General Intuition's $320M and why your data is the moat

General Intuition just raised $320 million at a $2.3 billion valuation to build AI agents that learn how to move through the physical world — by watching people play video games. Jeff Bezos and Eric Schmidt are in the round. The number is eye-catching, but the part worth stealing for your own business is quieter: the whole bet rests on a dataset almost nobody else can get.

What actually happened

On June 25, TechCrunch reported that General Intuition closed a $320M Series A led by Khosla Ventures, with General Catalyst, Bezos, Schmidt, and researchers from Google DeepMind and MIT participating. That brings total disclosed funding to $454M, on top of a $134M seed last October.

The company builds "world models" — systems that learn spatial-temporal reasoning, meaning how things move through space and time. Its edge is the training data: hundreds of millions of hours of gameplay from Medal, and critically, the action labels baked into those clips — records of exactly which buttons a player pressed and when. Founder Moritz De Witte argues rivals working from video alone have "insufficient" signal, because they can see what happened but not the input that caused it. The payoff they're chasing is transfer: the team says it took just eight minutes of real-world data to fine-tune a model to walk a quadruped robot. Most of the new money goes to compute.

Why it matters for your business

You are not going to buy a gameplay-trained world model. But the structure of this deal is a lesson you can use today. The models are commoditizing — there's a new frontier release every few weeks, and last quarter's leader is next quarter's discount tier. What isn't commoditizing is proprietary, well-labeled data. Investors just paid $2.3B for a company whose real asset is a dataset with the why attached, not the what.

Your business has the same kind of asset and probably throws it away. Every quote you win or lose, every support ticket and its resolution, every job you priced and how it actually went — that's your version of action-labeled data. Most operators keep it as unstructured text in an inbox, or not at all. Capture it in a structured form you own, and you can fine-tune, retrieve against it, or feed it to whatever model is cheapest next year. Skip that, and you're renting generic intelligence like everyone else.

We build the boring layer that makes this real: your operational data captured, labeled, and stored in systems you control — so the model stays swappable and the moat stays yours.

Sources: TechCrunch.