General Intuition

1.1 Identity, origin, and operating thesis

General Intuition is described by multiple independent publications as a New York-based AI research lab spun out of Medal, the gameplay-clip platform co-founded by Pim de Witte. Its stated product direction is a commercial API in which frames are inputs and predicted next actions are outputs, rather than a consumer application catalog. Reporting says the company uses gameplay clips with embedded button-and-time action labels to train models for perception, prediction, and action across virtual and physical environments. The company says its world model is a training environment, while the intended product is the agentic model. Public profiles disagree on whether the company was founded in 2014, 2015, or 2024; October 2025 is supportable only as the public launch date. This creates a coherent identity and product thesis, but public evidence does not yet establish scaled customer revenue or repeatable simulation-to-real-world transfer.[CO001, CO002, CO003, CO004, CO016]

1.2 Leadership, governance, and stakeholder map

The public leadership record centers on co-founder and CEO Pim de Witte and co-founders Eloi Alonso, Adam Jelley, and Vincent Micheli. Coverage links de Witte's Medal operating history to the company's proprietary-data strategy and identifies a chief product officer, Kent Rollins, in the operating team. The latest round was led by Khosla Ventures, with participation reported from General Catalyst, Jeff Bezos, Eric Schmidt, Nico Rosberg, and researchers at Google DeepMind and MIT; another account names General Catalyst as lead, so the lead attribution should be treated as conflicting until primary financing documents are available. The available evidence supports technical-founder fit and a concentrated stakeholder base, but not board composition, liquidation preferences, voting control, or key-person mitigation.[CO005, CO006, CO007, CO008, CO009]

1.3 Financing, scale, and milestone implications

General Intuition publicly announced a $133.7 million or approximately $134 million seed round in October 2025 and a $320 million Series A in June 2026 at a reported $2.3 billion valuation. Those rounds imply $453.7 million to $454 million of disclosed funding, but neither revenue, ARR, profitability, burn, runway, nor customer contract value is verified in the assigned evidence. The latest financing is reported as funding compute capacity, pre-training the next model, and a broader API release, while the company has demonstrated a gameplay agent and a quadruped robot in press coverage. The central diligence question is whether gameplay-derived data produces reliable physical-world transfer at scale; the evidence explicitly leaves that question open.[CO010, CO011, CO012, CO013, CO014, CO015]

1.4 Exhibits

2.1 Market boundary and buyer logic

The relevant market is not the entire robotics hardware market and not generic video generation. It is the software layer for action models, learned world models, simulation, and agent evaluation: systems that ingest visual state and control signals, predict consequences, or select actions before and during deployment. The adjacent spend includes robot foundation models, digital twins, simulation platforms, adaptive game agents, and autonomy testing. Excluded from this boundary are robot bodies, actuators, general cloud infrastructure, and passive video analytics unless they are purchased as part of an action-model workflow. Public sources describe a B2B API model and early customers across gaming, simulation, and robotics, which makes the immediate budget owner more likely to be a product, simulation, autonomy, or machine-learning engineering team than a consumer-gaming buyer. The buyer may be a robotics platform lead, game-studio AI lead, or simulation/evaluation owner; the user is the research or engineering team operating the model, and the payer remains unverified.[CM001, CM002, CM003, CM004, CM005]

2.2 Sizing lenses and what remains unproven

The public record supports multiple market lenses, but not a single additive market number. The data-supply lens is unusually concrete: reporting describes Medal as producing billions of gameplay videos or clips and as carrying action-labeled inputs, while other coverage cites 10 million or more than 17 million monthly users. Those figures describe data supply, not revenue or addressable spend. The workflow lens is narrower and more commercially useful: model access, learned simulation, policy testing, adaptive agents, and pre-deployment evaluation. The company has described a broader API rollout and a small number of customers, but public sources do not disclose API prices, contract values, renewal rates, customer concentration, or downstream performance uplift. The robotics transfer lens is therefore an adoption hypothesis rather than a market-size proof: a quadruped demonstration and an eight-minute fine-tuning claim show technical direction, not repeatable production demand. A valuation and funding round indicate investor appetite, not TAM or revenue. The market should remain evidence-constrained until comparable paid deployments and benchmarked savings are public.[CM006, CM007, CM008, CM009, CM010, CM011]

