The robotics and artificial intelligence worlds are converging at a pace that few predicted even a year ago, and the latest signal of that transformation is coming from a New York-based startup that believes the key to building truly generalized AI agents lies not in real-world data, but in hundreds of millions of hours of video game footage. General Intuition, the company training a foundation model meant to give AI agents the ability to move through space and time with human-like intuition, is in advanced talks to raise new funding at a $6 billion pre-money valuation, with participation from heavyweight investors including Valor Equity Partners, Point72 Ventures, and Seven Seven Six, according to sources familiar with the negotiations. The round, which is already oversubscribed, also includes follow-on investments from existing backers Khosla Ventures and General Catalyst. The new valuation represents a more than 2.5x increase from the company’s previous valuation of $2.3 billion, which was set just weeks ago when General Intuition raised $320 million. For a startup that only spun out of its parent company in October 2025, the trajectory is extraordinary — and it signals that institutional investors are making aggressive bets on which approach will win the race to build the operating system for physical AI.
A $6 Billion Bet on the Idea That Games Teach Machines Better Than the Real World
The core thesis behind General Intuition is both simple and deeply contrarian: instead of painstakingly labeling millions of real-world images and videos to train AI agents — the approach favored by companies like OpenAI, Google DeepMind, and Tesla — General Intuition believes that the richest, most scalable training data already exists in the form of video game footage. Specifically, the company is leveraging the hundreds of millions of hours of gameplay captured on Medal, the video game clip-sharing platform founded by General Intuition CEO Pim de Witte. What makes Medal’s data uniquely valuable, according to de Witte and his investors, is not just the video itself, but the accompanying “action labels” — precise records of which buttons a player pressed and exactly when they pressed them. These labels effectively serve as ground-truth annotations for intentional behavior in complex, dynamic environments. A player navigating a virtual city, dodging obstacles, aiming at targets, or solving puzzles is, from a machine learning perspective, generating a perfect training signal: here is an agent (the human player) acting intelligently in a complex world, and here is the exact sequence of inputs that produced that intelligent behavior.
The company’s name — General Intuition — is a direct reference to the capability its founders believe they can unlock. Vinod Khosla, whose firm Khosla Ventures is one of General Intuition’s earliest and most vocal backers, recently told TechCrunch that the action labels contained in Medal’s dataset will be a critical ingredient in what he calls the “emergence of intuition” — the moment when a model develops the ability to generalize across tasks it was never explicitly trained on, without requiring task-specific fine-tuning or massive amounts of new data. If that sounds like the holy grail of AI research, it is. The ability to build a single model that can navigate a warehouse, assemble a product, drive a vehicle, or perform delicate surgery — all without being retrained for each specific application — would fundamentally reshape the economics of robotics and automation.
From Medal to General Intuition: The Unlikely Origin Story
Pim de Witte is not a traditional AI researcher. He built Medal as a platform for gamers to clip and share short video highlights of their gameplay — essentially a Twitch-meets-YouTube for gaming moments. The platform grew rapidly, amassing a vast library of user-generated gameplay footage. What de Witte realized, likely earlier than most, was that Medal was sitting on a dataset that the world’s largest AI labs would have paid billions for. Every clip on Medal contains not just the visual stream of what the player saw, but the telemetry of what the player did. This combination — visual context plus behavioral ground truth — is extraordinarily rare at scale. Most video datasets (think YouTube or TikTok) have the visuals but lack structured action labels. Most robotics datasets (think Tesla’s Autopilot footage) have action labels but are limited in diversity and scale. Medal’s data, by contrast, is both enormous in scale and rich in behavioral annotation.
In October 2025, de Witte spun General Intuition out of Medal with a $134 million seed round, a sum that was itself remarkable for a company that had not yet shipped a product. The seed round valued the company at roughly $500 million, according to earlier reporting. By June 2026, the company had raised another $320 million at a $2.3 billion valuation, bringing total disclosed funding to over $450 million. Now, just weeks after that raise, General Intuition is already back in the market at a $6 billion pre-money valuation — a pace of dealmaking that reflects both the intensity of investor demand for physical AI exposure and the scarcity of credible approaches to building foundation models for embodied agents.
