The core training data behind a $2.3B embodied AI startup? Someone's Fortnite highlights.
Not a joke. The 2.3 billion game clips per year uploaded to Medal turned out to be the key ingredient that made a robot walk through an office on just 8 minutes of real-world data. No new data collection, no robotics lab. The answer was already in the clips.
It wasn't the video — it was the button logs
Medal lets gamers record and share gameplay clips. As of 2026, 10 million monthly active users upload 2.3 billion clips per year.
At first glance, it looks like a gaming version of YouTube Shorts. But Medal clips have something other video platforms don't: action labels — exact records of which buttons were pressed and when.
Medal operates at the game client level, capturing not just video but the controller input at every moment. To a viewer, it's "sick gameplay." To an AI, it's the answer to "what happened when I pressed left at this exact moment".
For an AI to act autonomously, it doesn't just need to see the world — it needs to know "what happens to the world when I take this action." Learning that requires paired video + action data, and collecting it at scale in the real world is expensive and slow.
| Regular game footage | Medal game clips | |
|---|---|---|
| Video data | ✓ Yes | ✓ Yes |
| Button action labels | ✗ No | ✓ Yes (input + timing) |
| AI training value | Visual pattern recognition | Action-consequence causal learning |
| Annual scale | Hundreds of millions | 2.3 billion clips |
How games teach a model to understand physics
General Intuition uses this data to build a "world model" — an AI that can predict what happens next given any action it takes.
After training on hundreds of millions of hours of gameplay, the model internalized the physics of in-game worlds:
- What a wall is — and that you can't pass through it
- How obstacles behave dynamically
- How shadows shift as the sun moves
- What happens when objects collide
And here's the thing — these aren't game-only rules. They reflect the same underlying physics that govern the real world. That's why the model transfers to robots and drones. As Silicon Canals noted, the thesis sounds absurd — that game data, not real robot telemetry, will produce the GPT of embodied AI. But the data says it's right.
"Gameplay is the largest untapped source of embodied-behavior training data. Agents trained on it develop spatial, temporal, and planning intuition that text-trained LLMs can never match."
— General Intuition core thesis
8 minutes to a walking robot
Too good to be true? It worked.
General Intuition's model can both play video games for hours and control a real robot. The real-world stat: 8 minutes of fine-tuning data → a quadruped robot walking autonomously through an office, with only a front-facing camera, no additional sensors.
Conditions were demanding: people walking through, objects added dynamically, zero extra sensing. Standard robotics training typically requires hundreds of hours of real-world trajectories.
These results brought in Jeff Bezos, Eric Schmidt, Khosla Ventures, and General Catalyst. $134M seed (October 2025), then $320M Series A at $2.3B — $454M in 8 months[[cite:4][cite:9]].
The ChatGPT moment for robotics?
In July 2026, TechCrunch published "This startup thinks robotics is about to have its ChatGPT moment." Just as text AI was trained on internet data, embodied AI may be trained on gameplay data — a paradigm now taking shape.
What this means for your data strategy
This isn't really an AI investment story. It's a data moat story.
What Pim de Witte had wasn't flashy AI research. It was Medal's action-labeled game data at a scale no competitor could replicate. And that became the basis for a $2.3B valuation.
What data does your company already have? Run through these 5 questions.
- Do you record user behavior in "pairs"?
Not just clicks or purchases — but "what situation triggered this action, and what was the outcome?" logged together. - Is your dataset irreplaceable at scale?
How many years would it take a competitor to match your volume? Two or more = real moat. - Are there hidden action labels in your data?
Like Medal's button logs, your data may contain metadata revealing user intent in ways you haven't leveraged yet. - Have you tried fine-tuning on it?
Fine-tune an existing foundation model on your proprietary data only. 8 minutes of data made a robot walk. - Are you building a data flywheel?
General Intuition's loop: deploy → collect real-world data → improve model → better deployment. It compounds.
