The Dutch AI company argues that keypress records embedded in hundreds of millions of hours of uploaded gameplay hold what each player was trying to do at every moment — the action signal raw video can't capture — and offers as proof one model that plays Fortnite for 100…
The video matters less than the keyboard underneath it. That, in one sentence, is the bet behind General Intuition's $2.3 billion valuation — on a $320 million raise, according to DutchNews — as reported on-site by TechCrunch's Brian Heater at the company's New York research floor.
The Dutch AI company, founded by Pim de Witte, has built a pretraining dataset it considers more valuable than the footage itself. The pipeline runs through Medal, the short-video platform for gamers that de Witte separately runs. Hundreds of millions of hours of uploaded gameplay come with something most archives throw away: a time-stamped record of every key the player pressed at each moment of the clip. General Intuition treats that stream not as metadata but as ground truth, the human motor intent that explains what the player was trying to do in each frame. In a field where most embodied-AI teams grind on expensive teleoperation and robot-collected trajectories, the pitch is that the humans already wrote the answer down, frame by frame, in their keypresses.
The proof, as Heater watched it on-site, is a single model. The same system that played Fortnite for 100 straight hours inside the company's demo loop was, minutes later, driving a four-legged robot through an unfamiliar room. The robot uses one camera, an exploration-default policy, and roughly eight minutes of real-world robotics data for fine-tuning. The behavior Heater described was not autonomy. The robot clipped chair legs and bumped trash bins. The interesting claim is not that the robot succeeded. The interesting claim is that the same model moved across two embodied domains with almost no new data, and that the cross-domain transfer survives changing the camera, the body, and the room.
That cross-domain transfer is what investors bought. General Intuition's funding arc runs from a $134 million seed in October 2025 to a $300 million round in talks at roughly a $2 billion valuation earlier this month, and now to the $2.3 billion figure at the top of Heater's reporting. A separate DutchNews piece pegs the round at $320 million, a figure that has not been reconciled with TechCrunch's $2.3 billion in public reporting; General Intuition did not respond to clarification requests before publication, and the gap is best read as a question of how the round was scoped (a single tranche, an extended vehicle, or cumulative capital raised) rather than a contradiction between the outlets. The capital is going into compute, into more clips, and into a small fleet of robots meant to test whether the cross-domain claim holds up outside a single demo floor.
The bet also has a market logic. Foundation models are commoditizing faster than the labs building them are comfortable with, and TechTimes reported this month that nearly $1.8 billion moved into inference- and world-model-focused startups in two days. General Intuition sits in that lane. Its pretraining work on 3D Gaussian-splat world models, the spatial-reasoning layer it first disclosed at the seed round, treats each gameplay clip as a way to teach a model how three-dimensional space behaves when a human is moving through it on purpose. A round like this is not just a bet on one startup. It is a vote that the next layer of value in AI sits on top of the foundation-model layer, in the data and the embodiment that turn a general-purpose system into one that can act.
The skepticism lives in the same demo. Eight minutes of fine-tuning is not a generalization claim in any peer-reviewed sense. The robot clipping chair legs is not a control result. And General Intuition's argument that competitors training purely on video are missing the action-label signal is, at this point, a competitive frame the company is advancing, not a finding the field has settled. The next honest question is whether the cross-domain transfer survives a benchmark instead of a demo room, and whether the keypress archive really does carry the motor intent its founder believes is encoded inside it.