Why a Gaming Startup’s $300M Raise Actually Makes Sense

I’ve been watching the embodied AI space for a while now, and most of the funding rounds feel like throwing money at a problem that’s still fundamentally unsolved. But this one from General Intuition caught my attention.

The startup is reportedly in talks to raise $300 million at a valuation around $2 billion. That’s a lot of zeros, but here’s the part that makes it interesting: they’re not building yet another robot that can open doors. They’re training world models using Medal’s dataset.

If you’re not familiar with Medal, it’s a gaming clip-capture platform with 10 million monthly active users. Those users generate about 2 billion videos per year. That’s not a typo. Two billion. Per year.

What General Intuition is doing with that data is actually clever. Instead of training AI on curated, sterile datasets — the kind you see in most robotics labs — they’re using real human gameplay. Millions of people making decisions, navigating 3D environments, reacting to unexpected events. That’s gold for training world models.

A world model, in case you’re not deep in the jargon, is basically an AI’s internal simulation of how the world works. If you can train one on billions of examples of cause and effect, you get something that can predict outcomes and plan actions. It’s the foundation for any embodied AI that needs to operate in the real world.

Now, I have some reservations. Gaming footage is not the same as real-world interaction. The physics are simplified, the consequences are fake, and the range of possible actions is constrained by game mechanics. But the scale here is hard to ignore. Most robotics datasets are tiny by comparison. A few thousand hours of teleoperated robot movements. Maybe a million simulated episodes. General Intuition has two billion videos per year.

That scale lets them do things other teams can’t. They can train models that generalize across different environments, different control schemes, different objectives. And because Medal’s users are generating this data naturally — they’re just playing games and clipping highlights — the cost of data acquisition is essentially zero.

I’ve seen this approach tried before in other domains. Tesla uses fleet data from millions of cars to train their self-driving models. Google uses search clicks to train ranking algorithms. The pattern is the same: find a massive, naturally occurring dataset and build your AI around it. General Intuition is just applying that playbook to embodied AI.

The $2 billion valuation is aggressive, no question. But compared to some of the other valuations I’ve seen in this space — companies with no revenue and a demo video — this one at least has a defensible thesis. They have exclusive access to a data pipeline that nobody else has. That’s a moat.

Whether they can actually bridge the gap from gaming to real-world robotics is the open question. I’m skeptical, but I’m also curious. If they pull it off, we’ll look back at this funding round as the moment embodied AI stopped being a lab experiment and started being a data problem.