General Intuition raised a $320M Series A in June 2026 at a $2.3 billion post-money valuation, led by Khosla Ventures with participation from General Catalyst, The Raine Group, Hedosophia, Bezos Expeditions, Innovation Endeavors, Eric Schmidt, and Nico Rosberg.

Before the Series A, the company raised a $134M seed round in October 2025 at launch. Total disclosed primary funding across the two rounds stands at approximately $454M.

General Intuition builds large action models and world models that perceive, predict, and act across virtual and physical environments. The product has two parts: a world model that learns how an environment evolves over time as an agent takes actions, effectively an internal simulator trained from video and control data rather than a hand-coded physics engine, and an action model that decides what to do next in that simulated or real environment to move toward a goal.

The clearest public example is MIRA, a playable multiplayer world model built with Kyutai and Epic Games and trained on 10,000 hours of Rocket League gameplay. MIRA runs in real time at 20 frames per second, responds to the keypresses of multiple simultaneous players, and maintains coherent physics-like behavior, including boosting, collisions, and scoring, with no underlying physics engine or 3D renderer. The system learned those behaviors from action-conditioned video.

MIRA is less important as a Rocket League application than as evidence for the training approach. The models learn from action-labeled gameplay clips sourced from Medal, the game-clipping platform from which the company spun out. Unlike plain video scraped from the internet, these clips include ground-truth records of which controls produced which outcomes, so the model can learn causal relationships between inputs and consequences rather than visual patterns alone.

The company uses the world model as a training gym rather than a standalone product. In practice, the learned simulation environment lets agents be trained, evaluated, and iterated on before touching real hardware, while the agents themselves, the action models produced through that process, are the commercial output.

For robotics developers, the workflow is to provide environment data or a robot task specification, use General Intuition's API to pretrain or fine-tune an action model inside the learned simulator, and then deploy that model to a physical robot with less real-world data collection than would otherwise be required. The company demonstrated this with a quadruped robot that needed only eight minutes of real-world data to fine-tune after pretraining in its broader system. For game studios, the workflow is to integrate the API to power adaptive NPCs, QA bots, or interactive simulation environments that respond to player behavior in ways a scripted system cannot.

General Intuition is a B2B foundation-model platform that sells access to its world and action models through a commercial API rather than building or selling robots directly. It monetizes through usage-based API access, with enterprise pilot pricing for early partners and a billing-grade infrastructure layer built to scale from the first hundred users to tens of thousands of concurrent ones.

The company is compute-intensive by nature, and most of its Series A capital is going toward GPU capacity, including a production relationship with CoreWeave, but it is asset-light relative to vertically integrated robot manufacturers because it does not own hardware fleets or field service operations. Gross margins are likely lower than pure software because of the compute intensity of training and serving large models, but the model scales better than hardware-centric robotics companies because inference economics improve as the model matures and distillation reduces serving costs.

A key structural feature of the business is the Medal flywheel. Medal generates billions of action-labeled gameplay clips per year from over 17 million monthly active users, and that data flows into General Intuition's training pipeline, improving the coverage and causal richness of its world and action models. Better models can improve Medal's own platform features, while commercial API revenue funds more compute and research. The loop is difficult for competitors to replicate from a standing start.

The strategic bet is that owning the pretraining layer, where proprietary action data creates a durable prior over intent and consequence, lets General Intuition capture recurring software-like economics without owning the physical deployment stack.

General Intuition competes across three overlapping fronts: horizontal foundation-model labs building similar world and action model stacks, vertically integrated robot companies that own the full loop from model to hardware to deployment data, and infrastructure incumbents that could commoditize the underlying platform. The competitive question is whether value accrues to a standalone model layer or shifts to companies that control deployment data, hardware integration, or the broader tooling stack.

Physical Intelligence and Skild AI are the closest independent competitors. Physical Intelligence is building generalist control models aimed at any robot doing any task, with increasing emphasis on real-robot deployment and partner integrations that give it a tighter feedback loop than a pure simulation-first approach. Skild AI targets the same data bottleneck General Intuition does, arguing that large-scale pretraining reduces dependence on scarce robot demonstrations, but frames its pitch around cross-embodiment breadth across quadrupeds, humanoids, and tabletop arms, which maps more directly to enterprise robotics buyers evaluating specific hardware.

Google DeepMind competes specifically on the games-as-training-ground axis, with Genie 3 for interactive world generation and SIMA 2 for agents that reason and act in virtual 3D environments. DeepMind's advantages are frontier-model scale, multimodal integration across Gemini, and existing OEM partnerships. General Intuition's counter is more explicit access to action-labeled interaction data through Medal, where the causal signal between control inputs and outcomes is richer than generic video.