2.3 Adoption drivers, constraints, and diligence implications

Adoption is pulled by the cost and scarcity of real-world robot data, the desire to test policies in safer digital environments, and the prospect of reusing one action-oriented model across embodiments. Action labels can give buyers information that passive video lacks, and a model provider can sit upstream of many robot, drone, simulation, and game workflows. The constraints are equally material. Games simplify physics and operating conditions; public coverage explicitly identifies transfer to real-world conditions as unresolved. Customers will need measurable transfer benchmarks, integration into existing simulators and deployment stacks, safety controls, audit logs, and clear data rights. Compute intensity can make gross margin and price-performance difficult to prove, while vertically integrated robotics companies may keep training data and model economics inside the hardware stack. The near-term adoption trigger is therefore a credible API or pilot that reduces data collection or evaluation cost for a named workflow. The key diligence asks are paid-customer count, buyer and payer identity, conversion from pilot to production, benchmarked reduction in real-world data, usage-based pricing, and the failure rate under messy physical conditions.[CM012, CM013, CM014, CM015, CM016]

2.4 Exhibits

3.1 Competitive landscape and categories

General Intuition competes across several layers rather than against one uniform product. TechCrunch describes DeepMind and World Labs as world-model makers selling models for agent training and content creation, while General Intuition says world models are an enabling layer and emphasizes agents and applications. The same source places the company against deterministic game bots in gaming and against robotics or drone systems in physical settings. Latent Space describes World Labs, Yann LeCun's work, and other world-model efforts as part of a broader spatial-intelligence race. The practical landscape therefore includes direct world-model peers, embodied-AI model builders, multimodal video systems, game-engine and bot infrastructure, and status-quo internal behavior trees. Publicly fetched evidence does not establish that every named peer sells a directly comparable API, so the profiles below distinguish direct overlap from adjacency and mark unverified fields rather than filling them with estimates.[CP001, CP002, CP003, CP004]

3.2 Capability, distribution, and trust comparison

The most defensible capability distinction is the input-to-action loop. TechCrunch reports that General Intuition's agents use visual input and controller actions, can understand environments not seen in training, and are intended to transfer to robotic arms, drones, and autonomous vehicles. Latent Space describes a fully vision-based approach, action prediction, partial observability, and a frames-in/actions-out API for games, engines, and controller-based robots. This creates a potential advantage over content-oriented world-model products when the buyer needs control rather than video generation, but it also creates a harder proof burden: public commercial models have not yet been shown on a common benchmark, and the fetched sources do not provide comparable latency, accuracy, price, uptime, or customer-retention data. Distribution is similarly unresolved; General Intuition has early-access positioning and planned broader API availability, while incumbents and larger labs may have stronger channels, compute, and developer ecosystems.[CP005, CP006, CP007, CP008]

3.3 Moat durability and switching costs

General Intuition's moat case combines proprietary or preferential access to Medal's gameplay corpus, action labels attached to video, and the founders' world-model research credentials. TechCrunch reports that Medal had 2 billion videos per year from 10 million monthly active users across tens of thousands of games and says the company views the selection of positive and negative gameplay examples as useful training data. Latent Space characterizes the clips as episodic memory for simulation and describes a transfer path from games to realistic games, real-world video, and robotics. Those assets can raise switching costs for a model buyer that values a distinctive data pipeline, but they do not by themselves prove exclusive rights, sustained data quality, physical-world transfer, or customer lock-in. A buyer can also multi-home across model providers, use internal behavior trees, or commission task-specific robotics models. Pricing, contract length, integration cost, and partner exclusivity remain unreported in the assigned evidence.[CP009, CP010, CP011, CP012]