Who Is Backing General Intuition — and Why It Matters
The list of new investors in this round reads like a who’s-who of institutional capital with a focus on deep tech and long-duration bets. Valor Equity Partners is perhaps the most intriguing name on the cap table. Valor is best known for being one of the earliest and most consistent institutional backers of SpaceX, and the firm has historically been highly selective about its investments in AI. In fact, General Intuition would be the first AI lab that Valor has invested in since SpaceX, according to sources familiar with the firm’s portfolio. The parallel is instructive: just as Valor bet early on the idea that reusable rockets could fundamentally change the economics of space access, the firm is now betting that a foundation model trained on video game data could fundamentally change the economics of robotics. Both bets rest on the conviction that a step-change improvement in a core technology can unlock entirely new markets.
Point72 Ventures, the venture arm of Steve Cohen’s Point72 Asset Management, brings a different kind of credibility. Point72 has been building a substantial AI and deep tech practice, and its participation signals that the crossover between quantitative finance and physical AI is becoming more recognized. Seven Seven Six, the venture firm founded by Alexis Ohanian, adds a consumer-internet and platform-economics perspective to the investor group. Existing backers Khosla Ventures and General Catalyst are both doubling down, which in venture capital is one of the strongest signals of conviction. When existing investors are willing to increase their pro-rata at a valuation that has more than doubled in a matter of weeks, it indicates that the company’s internal milestones are being met or exceeded.
What General Intuition Is Actually Building: A Foundation Model for Movement
To understand why investors are placing such large bets on General Intuition, it helps to understand the technical problem the company is trying to solve. Today’s AI models are remarkably capable at certain kinds of tasks — language generation, image recognition, code writing — but they are fundamentally disembodied. A large language model like GPT-4 or Claude can write a poem about a robot, but it cannot navigate a room, open a door, or pick up an object. The gap between intelligence that lives in text and intelligence that can act in the physical world is one of the most significant unsolved problems in AI.
General Intuition’s approach to bridging that gap is to train a foundation model for spatial-temporal reasoning — a model that understands not just what objects are, but where they are, how they move, how they relate to each other, and what sequences of actions lead to desired outcomes. The company’s initial training data comes from Medal’s gameplay library, but the goal is not to build a model that plays video games. The goal is to build a model that learns from video games the fundamental principles of physics, navigation, object manipulation, and goal-directed behavior — and then transfers that knowledge to the real world. This is the same basic strategy that DeepMind used with AlphaGo and AlphaZero (learning from game simulations and then transferring to real-world problems), but General Intuition is applying it at a much larger scale and with a much broader scope.
Video Games as a Robotics Training Ground
The idea that video games can serve as a training environment for AI agents is not new. OpenAI’s Dota 2 bots, DeepMind’s StarCraft agents, and Waymo’s simulated driving environments all used game-like or simulation environments to train models. What sets General Intuition apart is its reliance on human gameplay data rather than synthetic or self-play data. This distinction is important for several reasons. First, human gameplay data contains a diversity of strategies, behaviors, and failure modes that are difficult to replicate in self-play. Second, the action labels in Medal’s dataset are genuinely human — they reflect the messy, intuitive, imperfect decisions that real people make under time pressure, which may be more transferable to real-world robotics than the perfect-but-brittle behaviors learned through reinforcement learning from scratch. Third, Medal’s dataset is uniquely large: hundreds of millions of hours of footage covering thousands of different games, each with its own physics, objectives, and control schemes. The scale and diversity of this dataset is something that no other AI lab can easily replicate.
How General Intuition Plans to Spend $6 Billion
According to sources close to the company, General Intuition intends to use the proceeds from this round to double down on two areas that are essential for its long-term strategy: compute infrastructure and talent acquisition. Both are massive cost centers for any company attempting to train frontier AI models, but they are especially acute for a company working on physical AI, which requires not just language-scale compute but also simulation, rendering, and real-time interaction capabilities. General Intuition has an existing partnership with CoreWeave, the GPU infrastructure company that has become one of the most important suppliers of compute for AI startups. The new funds will likely allow General Intuition to significantly expand its cluster, potentially putting it in competition for scarce H100 and B200 GPU capacity with companies many times its size.