Figure, with its Helix onboard control stack, is the clearest structural threat to General Intuition's model-layer position. Vertically integrated robot companies collect proprietary closed-loop deployment data at scale, improve policies in the field, and capture the full economics of the system, leaving less budget and strategic need for an upstream model vendor. The same dynamic applies to the Google DeepMind and Apptronik ecosystem, where Gemini Robotics is positioned as a VLA capable of controlling any robot type and is already in private preview with hardware partners.

General Intuition is strongest if the market stays modular and buyers want a standalone world and action model supplier. Its position weakens if the winning loop becomes deployment data feeding model improvement feeding better hardware behavior, a cycle that vertically integrated players are better positioned to run.

NVIDIA Cosmos is the most serious platform-level threat because it is not just a competing model but an ecosystem: open-weight world foundation models, tokenizers, post-training recipes, NIM containers, and deep integration with Omniverse and Blackwell infrastructure. NVIDIA's partner list already includes Figure, Skild AI, and Toyota Research Institute, which means it can become the default substrate partners build around rather than General Intuition's API.

World Labs, after acquiring SceniX in July 2026, is expanding from spatial world generation toward robotics enablement, attacking the same adjacency from a stronger 3D scene-construction position. Meta's V-JEPA 2 advances the thesis that video pretraining plus modest robot data can unlock strong physical planning, and its open research posture reduces the novelty premium around proprietary world-model claims for any well-resourced team that wants to build internally.

General Intuition's expansion logic rests on one thesis: if game-derived action intelligence transfers to physical environments, the same stack can serve games, simulation infrastructure, and robotics, with one pretraining investment spread across multiple markets.

The nearest expansion is from selective partner access to a broadly available developer API, which the company targeted for the end of summer 2026. That transition would expand the customer base from a handful of design partners to a larger pool of game studios, simulation companies, and robotics developers that need controllable agent behavior in dynamic environments.

Beyond API access, the world model creates a second product surface: learned simulation environments that can serve as sandboxes for policy testing, synthetic trajectory generation, and offline agent evaluation. That puts General Intuition into the simulation infrastructure layer that NVIDIA, Siemens, and others identify as a bottleneck for physical AI adoption, not by replacing physics-based simulators, but by augmenting them with learned dynamics in partially observable or multi-agent settings where hand-coded simulators break down.

Games are the initial beachhead because the Medal data asset and the MIRA collaboration with Kyutai and Epic Games give General Intuition credibility with game-native buyers. Within games, the expansion path runs from research partnerships toward production use cases such as adaptive NPCs, QA automation, replay analysis, and interactive simulation tools that game studios would pay for at scale.

The larger TAM expansion is robotics. The company's eight-minute fine-tuning demonstration with a quadruped robot is the clearest evidence that gameplay pretraining can reduce the cost of real-world data collection, a bottleneck in physical AI economics. If that transfer holds across more embodiments and tasks, General Intuition can sell into robot foundation-model startups, industrial automation OEMs, warehouse and logistics developers, and humanoid manipulation labs as a pretraining layer that makes their own data collection more productive.

Simulation and industrial software vendors are a third customer bucket. Companies building digital twins or validation pipelines for manufacturing and autonomy already have distribution, and General Intuition's learned world models can slot in as a new layer of realism and controllability on top of existing stacks. That is a faster path to enterprise adoption than displacing those stacks entirely.

General Intuition's strongest structural expansion move is deepening the Medal data pipeline. Coalition Capital described the asset as billions of decision-labeled gameplay clips from over 17 million monthly active users, and CoreWeave's infrastructure case study references training on over one billion hours of video from Medal. As the model improves, it can extract more signal from that existing corpus, while Medal's ongoing clip generation expands coverage of environments, edge cases, and action patterns.

The next step would be acquiring real-world interaction pipelines such as robotics data collection teams, teleoperation infrastructure, or fleet learning capabilities. That would shorten the path from game pretraining to physical fine-tuning and close the sim-to-real loop with proprietary field data rather than relying entirely on transfer from virtual environments.

Sim-to-real transfer: General Intuition's commercial thesis depends on action priors learned from gameplay transferring into physical robot deployments, a bottleneck that industry analysts and competing labs still describe as unsolved at production scale, which could leave the company's valuation disconnected from monetizable robotics revenue for longer than its funding timeline implies.

Data governance: The company's training advantage runs through Medal's gameplay clips under a joint-controller arrangement with user opt-out rights and onward-sharing constraints, creating exposure to regulatory pressure, platform-policy changes, or shifts in user consent norms that could erode its proprietary data moat without any corresponding change in model quality or competitive position.

Platform commoditization: NVIDIA's Cosmos ecosystem, open-weight world foundation models, post-training recipes, and deep Omniverse integration with a partner list that already includes Figure and Skild AI, could become the default physical-AI middleware layer, forcing General Intuition to compete as one model provider among many on a platform it does not control rather than setting the standard itself.