3.4 Displacement risk and diligence gaps

The adverse case is not that General Intuition lacks a technical thesis; it is that the thesis may remain difficult to monetize before competitors, incumbents, or internal teams close the same capability gap. TechCrunch says the startup is deliberately not trying to produce models that compete with game developers and frames gaming applications around replacing deterministic bots with agents whose difficulty can scale. That focus narrows the product claim but leaves questions about procurement, safety, deployment economics, and willingness to replace established tooling. The assigned sources also report rapidly changing valuation and financing narratives, but none supplies a neutral head-to-head benchmark, public customer contracts, realized pricing, renewal data, or a quantified transfer result from gameplay to physical robots. These gaps make the competitive conclusion conditional: data and research are credible differentiators, while durability and commercial displacement remain unresolved.[CP013, CP014, CP015, CP016]

3.5 Exhibits

4.1 Revenue model and commercial proof

The disclosed commercial model is B2B access to agents through selective early-access arrangements and a commercial API, rather than sale of the underlying world model itself. GamesBeat reported that initial partners in gaming, simulation, and robotics had been onboarded, while MLQ and Toflow attribute approximately $40 million of 2025 revenue to early-access arrangements. That figure is useful evidence of willingness to pay, but it is neither audited nor identified as ARR, and the retained sources do not disclose customer concentration, contract duration, recognition policy, or renewal performance. The investable interpretation is therefore early contracted demand, not proven repeatable software economics.[CI001, CI002, CI003, CI013, CI015]

4.2 Pricing and go-to-market

Commercialization appears deliberately narrow. June 2026 reporting described a handful of early partners and a planned broader API release, but the allocated evidence contains no list price, usage rate, minimum commitment, discount schedule, revenue-share term, or realized price. The team was also described as roughly 25 people, indicating a founder- and research-led motion rather than a scaled sales organization. Until management provides pipeline conversion, sales-cycle, churn, expansion, and channel economics, CAC and payback cannot be estimated responsibly. The broader API plan is a distribution milestone, not evidence that self-serve or repeatable demand has materialized.[CI002, CI003, CI009, CI012]

4.3 Cost structure and unit economics

Public evidence points to compute as the dominant near-term cost driver. CEO Pim de Witte said GPUs are expensive, and GamesBeat reported that the company expected to rent capacity under long commitments; Toflow separately described the financing as supporting compute scaling and model pre-training. This suggests a cost base with substantial committed infrastructure before broad product availability. It does not establish gross margin, because the sources do not separate training from inference, disclose utilization, or show whether customer pricing passes through compute. Revenue per customer, inference cost, support burden, gross margin, CAC, payback, and net retention all remain unavailable.[CI008, CI012, CI013, CI014]

4.4 Capital adequacy and financial verdict

The latest completed financing in the retained evidence is the $320 million round that closed in January 2026 and was announced in June at a $2.3 billion valuation, bringing disclosed funding to about $454 million. This is financing capacity, not cash available: none of the reviewed materials discloses cash on hand, burn, runway, debt, project-finance obligations, or remaining GPU commitments. GamesBeat also reported conversations about an additional Series B, reinforcing financing dependence as model progress accelerates. The adverse case is material: Santage argued that investors were pricing the technical thesis rather than revenue and highlighted the still-open sim-to-real gap. The financial verdict is promising revenue optionality, unproven revenue quality, unknown margins, and high capital intensity.[CI004, CI005, CI006, CI007, CI010, CI011, CI016]

