Intuition’s moat is relatively concrete: the game data asset is extremely difficult for others to replicate. The billions of hours of footage accumulated by Medal are the result of years of buildup—not something that can be obtained simply by starting to collect today and having it ready tomorrow. This is precisely why OpenAI was willing to pay $500 million for it back then—data itself is a barrier.

The company revealed that most of the new funding will be allocated to expanding computing power, with plans to open its API to more developers by late summer this year. Transitioning from closed R&D to an open ecosystem marks a crucial step for General Intuition in evolving from a “fascinating research project” into a “platform-level company”—and it is the true test of whether its $2.3 billion valuation can be sustained.

Regarding another possible concern from game developers, de Witte also addressed this in a podcast interview: he believes fears that AI models will replace developers are "overblown," as training AI agents in spatial reasoning and generating game content are two entirely different paths.

The game screen is raw material, not a competitor.

How far is this road?

The idea of moving virtual physics from games into the real world is not something new.

Game engines have long been widely used in robotics simulation, with NVIDIA Isaac and Unity Robotics doing similar work. What sets General Intuition apart is that it doesn’t possess the engine itself, but rather the traces of human actions within games—intentional data, not just pure physics simulation outputs.

This difference can be understood through a simple analogy: teaching a student to drive, you could show them all the formulas about friction and turning radius from a physics textbook, or you could have them sit beside an experienced driver and watch a thousand hours of real driving footage, feeling the steering wheel’s resistance in their hands. The latter may not be more efficient, but it builds an ability closer to instinct. This is precisely where game footage holds value for robotic spatial reasoning.

Of course, there is a gap between driving games and real driving, just as there is a gap between the physics engine in Fortnite and the laws of physics in the real world. The size of this gap and how much data can bridge it are questions that General Intuition will need to answer through products and benchmarks, not through demonstrations or valuations.

In the wave of funding for world models, every company is telling a grand story about "physical AI." But the real turning point often comes after the API is opened—developers will quickly show what they can and cannot build with it. General Intuition’s late-summer API launch is likely a more critical milestone to watch than this funding round.

The gaming industry has spent decades training human brains to process three-dimensional space; now, that data is being used in reverse to train robots' brains—and this cycle alone is already fascinating.