Bezos-Backed General Intuition Eyes $300M for Game-Trained Agents
A New York startup teaching machines to reason about space and time from billions of gameplay clips is in talks for $2B - and India's robotics builders should be watching closely.
The News
General Intuition, a New York startup spun out of the video-clip platform Medal, is in talks to raise around $300 million at a roughly $2 billion valuation, according to TechCrunch. The investor list reads like a who's-who of deep-tech capital: Jeff Bezos, former Google chief Eric Schmidt, Khosla Ventures and General Catalyst.
The round, still being negotiated, would land just eight months after the company closed a $134 million seed in October 2025. That cadence - a near-tenfold jump in headline valuation inside a single year - signals how aggressively investors are chasing the next frontier of machine reasoning.
Led by founder Pim de Witte alongside co-founders Eloi Alonso, Adam Jelley and Vincent Micheli, General Intuition is building foundation models that train AI agents in what it calls spatial-temporal reasoning. The fresh capital is earmarked to scale compute and ship a new product by the end of summer or early autumn.
Why It Matters
Most of the money pouring into AI over the past three years has chased language. General Intuition is betting on a different axis: teaching machines to perceive, anticipate and act inside a moving, three-dimensional world. Its raw material is unusual - Medal's library of roughly 2 billion gameplay videos uploaded annually by 10 million monthly active users. First-person game footage turns out to be a dense, cheap source of the embodied data that robots and autonomous systems desperately need.
The wager is that agents, not world models sold to third parties, are the product. That puts the firm in the same arena as Runway, Decart and Fei-Fei Li's World Labs, each racing to crack physical-world intelligence.
The last time a data-rich consumer platform pivoted into foundational AI on this scale, it was DeepMind training agents on Atari and StarCraft a decade ago - work that seeded much of today's reinforcement-learning playbook. The difference now is the capital intensity. A $2 billion valuation for a pre-product company underlines how investors are pricing embodied AI as the successor wave to the chatbot boom.
Indian Angle
For India, the interesting thread is not the dollar figure but the data thesis. India is one of the world's largest gaming markets by users, with hundreds of millions of mobile players generating exactly the kind of interaction footage General Intuition mines. Homegrown platforms such as Krafton-backed studios, Mobile Premier League and Loco sit on troves of gameplay data that, until now, have been monetised only through advertising and tournaments. The General Intuition model suggests a second, far larger use: training embodied AI.
It also sharpens the question for India's nascent robotics and world-model builders. Startups like Bengaluru-based humanoid venture and the robotics arms emerging from the IITs need vast quantities of spatial-temporal training data, and licensing it from abroad in dollars is expensive for rupee-funded teams. A domestic equivalent that converts Indian gaming exhaust into training corpora would be both a cost play and a data-sovereignty play - the sort of thing MeitY's IndiaAI mission has flagged as strategically important.
Finally, talent. A meaningful share of the reinforcement-learning and simulation engineers behind firms of this kind are of Indian origin or India-trained. As embodied AI heats up, expect the same competition for that talent pool that language-model labs triggered, with Indian compensation benchmarks rising in step.
FAQ
How much is General Intuition raising and at what valuation?
The company is in talks to raise around $300 million at a roughly $2 billion post-money valuation. The round is not yet closed, and terms could shift before it is finalised. It follows a $134 million seed round completed in October 2025.
What does General Intuition actually build?
It builds foundation models that train AI agents in spatial-temporal reasoning - teaching machines to perceive, anticipate and interact in real time within simulated environments. The agents themselves are the product, rather than world models licensed to outside buyers.
Where does its training data come from?
From Medal, the gaming-clip platform it spun out of. Medal collects roughly 2 billion videos a year from about 10 million monthly active users, supplying first-person gameplay footage as embodied training data.
Why should Indian readers care?
India's enormous base of mobile gamers generates similar interaction data, and India's robotics and world-model startups need exactly this kind of spatial-temporal corpus. The model points to a potential domestic opportunity in both data and embodied-AI tooling.
Where can I read the original report?
The story was first reported by TechCrunch; the link is in the attribution note below.
This story was reported by TechCrunch. Read the full original coverage at TechCrunch.