4.5 Exhibits

5.1 Product surface and user workflow

General Intuition's public product concept has two linked components: a world model that predicts how an environment evolves after actions, and an action model that chooses the next action toward a goal. The immediate user workflow is not a consumer robot sale. For game developers, the proposed surface is adaptive non-player characters, QA bots, and interactive simulation; for robotics developers, the proposed workflow is to supply environment data or a task specification, pretrain or fine-tune inside a learned simulator, and then deploy to a physical system. This is a credible workflow description, but the public evidence still describes early-access or design-partner activity rather than a generally available, independently benchmarked production product. The company has also positioned world models as training infrastructure rather than as a game-content product, partly to avoid competing with game developers.[CE001, CE002, CE003, CE004]

5.2 Data and model architecture

The technical differentiator is the representation of gameplay as visual frames paired with action labels rather than raw keystrokes. In the Naavik interview, Pim de Witte says Medal converts controls to action labels in memory, preserving the action-to-outcome loop while avoiding capture of the literal key or controller input. TechCrunch separately reports that the corpus contains large volumes of gameplay clips and that the company views their first-person perspective and edge-case selection as useful for training. MIRA provides the clearest public system demonstration: a playable multiplayer world model trained on Rocket League gameplay that reportedly runs at 20 frames per second and reproduces several game dynamics without a conventional physics engine or 3D renderer. These facts support a specific architecture thesis, not a claim that the same performance transfers to physical environments.[CE005, CE006, CE007, CE008, CE009]

5.3 Deployment, differentiation, and diligence gaps

General Intuition's moat case rests on the combination of Medal's distribution and data flywheel, the action-conditioned representation, and research progress from game environments toward physical systems. The company has publicly discussed game-developer feedback on models and an internal goal of a broadly available API, while the recruiting surface shows a small product team building desktop and mobile experiences with design-system and accessibility responsibilities. The principal implementation risks are practical: the Naavik discussion identifies multiplayer consistency, long-horizon coherence, and inference cost as unsolved limitations; it also says game bots were not yet deployed at the time of the interview. Public material does not establish API uptime, latency targets, access controls, incident history, model-evaluation protocols, data-rights documentation, or customer production outcomes. Those omissions should be treated as diligence blockers rather than filled with marketing claims.[CE010, CE011, CE012, CE013, CE014]

5.4 Evidence exhibits

The exhibits separate what is publicly demonstrated from what remains a roadmap or diligence request. The module matrix emphasizes maturity and user context; the workflow table distinguishes a proposed operating path from measured customer benefit; and the architecture table records dependencies and technical risks. The stack figure is a compact map of the disclosed data-to-deployment path, not a claim that every layer is production-ready. Readers should treat the unverified controls and deployment fields as explicit diligence asks, not as implied capabilities.[CE003, CE006, CE010, CE012]

5.5 Exhibits

6.1 Buyer and use-case segments

Public materials describe a B2B foundation-model platform whose prospective buyers are game studios, simulation providers, and robotics developers. Sacra describes a workflow in which a customer supplies an environment or robot task and uses the API to pretrain or fine-tune an action model, while gaming use cases include adaptive NPCs, QA bots, and interactive simulations. I-Scoop says the first API customers span gaming, simulation, and robotics, but does not name them or disclose contract values. This is materially different from proof of broad production adoption: the available evidence supports a multi-vertical target market and early partner activity, not a verified customer count or repeatable sales funnel. The buyer may be a model or simulation team, the end user may be a game player or robot operator, and the economic payer is not publicly identified.[CU001, CU002, CU003, CU004]

6.2 Adoption and usage evidence

The clearest measurable usage signal is upstream: TechCrunch reported that Medal had 10 million monthly active users and two billion videos per year in its 2025 coverage, while Sacra later described more than 17 million monthly active users and billions of clips per year. These figures describe the data-source platform, not General Intuition paying accounts, and the differing periods and denominators should not be merged into a customer-growth series. General Intuition's own model demonstrations, including a quadruped transfer using eight minutes of real-world fine-tuning data, demonstrate technical use rather than customer retention. The company has stated that a broader API release was planned for summer or early autumn 2026, so the transition from selected partners to a wider developer base was a key adoption milestone rather than an observed outcome in that coverage. By October, InvestGame reported a commercial API serving a small number of partners and a partner waitlist opened on September 29, which is evidence of a controlled commercial rollout but still not proof of broad adoption.[CU005, CU006, CU007, CU008, CU015]

