General Intuition Raises $220M at $6.2B Valuation

The Startup Training Robots on Your Game Clips Just Nearly Tripled in Value — in Three Months

Update, Sept. 30: the $40M revenue line and the OpenAI offer de Witte refused

Two details in DutchStartup.ai's reporting reframe the raise. General Intuition generated roughly $40 million in 2025 revenue from early-access partner agreements — unusually large for early deeptech — and de Witte turned down a reported $500 million acquisition offer from OpenAI for Medal's data in 2024, before spinning the lab out. Medal crossed two billion cumulative uploaded clips in August 2026, with about 195 million new clips arriving each month. On the technical side, the official MIRA repository clarifies the demo was trained on roughly 10,000 match-hours of bot self-play 2v2 Rocket League matches — no human gameplay — with physics state logged via BakkesMod for evaluation only, and released under Apache 2.0. The company is legally registered in the Netherlands and operates as a public-benefit corporation.

General Intuition has raised another $220 million at a $6.2 billion valuation, the company announced on September 29 through a post on X by co-founder and CEO Pim de Witte. The round lands barely three months after the company's $320 million Series A priced it at $2.3 billion — which means one of the hottest startups in "physical AI" has nearly tripled its paper value in less than a fiscal quarter, on a thesis that may be the most audacious in robotics: the missing training data for robots is not more text, and not more lab footage. It is gameplay.

File photo: esports competitors at the 2015 League of Legends World Championship in Paris. General Intuition trains its models on billions of action-labelled gameplay videos sourced from Medal’s upload corpus — not on esports broadcasts. Photo: Bruce Liu, CC BY-SA 3.0, via Wikimedia Commons.

The round was led by Valor Equity Partners and Atreides, with participation from 776, Point72, Khosla Ventures, and General Catalyst, according to coverage from GamesBeat, RuntimeWire, PocketGamer.biz, FinSMEs, and DutchStartup.ai. General Intuition says it has now raised more than $650 million since last October. In an AI market where late-stage investors are bidding up anything with a defensible data moat, this is the clearest signal yet that gameplay footage with recorded player inputs — what researchers call "action labels" — has become a first-class training asset.

The $220M round: the numbers behind the $6.2B valuation

The fundraising velocity is genuinely unusual. General Intuition's first disclosed round was a $133.7 million seed in October 2025, led by Khosla Ventures and General Catalyst. Then came the $320 million Series A on June 25, 2026 at a $2.3 billion valuation. RuntimeWire calculates the disclosed total at roughly $673.7 million including the new $220 million, and notes the announcement does not specify whether the $6.2 billion valuation is pre- or post-money.

There was also a public trail before the announcement. In late August, TechCrunch reported that General Intuition was finalizing a round at a $6 billion pre-money valuation led by new investors Valor Equity Partners and Point72 Ventures — a round one source close to the deal described as oversubscribed. That report, which also named Seven Seven Six as joining, reads in hindsight as this exact deal closing. The final $6.2 billion figure landed slightly above the rumored number.

Valor's participation is notable on its own terms: TechCrunch noted this would be the fund's first AI lab investment since SpaceX. For a firm known for backing SpaceX, writing the lead check here is a statement that "world models" have graduated from research curiosity to venture-scale infrastructure.

Why games: the action-label thesis

The company's entire bet rests on a simple observation that de Witte has repeated across announcements. "Humans play video games and (tele)-operate robots using the same inputs and outputs, namely vision & controller actions," he said in the funding post, per PocketGamer.biz. "We train directly on the largest dataset of these, similarly to how Tesla trained FSD, but across more actions and environments."

The mechanism is Medal, the gameplay-clips platform de Witte co-founded before spinning General Intuition out in 2025. Medal is where players upload clips of their gameplay — and it is on track for three billion uploaded videos per year, according to GamesBeat's reporting on the announcement. Each clip is not just video; it carries the player's exact inputs, button by button, millisecond by millisecond. That pairing of vision with recorded action is what the field calls action labels, and General Intuition's claim — per de Witte's LinkedIn post cited by GamesBeat — is that leading robotics and world models are trained on less than one percent of the action data it can access.

De Witte framed the statistical argument in unusually vivid terms: "On any given day, we see more accidents in sim than across the entire USA," he said. "And at any moment, more people play with steering wheels than Waymo has cars on the road. This allows us to create General Agents and World Models, capable of handling hard to predict events."

The point is edge cases. Robots fail not on routine motion but on rare, dangerous, hard-to-predict events — and gathering millions of examples of such events in the physical world is slow, expensive, and sometimes unsafe. Simulated worlds where humans constantly crash cars, fumble physics puzzles, and improvise under pressure offer an effectively infinite supply of exactly those situations.

MIRA: the 5B-parameter world model trained on self-play

The concrete demonstration of the thesis is MIRA, a multiplayer interactive world model that a full 2v2 Rocket League match can be played inside — in real time, at 20 frames per second, with no explicit physics or rendering engine, just a neural model generating each frame from the players' actions. The model was built by General Intuition together with the French AI lab Kyutai and in collaboration with Epic Games, and its code and technical report have been published openly on GitHub under the mira-wm repository.