The second major use of funds is hiring. Building a foundation model for physical AI requires talent that is simultaneously scarce and expensive: researchers who understand both modern deep learning (transformers, diffusion models, reinforcement learning) and classical robotics (control theory, kinematics, sensor fusion). The number of people in the world who genuinely excel at both can be counted in the hundreds, and most of them are already employed at DeepMind, OpenAI, Tesla, or Boston Dynamics. General Intuition will need to offer compelling compensation, research freedom, and the promise of working on one of the most ambitious problems in AI to attract that talent. The $6 billion valuation gives the company significant equity currency to make those offers.
The Robotics Embodiment Question: When Does the Model Meet the Machine?
One of the most closely watched aspects of General Intuition’s strategy is its approach to robotic embodiment. The company has stated that it intends to focus on improving its general model specifically for robotic embodiments, but it has not yet revealed whether it plans to build its own hardware, partner with existing robotics manufacturers, or license its model to third parties. Each path comes with trade-offs. Building hardware is capital-intensive and slow, but it allows for tight integration between the model and the physical platform. Partnering is faster and cheaper, but it risks ceding control over the user experience and the feedback loop. Licensing is the most capital-efficient approach, but it depends on there being a robust market of robotics companies that are ready to integrate third-party AI models — a market that is still in its infancy.
The fact that General Intuition is raising at a $6 billion valuation before answering these embodiment questions suggests that investors are betting primarily on the model rather than on any specific go-to-market strategy. The assumption is that if General Intuition can build a truly general foundation model for spatial-temporal reasoning and physical action, the embodiment question will solve itself — because every robotics company in the world will want to use that model. This is the platform play, and it is the same logic that drove early investments in companies like OpenAI (the model would be the platform) and Android (the operating system would be the platform). Whether that logic holds for physical AI is one of the most important questions in technology investing today.
What Is the “Emergence of Intuition” and Why Does It Matter?
Vinod Khosla’s phrase — the “emergence of intuition” — deserves closer examination, because it captures both the promise and the difficulty of General Intuition’s approach. In AI research, emergence refers to the phenomenon where a model develops capabilities that were not explicitly programmed or trained for, simply as a byproduct of scale and training on diverse data. For example, large language models were trained to predict the next word in a sentence, but they eventually developed the ability to translate languages, write code, and solve logic problems — tasks that were not part of their original training objective. Khosla and General Intuition are betting that a similar emergence will happen with spatial-temporal reasoning: a model trained on video game action labels will, at sufficient scale, develop the ability to generalize to real-world physical tasks that it has never seen before.
If this bet pays off, the implications are profound. It would mean that a single model could control a robot arm in a factory, a drone in a warehouse, a delivery vehicle on city streets, and a humanoid robot in a home — all without being retrained or fine-tuned for each specific environment and task. It would mean that the cost of deploying physical AI could drop by orders of magnitude, because the expensive part (training the foundation model) would be amortized across millions of applications. It would also mean that the company that owns that foundation model would be in a position of extraordinary leverage, much as OpenAI and Anthropic are today in the language model space.
The Competitive Landscape: Who Else Is Racing to Build Physical AI?
General Intuition is far from the only company pursuing physical AI, but its approach is distinct enough that it occupies a unique position in the competitive landscape. OpenAI has shown interest in robotics but has not made it a primary focus since shutting down its robotics division in 2021; the company’s recent investment in humanoid robotics startups like Figure suggests it is re-engaging, but its approach is likely to be model-centric rather than data-centric. Google DeepMind has been working on robotics foundation models for years — RT-1, RT-2, and the PaLM-E family of models — and has the advantage of owning both cutting-edge AI research and a robotics team with deep expertise. However, DeepMind’s approach has been more focused on real-world data and simulation, and it has not publicly embraced the video game data thesis that is central to General Intuition.