6.3 Named proof, durability, and concentration

Public reporting names technical collaborators and describes early API customers, but it does not provide a customer roster with production status, contract length, renewal dates, annual recurring revenue, gross retention, or net revenue retention. Sacra names Kyutai and Epic Games in connection with the MIRA research demonstration, and describes a commercial API for game studios and robotics developers; those are useful relationship signals but do not establish paid production deployments. The available sources also do not identify the largest customer, the share of revenue represented by any partner, or whether Medal-related data access is economically separate from customer revenue. Accordingly, the customer thesis rests on plausible use cases and a potentially deep data flywheel, while durability and concentration remain diligence questions.[CU009, CU010, CU011, CU012]

6.4 Exhibits

7.1 Ranked residual-risk assessment

The most consequential risk is technical transfer: independent coverage still describes the move from simulated gameplay to unpredictable physical settings as the central challenge, while public comparable benchmarks for the commercial models remain limited. The next tier combines commercialization and capital intensity. Public sources describe a gradual or selective API rollout, a large compute commitment, and heavy pre-training expenditure, but do not provide verified revenue, ARR, profitability, burn, or runway. The $320 million Series A provides financing capacity, not proof that model economics, product reliability, or customer demand are durable. The heatmap therefore scores residual exposure, not incident history, and treats absent public proof as a diligence gap rather than evidence that a risk has not occurred.[CR004, CR005, CR007, CR008, CR009, CR013, CR016]

7.2 Regulatory, legal, and safety exposure

The assigned evidence pool contains no regulator filing, court record, certification, trust center, data-processing agreement, incident report, or formal training-data rights analysis for General Intuition. TechCrunch reported that management was avoiding content generation use cases in part to reduce copyright issues, which is a useful design signal but not a legal clearance for using action-labeled gameplay clips or deploying models in robotics. APIs.io applies an AI-and-biometrics regulatory posture to the provider and reports no machine-readable API contract, but that third-party classification is not an official compliance assessment. Before investment, counsel should trace gameplay-data consent and license rights, model-output ownership, privacy obligations, and allocation of product and bodily-injury liability across each planned deployment.[CR001, CR002, CR014]

7.3 Operational, quality, and security risk

Operational readiness is difficult to separate from research progress. A gameplay-derived model can demonstrate coherent behavior without establishing robustness under sensor noise, distribution shift, adversarial inputs, latency constraints, or safety-critical failure modes. Public material also leaves the commercial API's service levels, authentication controls, vulnerability process, data retention, evaluation suite, and incident response unverified. The concentration of spending on CoreWeave capacity and pre-training increases exposure to infrastructure availability, model-training setbacks, and unfavorable inference economics. Required diligence should include out-of-distribution tests, failure logs, red-team results, uptime history, disaster-recovery evidence, and cost-per-task curves from representative customer workloads.[CR005, CR007, CR008, CR011, CR013, CR014, CR016]

7.4 Dependencies, financial-model risk, and thesis breaks

The strategy depends on Medal for differentiated action-labeled data, external compute for frontier-scale training, and a modular market in which robotics and game developers buy an upstream model rather than build an integrated data-model-hardware loop. Sacra explicitly argues that vertical integration weakens the standalone model-layer position. Public funding data shows substantial capital raised, yet the available sources do not verify gross margin, burn, runway, contracted backlog, renewals, customer concentration, or financing terms; an estimated round dilution is not a cap-table fact. Thesis-break triggers should include failure to reproduce transfer gains on customer-owned hardware, continued absence of production API evidence, loss or material restriction of Medal data rights, sustained compute cost above customer willingness to pay, or financing terms that imply the operating plan cannot be funded on acceptable dilution.[CR001, CR003, CR004, CR008, CR009, CR010, CR011, CR012, CR015, CR016]