The technical figures are striking. The official repository describes MIRA as a 5-billion-parameter latent diffusion model; press coverage of the announcement, including Tech Startups' September 30 writeup, cites 5.6 billion parameters and notes it renders interactive video at 20 frames per second on a single Nvidia B200 GPU — a fraction of the trillion-parameter scale that dominates headlines. The demo site at mira-wm.com lets anyone play the model live. Coverage also reports it can run indefinitely without diverging, generating long, seamless simulations of fast-moving objects where most video-generation models degrade within seconds.

To be clear about scope: the public demo centers on a single game. But the company positions MIRA as a step toward "physical AI" — world models that learn how environments respond to actions and can later be pointed at robots, simulators, and autonomous systems. AIWeekly's earlier reporting described the pitch more concretely: an agent trained on gameplay that then controlled a quadruped robot after roughly eight minutes of real-world fine-tuning data. That transfer — game world to physical world — is the result the $6.2 billion valuation is really pricing.

Who is betting, and why

The investor roster tells its own story about where the round sits in the market. Valor and Atreides led; 776 (Alexis Ohanian's firm), Point72 Ventures, Khosla Ventures, and General Catalyst joined. Khosla and General Catalyst were already in from the seed and Series A — insiders doubling down as the price nearly tripled.

The conviction language from investors is unusually philosophical. Vinod Khosla, quoted in earlier coverage, said: "In world models, I think the quantum leap is the emergence of intuition in the AI," driven by human action data from games. The company's name is the thesis: that intuition — the ability to act sensibly in situations never explicitly trained for — emerges from watching humans act, at scale.

Timing matters too. The round lands in a quarter where physical-AI startups have been vacuuming up capital; one industry tracker cited in recent coverage put H1 2026 physical-AI funding at $47.4 billion. General Intuition is positioning itself as the data-and-models layer for that wave, with a compute partnership with CoreWeave (the company described as "neolab CoreWeave" in earlier coverage) to scale training.

The skeptic's case

The valuation now demands scrutiny, and the company's own disclosures invite it. The models are being made available to external partners through a waitlist, per PocketGamer.biz — until now, access was a closed API with a small group of companies in gaming, simulation, and robotics, per DutchStartup.ai. So the product is real, but broadly unproven outside early partners. Tech Startups' analysis put it plainly: MIRA remains a research demonstration centered on a single game, and its broader commercial capabilities have not yet been established.

There is also a substantive research question the market is glossing over: whether action labels learned in game worlds actually transfer to physical robots at the scale and reliability industry needs. The quadruped demo is suggestive, not conclusive. If the transfer works, Medal's user-generated corpus is a compounding moat no lab can replicate without an equivalent content business. If it does not — if billions of gameplay videos turn out to be less useful outside simulated environments — the valuation is pricing a dataset whose primary asset is entertainment.

Note also the careful wording in the announcement: neither pre- nor post-money basis was specified for the $6.2 billion figure, per RuntimeWire. Comparison against the June $2.3 billion figure is directionally clear — roughly a 2.7x increase in under three months — but the exact basis is not.

General Intuition FAQ

What is General Intuition? A 2025 spinout of Medal, the gameplay clip platform, founded by Pim de Witte. Its thesis — the "action-label" thesis — is that watching humans act at scale teaches AI systems the intuition to act in situations they were never explicitly trained for.

What is MIRA? A playable 2v2 Rocket League world model: a 5B-parameter diffusion model plus a 600M-parameter codec, trained on roughly 10,000 match-hours of bot self-play — no human gameplay — and open-sourced under Apache 2.0.

Is the $6.2 billion valuation pre- or post-money? The announcement does not specify whether the $6.2 billion valuation is pre- or post-money.

How do games train robots? Human gameplay carries action labels — what the player did in each frame — which transfer to physical tasks: the company demonstrated a quadruped robot fine-tuned on the approach in roughly eight minutes of training, per AIWeekly.

What's next

General Intuition says it will scale the team in New York and in Europe — "If you're a top researcher or engineer, reach out," de Witte said, per GamesBeat — and broaden partner access to its models beyond the closed-API phase. The company's founders alongside de Witte are Eloi Alonso, Adam Jelley, and Vincent Micheli, and DutchStartup.ai notes offices in New York, Geneva, Paris, and London, with data and IP managed from the Netherlands.

The deeper significance for the industry is the emerging two-track race in physical AI. On one track, companies like AMD — which just placed an $8.2 billion bet on World Labs — are buying their way into world models through acquisition. On the other, General Intuition is building the training substrate from an asset only it owns: three billion gameplay clips a year, each one a record of a human acting in a complex world. Which approach generalizes to real robots first will decide whether this $6.2 billion is the opening bid or the peak.

Sources: GamesBeat; RuntimeWire; PocketGamer.biz; FinSMEs; DutchStartup.ai; Tech Startups; TechCrunch (Aug. 24 reporting); AIWeekly; the official mira-wm GitHub repository and technical report; General Intuition's September 29 X announcement via de Witte. De Witte quotes via GamesBeat and PocketGamer.biz reporting on the announcement post.