General Intuition Raises $220 Million at $6.2 Billion, Nine Weeks After Its Series A
A world-model startup spun from a gaming platform raised $220 million just nine weeks after its Series A, nearly tripling its valuation.
General Intuition, a New York–based startup that builds world models for training artificial intelligence agents, raised $220 million in September at a $6.2 billion valuation. The round came nine weeks after the company completed its Series A at $2.3 billion, a nearly threefold jump that reflects less explosive growth in startup multiples and more a sharp pivot by major investors toward companies controlling high-quality training data.
The company spun out in 2025 from Medal, a platform where gamers record and share video clips, giving General Intuition access to roughly 10 million monthly active users and 2 billion recorded video clips annually. Those videos, showing first-person interactive gameplay, became the core asset that convinced a group of prominent investors—Valor Equity Partners, Atreides, Seven Six, Point72, Khosla Ventures and General Catalyst—to close a round so quickly after the previous funding announcement. The speed between rounds signals that investors saw early signs of customer interest in the technology.
Why world models matter in 2026
Large language models like GPT-4 process text tokens, but they cannot plan actions or navigate physical or simulated spaces. World models solve that gap: they predict what will happen next in video, allowing AI systems to anticipate consequences before acting. An AI agent trained on a world model understands timing, spatial reasoning and cause-and-effect. This capability is essential for robotics, autonomous vehicles, and game-playing agents that must make decisions in real-world or simulated environments.
Training world models requires video of agents interacting with environments—preferably lots of it, and ideally labeled with the actions that caused what happened on screen. Medal's library of gameplay footage is nearly ideal for that purpose. Unlike spectator video from YouTube or Twitch, Medal captures first-person, interactive video showing a player's decision-making in real time. The platform records roughly 2 billion clips annually from its user base.
That scale of interactive data has become the competitive focus for major technology companies. Every large AI laboratory is now hunting for the training data that will let its models simulate and predict environments. Companies without their own data source face a lengthy, expensive acquisition problem. General Intuition's possession of Medal solved that problem built-in, and the platform's continued growth creates an expanding moat that competitors cannot easily replicate.
The MIRA breakthrough and its limits
In June 2026, General Intuition released MIRA, a world model algorithm that ran at 20 frames per second in 720x576 resolution using only a single graphics card and 5.6 billion parameters. More importantly, the model could "run infinitely without diverging," meaning it could predict future frames far into the future without accumulating error. Previous approaches either compressed video into tokens before prediction or processed raw video frames directly, which required more memory. MIRA instead relies on latent diffusion, running calculations on a compressed latent-space representation of video rather than raw frames, which cuts memory use and speeds processing.
Infinite rollout capacity matters because it lets AI agents plan longer action sequences. Most previous world models degraded after a few seconds—their predictions became increasingly unreliable the further into the future they projected. An agent trained on a world model that becomes inaccurate after three seconds can only make immediate, short-term decisions. One that remains accurate for minutes or hours can plan complex multi-step actions. MIRA's ability to sustain prediction without collapse was a tangible step forward, not a marginal improvement in existing approaches.
MIRA remains a research demonstration: in its current iteration, the model can only generate synthetic footage of a single video game. General Intuition is testing a commercial version of the technology with a limited number of customers in robotics, simulation and entertainment, through a waitlist it opened alongside the funding announcement.
General Intuition had already begun engaging commercial customers before the Series B. The company was running a limited early-access program with customers in robotics, simulation and entertainment, and opened a public waitlist for the offering alongside the funding announcement. That early interest meant the Series B was not purely speculative: investors saw both working software and real customer engagement.
A different business model than competitors
Competitors like Decart and Google's Project Genie sell world models as products for other AI companies to license and use. General Intuition pursues a fundamentally different strategy: the company builds world models to train its own AI agents, then sells the agents rather than the models themselves. Revenue ties to agent capability, not simulation quality. The distinction is subtle but consequential for valuation and competitive position.
That choice has two immediate implications. First, General Intuition retains control of the world model technology, avoiding commoditization. If world models eventually become cheap infrastructure—like computing itself—selling them would become a low-margin business. Building products on top of them preserves margin and creates lock-in once customers build agents using the company's specific model architecture.
Second, the company can keep Medal's data proprietary. Selling world models would require licensing the underlying video, inviting competition and data-sharing agreements that weaken the source. By keeping the models internal and selling only agents trained on those models, General Intuition keeps the data advantage private and defensible. Competitors can build world models from public footage, but they cannot replicate Medal's scale and specificity without months or years of data collection.
This strategy also aligns General Intuition with how Nvidia and other infrastructure companies have historically captured value. Rather than selling raw capability as a commodity, the company layers proprietary products on top of the core technology. For robotics applications, that could mean selling specialized agents for specific tasks. For simulation, it could mean selling simulation engines that run on the company's world models. The business model remains in early stages, but the direction suggests a path to higher margins than selling world models as a service.
“Every large AI laboratory is now hunting for the training data that will let its models simulate and predict environments.”
Why the September round closed so quickly
The nine-week gap between Series A and Series B is unusual but not unprecedented when companies hit inflection points. General Intuition's situation met three conditions that justify accelerated follow-on investing: a defensible data asset that was still growing, technical proof of progress with a working algorithm, and demonstrated customer interest from early testing.
By September, the company had been testing MIRA with a limited number of customers in robotics, simulation and entertainment. Investors could see whether the algorithm's efficiency held up with real customers. They could also weigh Medal's continued growth—roughly 10 million monthly users generating about 2 billion clips a year—as evidence the data pipeline kept expanding. Growing data supply meant the next generation of models could be trained on an even larger dataset.
General Intuition said it will use the $220 million to hire more AI researchers.
The valuation in context
At $6.2 billion, General Intuition is now valued at roughly 28 times the $220 million Series B raise. That multiple is high but not exceptional for venture-stage artificial intelligence companies backed by marquee investors and demonstrating rapid progress.
The answer lies in the exclusive data asset and the pace of progress. Medal's roughly 2 billion video clips generated annually by more than 10 million monthly users represent a data pipeline that competitors cannot replicate quickly. The jump from Series A to Series B happened not because the company raised more money (Series B was smaller than Series A), but because the company had begun testing its technology with early customers, signaling that demand existed rather than merely being speculative.
What matters now is whether the company's technology can eventually justify the valuation. General Intuition has not disclosed specific customer contracts, customer counts, or revenue figures. The company must convert technical capability into products that AI labs, robotics companies, and simulation platforms will pay for at scale. Whether MIRA's performance translates to sustainable market demand beyond its current early-access testing remains to be seen. If the company can convert that testing into paying customers across new verticals, the valuation becomes reasonable. If early interest fails to convert, investors will reassess.
Related coverage: Clay's $7.1 billion round shows what New York investors will pay for AI sales tools; Profound's $180 Million Round Shows What Investors Pay for AI Search Visibility.