7.5 Exhibits

8.1 Recommendation and valuation stance

General Intuition's public financing record shows unusually rapid mark expansion: Axios reported a $320M Series A at a $2.3B post-money valuation on June 26, 2026, while a September 29 financing was reported by Pomegra at $220M and $6.2B. The latter is a reported financing datapoint, not a public audited valuation mark, and the source does not disclose revenue. The valuation case therefore rests on the scarcity and strategic value of Medal's action-labeled gameplay data, the founding team's spatial-temporal research credentials, and the possibility that game-trained models transfer into robotics, simulation, or autonomous systems. Against that, the available evidence does not establish paid customer scale, recurring revenue, margins, cash burn, liquidation preferences, or a comparable operating multiple. I recommend research-more, with a high risk rating and an expensive stance at the reported September price. A lower entry price, verified contracted revenue, or demonstrated transfer economics could move the call toward track or buy; a failed transfer milestone, down round, or weak customer conversion would reinforce avoid.[CV001, CV002, CV003, CV004, CV005]

8.2 Investment thesis and anti-thesis

The positive thesis is that General Intuition combines a hard-to-replicate data position with a research team that has already demonstrated world-model work, and that its financing syndicate has repeatedly backed the company. TechCrunch described Medal's platform as holding roughly 2 billion videos per year from 10 million monthly active users, with player inputs attached to clips, while the company is pursuing agents that reason over space and time. That is a credible strategic asset, but not yet a proven commercial moat. The anti-thesis is that action labels and gameplay diversity may not transfer cleanly to physical environments, and the latest step-up in valuation occurred without public revenue or named customer evidence. Pomegra explicitly frames the mark as a bet on scarce embodied-AI training data and says the company has disclosed no revenue. The key diligence question is whether customer pilots convert into repeatable, margin-positive contracts before the next financing.[CV006, CV007, CV008, CV009, CV010]

8.3 Scenarios and comparable evidence

A conventional revenue-multiple valuation cannot be applied responsibly because no retained source verifies current revenue, ARR, gross margin, or a customer count. The scenario table therefore uses valuation bands as decision thresholds, not forecasts or fair-value estimates. The bull case requires evidence that the $6.2B round reflects a durable financing market and that gameplay-trained models reduce customer data or deployment costs in robotics, simulation, or entertainment. The base case assumes the financing is genuine but the commercial model remains early and disclosure remains sparse; in that case, the reported price is stretched or expensive for an investor without private terms. The bear case assumes a domain-transfer shortfall, delayed API monetization, or multiple compression; the public record does not permit a probability estimate, so the signals are qualitative. Public-company yardsticks in the Silicon Valley Investor Club tear sheet are historical IPO reference points, not a peer set with comparable revenue, stage, or business model.[CV004, CV011, CV012, CV014]

8.4 Entry discipline and thesis-break triggers

Before underwriting a new entry, obtain the signed financing documents, post-money capitalization, option-pool treatment, liquidation preferences, pro-rata rights, debt or compute commitments, and the exact closing date for the reported Series B. Verify revenue by customer and contract type, renewal and gross-margin data, compute cost per inference or training milestone, cash balance, burn, runway, and any concentration in Medal or a small number of partners. The public evidence also leaves open whether the broader API release occurred and whether early-access partners are paying customers rather than pilots. A thesis break would be a failed independent transfer test from gameplay to a physical or simulated task, a material down round, customer pilots that do not renew, or economics that require continued financing faster than commercial conversion. Conversely, independently verified paid deployments, repeatable gross-margin improvement, and disclosed terms that protect new investors would support a less defensive stance.[CV015, CV016]