Tesla is arguably the most vertically integrated player in physical AI, with its own AI hardware (Dojo), its own foundation model work (the same system that powers Full Self-Driving), and its own robotics platform (Optimus). But Tesla’s data comes primarily from its fleet of vehicles, which is powerful for driving but may not generalize well to the broader range of physical tasks that General Intuition is targeting. Boston Dynamics has best-in-class hardware but has historically been less focused on foundation models and more on control systems and classical robotics. Covariant and Physical Intelligence are two startups that, like General Intuition, are building foundation models for robotics, but they have taken different data and architecture approaches. The field is crowded, but it is still early, and no single approach has yet emerged as dominant.
Why Action Labels May Be General Intuition’s Moat
If there is one factor that could give General Intuition a durable competitive advantage, it is the action label dataset from Medal. Unlike synthetic data (which can be generated by anyone with compute), and unlike real-world data (which can be collected by anyone with robots), Medal’s human gameplay data with action labels is a proprietary resource that cannot be easily replicated. A competitor would need to acquire or build a platform with similar scale and similar behavioral annotation — a process that would take years and cost billions. This data moat is one of the reasons investors are willing to assign a $6 billion valuation to a company that is still in its early stages of development. The data is the asset, and the model is the derivative.
Of course, data moats have a mixed track record in AI. Some datasets have proven to be enduring advantages (Google’s search index, Meta’s social graph), while others have been rendered obsolete by new architectures, new data sources, or new training techniques. It is possible that advances in synthetic data generation, simulation, or self-supervised learning will reduce the importance of Medal’s human gameplay data over time. But for now, General Intuition has something that no other AI lab has: a dataset of human intentional behavior in complex environments at a scale that is unmatched in the industry. That alone justifies a significant premium.
The Oversubscribed Round: What It Tells Us About the Market for Physical AI
The fact that General Intuition’s latest round is oversubscribed — with more demand from investors than the company is willing to accept — is a signal not just about this specific startup, but about the broader market for physical AI investing. Institutional investors are increasingly viewing physical AI as the next major wave of technological disruption, following the language model wave that crested in 2023 and 2024. The logic is straightforward: if AI can think (language models) and see (vision models), the next frontier is for AI to act (physical models). The market for physical AI — in manufacturing, logistics, construction, healthcare, agriculture, and the home — is potentially larger than the market for language AI, because physical work represents a much larger share of global economic output than knowledge work.
But investing in physical AI is also riskier and more capital-intensive than investing in language AI. The path to a return is longer, the technical challenges are harder, and the risk of hardware or safety failures is higher. The oversubscription for General Intuition suggests that a significant cohort of institutional investors has decided that the risk-reward calculus is favorable, and that they are willing to pay up for access to the most promising companies in the space. It also suggests that the window for investing in physical AI at reasonable valuations may be closing, as more capital chases a limited number of credible teams and approaches.
A Story Still in Development: What to Watch Next
As the company notes in its communications, this story is still developing. The round has not yet closed, and terms could change. The participation of Valor Equity Partners — pending official confirmation — would be a significant milestone, signaling that investors with deep experience in hard tech and long-duration bets see General Intuition as a generational opportunity. The company’s ability to raise at a $6 billion valuation just weeks after a $2.3 billion round is a testament to the intensity of investor demand, but it also raises the stakes: General Intuition will now be expected to deliver on its technical promises at a pace that matches its valuation growth. The company will need to show meaningful progress on model performance, robotic embodiment integration, and real-world deployment within the next 12 to 18 months to justify the current valuation — let alone the next one.
For the broader AI and robotics ecosystem, General Intuition’s rapid ascent is both validating and destabilizing. It validates the thesis that video game data is a powerful resource for training physical AI, and that investors are willing to make large bets on that thesis. But it also destabilizes the existing competitive order, because a startup that did not exist a year ago is now sitting on a valuation that rivals or exceeds that of established robotics companies with decades of history and hundreds of deployed systems. Whether General Intuition can translate its valuation advantage into a real technical advantage will be one of the most closely watched stories in AI over the next several quarters. The company has the data, the investors, and the ambition. Now it needs to build the model that justifies the belief.
