Startup Raises $320M to Train AI Agents on Millions of Hours of Gameplay

The bottleneck in building AI that can act in the world is not smarter algorithms; it is data showing what to do, not just what things look like. General Intuition just raised $320 million on the bet that the richest, cheapest source of that data is sitting in video game replays.

Key takeaways

• General Intuition raised $320 million at a $2.3 billion valuation, led by Khosla Ventures.
• The round brings total disclosed funding to about $454 million, after a $134 million launch round in October 2025.
• The company trains AI on hundreds of millions of hours of gameplay video and the player inputs that produced it.
• The goal is "spatial-temporal reasoning": understanding how to move through space and time.
• Investors include General Catalyst, Hedosophia, Bezos Expeditions, Innovation Endeavors, and former F1 champion Nico Rosberg.

What happened

General Intuition was spun out of Medal, a service that lets gamers upload and share video clips. Medal's library of gameplay footage, hundreds of millions of hours of it, gives the new company a vast, ready-made dataset. Crucially, that footage often comes paired with the player inputs, the actual button presses and movements, that generated each clip.

That pairing is the whole point. Most AI models learn from text or images. General Intuition wants to learn from action: what a player saw, what they did, and what happened next. The company calls the target capability spatial-temporal reasoning, the ability to understand how to navigate an environment over time.

Why gameplay data is different

To see why investors paid this much, compare the data sources available for training agents that act.

Data source	Cost to collect	Action labels	Risk	Variety
Gameplay video + inputs	Low (already exists)	Yes, paired button presses	None	Vast, many genres
Robot teleoperation	Very high	Yes	Physical, hardware wear	Limited
Self-driving fleets	Very high	Partial	Safety-critical	Narrow domain
Text and images	Low	No actions at all	None	Huge but action-free

The standout column is action labels paired with low cost and zero physical risk. Text and image data is plentiful but contains no record of decisions and outcomes. Real-world robot and vehicle data has those decisions but is expensive, slow, and sometimes dangerous to collect. Gameplay sits in the rare sweet spot: cheap, varied, and full of the perception-to-action loop that embodied AI needs.

Why it matters

Training agents to act in the world is a notorious bottleneck. The founder's framing, as reported by TechCrunch, is striking: "the same brain powering the agent playing the game is powering the robot." If a model can learn general spatial intuition from games, that capability could in principle transfer to robotics and other embodied tasks. The investor list, which includes Bezos Expeditions and a former Formula 1 world champion, signals serious belief in that thesis.

The approach connects to a broader shift in AI from passive text models to active agents. That theme runs through much of this year's research and investment, including the work we covered on building reliable AI agents and the rise of small models running on-device.

The hard part: transfer

The valuation implies enormous expectations, and the bet hinges on one unproven leap: that skills learned in games generalize to messier real environments. Games have clean rules, clear rewards, and forgiving physics; the real world has none of those guarantees.

1. Transfer to the real world. Whether game-trained spatial reasoning holds up in robotics or physical tasks, not just virtual ones, is the central open question.
2. Data advantages. How durable Medal's gameplay dataset is as a moat versus rivals collecting their own.
3. Concrete products. What General Intuition ships beyond research, and which industries adopt it first.
4. Safety and control. How action-trained agents are constrained as they grow more capable.

Who is backing it, and why that signals belief

Funding rounds are partly about money and partly about who shows up. This one drew an unusually pointed mix of investors, and the list reads like a bet on embodied AI specifically rather than AI in general.

Khosla Ventures led, the firm with a long record of backing contrarian deep-tech bets early. General Catalyst and Hedosophia bring later-stage growth conviction. Innovation Endeavors, founded by former Google chairman Eric Schmidt, has a thesis-driven interest in AI and the physical world. Then come the names that stand out: Bezos Expeditions, Jeff Bezos's personal investment vehicle, which has repeatedly backed robotics and space ventures, and Nico Rosberg, a former Formula 1 world champion now active as a sustainability and tech investor. The presence of a racing champion is not a gimmick; spatial-temporal reasoning, knowing how to move through space and time under pressure, is exactly the intuition a driver lives by.

What the mix signals is that serious capital believes the games-to-robots transfer is at least plausible enough to fund at a $2.3 billion valuation on a Series A. That is an early-stage round commanding a late-stage price, which only happens when investors think the data moat and the thesis are both rare.

How world models and action models fit together

The two technical terms General Intuition keeps using, world models and large action models, describe the two halves of an agent that acts. A world model is the AI's internal simulation of how an environment behaves: drop a ball, it falls; open a door, you can walk through. A large action model is the decision-maker that, given a goal and a situation, picks what to do next.

Gameplay data is valuable precisely because it contains both signals at once. Every clip shows an environment behaving (the world model side) and a human deciding how to respond to it (the action model side), with the button presses making those decisions explicit and machine-readable. Most datasets give you one or the other. Text describes the world but records no embodied decisions. Robot logs capture decisions but are expensive and narrow. Paired gameplay gives you the perception, the action, and the outcome in a single tightly coupled loop, which is the raw material agents need.

Frequently asked questions

What does General Intuition actually build?

It trains AI models on gameplay video and the player inputs behind it to develop spatial-temporal reasoning, the ability to move through and act in environments over time, with the aim of transferring that skill to agents and robots.

Where does the training data come from?

From Medal, a gameplay-clip sharing service the company was spun out of, which provides hundreds of millions of hours of footage often paired with the player inputs that produced it.

How much is the company worth?

The June 2026 Series A valued General Intuition at $2.3 billion, with total disclosed funding of about $454 million after a $134 million round in October 2025.

Why use games to train AI?

Games offer vast, varied, low-risk environments with clear actions and outcomes, making them far cheaper and faster than collecting real-world robot or vehicle data for training agents that act.

What is the biggest risk to the thesis?

Transfer. Game environments are clean and rule-bound, while the real world is noisy and unpredictable. Whether spatial skills learned in games actually carry over to robotics is unproven, and the $2.3 billion valuation assumes they will.

Whether the gameplay-to-robot leap pays off is unproven. But at $2.3 billion, a meaningful slice of the industry is now betting that the path to capable AI agents runs straight through the games people already play.