How Startups Are Building Models of the World

AI models are beginning to move beyond generating content or predicting isolated outcomes. A growing group of startups is trying to build representations of entire systems that can anticipate how they will evolve, ideally in response to an action or intervention. The underlying bet is that AI will become more useful when it can reason about what may happen next rather than only interpret what exists now. This landscape examines 22 startups pursuing that idea across virtual, physical, biological and social systems.

(By the way, the idea for this landscape came after watching this YC video: World Models, Explained.)

What are world models and how are they different from LLMs?

A world model is an AI model that learns how an environment or system changes over time. It does not only identify what is currently happening. It tries to predict what may happen next, particularly when an action is taken.

• Learning how a system changes rather than only describing its current state. A model may observe a robot standing in front of a table, but a world model should also predict what will happen when the robot pushes an object or attempts to pick it up.

• A “world” does not need to mean the entire physical world. It can refer to any system whose future depends on its current state and the actions or events affecting it. This is why you can have world models about the weather, an industrial facility or a biological cell.

• Building an internal representation of the system. The model learns a compressed representation containing the information it needs to predict future states. An AI agent can then use this representation to explore possible outcomes before acting in the real environment.

• This creates the possibility of planning. Instead of immediately taking an action, an agent can estimate what may happen under several options and select the one most likely to achieve its objective.

• Learning dynamics rather than relying only on programmed rules. A traditional simulator is normally built by engineers who define the relevant objects, variables and physical equations. A world model attempts to learn at least part of these relationships from data.

• LLMs primarily learn patterns in language. A major difference with large language models is that they are trained to predict what is likely to come next in a sequence of text. Through this process, it learns large amounts of information about concepts, facts and how people describe the world.

• A world model is trained around a different prediction problem. It tries to estimate how the state of a system will change after time passes or an action occurs. That state might be represented through video, spatial information, sensor readings or biological measurements.

• The field remains early. Models can lose consistency as they predict further into the future, produce visually convincing but physically incorrect outcomes, or reproduce patterns without understanding the causal relationship behind them.

Category: General & Spatial Worlds

What is this category about?

• Building models that represent an environment beyond its current appearance. These companies want AI to do more than generate an image or continue a video. Their models aim to maintain an internal representation of a space as time passes, the viewpoint changes or an action is taken, allowing the output to behave more like a persistent world than a sequence of disconnected visuals.

• Betting that spatial intelligence will become a foundation for more capable AI. Language models learn from how the world is represented in text. Companies in this category believe models must also learn how objects occupy space, relate to one another and change over time. This could provide a different form of intelligence for planning, creation and eventually physical action.

• Treating world generation as a new computing interface. Digital environments currently require people to create assets, construct scenes and program interactions through specialized software. The underlying bet is that users may eventually describe or show a world and let a model generate much of the environment directly, turning interactive spaces into a more accessible and programmable medium.

• Using generated environments as places where AI agents can learn. A sufficiently consistent world model could allow an agent to test actions, encounter unusual situations and learn from failure before operating in reality. Some of these environments may therefore be designed primarily for other AI systems rather than for people to watch or explore.

• Pursuing generality rather than modelling one specific system. Unlike world models focused on robotics, weather or biology, this category aims to represent many types of virtual or physical spaces. That creates a wider range of possible applications, but also a harder technical problem because the model must remain useful across very different objects, environments and interactions.

What do products in this category do?

• A first approach is to create a world as a reusable spatial asset. World Labs’ Marble and SpAItial’s Echo turn a prompt or visual input into a persistent 3D environment that can be revisited, edited and viewed from new angles. The reason for giving the world an explicit spatial structure is that it can then enter existing creative, design or simulation workflows rather than disappearing once the initial output has been generated.

• A second approach treats the world as a process that is generated continuously. Odyssey and Decart produce the next frames of an environment in response to the user’s actions, allowing the experience to continue without every location and interaction being constructed in advance. This product model is closer to an AI-native game engine: the model does not retrieve a finished world but imagines what should happen next as the user moves through it.

• Some companies are trying to make these continuously generated worlds more accessible and personal. Overworld optimizes its Waypoint models for consumer GPUs and distributes them through the open-source Biome application. The underlying bet is that interactive world generation could become something people run locally, modify and build upon, rather than a capability controlled entirely through expensive cloud infrastructure.

• Another product approach begins with environments where actions are already recorded. General Intuition learns from gameplay because each recording connects what happened on screen with the inputs that caused the world to change. MIRA and Odyssey’s Agora extend this idea into shared environments where several people or agents can act at once. The objective is not merely to recreate the appearance of a game, but to learn the relationships between actions, participants and changing world states.

• Existing generative media platforms are entering the category through progressively more controllable models. Runway can build its general world models on top of the video-generation technology and commercial distribution it already developed for creators. This route starts with a useful media product and gradually adds real-time interaction, environmental memory and action control, rather than beginning with a standalone simulator that must still find its first market.

What are the use cases?

• Creating 3D environments for creative production. World models can turn a prompt, image or video into an explorable space that filmmakers, game designers and other creators can use for early concept development, virtual production or scene design. World Labs and SpAItial already allow generated environments to be explored, edited and exported, although these workflows remain less mature than established 3D production tools.

• Producing varied environments for robotics training and evaluation. A world model can change the visual appearance or layout of a simulated environment without requiring each variation to be created manually. Robotics teams can then expose their agents to more situations before deployment. World Labs has demonstrated this approach by connecting generated environments with established robotics simulators, while Decart increasingly positions Oasis around physical-AI use cases.

What are the major trends shaping this category?

• Generative AI is moving from producing finished media toward producing interactive experiences. Image and video models create an output that remains the same when it is replayed. World models allow the user to influence what happens next, making the result behave more like software than content. Google DeepMind’s Genie 3 and Microsoft’s Muse show that this shift is attracting major research investment from both frontier AI laboratories and the gaming industry.

• The potential market is expanding beyond worlds created for human entertainment. Google DeepMind is already connecting Genie generated environments with agents that learn to complete goals, while NVIDIA is positioning virtual worlds as training and validation infrastructure for robots, vehicles and other physical AI systems. The long term opportunity may therefore be larger in machine experience than in human consumption: world models could generate the situations through which other AI systems learn how to act.

• As impressive demonstrations become more common, value is likely to shift toward distribution and compatibility with the existing 3D ecosystem. A generated world becomes more useful when it can be exported, edited, embedded through an API or connected with established simulation and content tools. OpenUSD and NVIDIA Omniverse are providing a more accessible common infrastructure for these workflows, while products such as the World API are making generation programmable. This should favour companies that can make their worlds useful inside other products rather than keeping them inside a proprietary demonstration.

How does the funding environment look?

• The amount of capital entering the category is exceptional for such an early market. The latest headline rounds announced by AMI Labs, World Labs, Luma, General Intuition, Runway, Odyssey and Decart alone exceed $4 billion. This excludes their earlier financing and the capital raised by smaller companies in the landscape.

• Investors are financing the possibility of a new foundation-model layer before the commercial market has been proven. AMI Labs raised $1.03 billion at launch without a public product, while World Labs raised another $1 billion shortly after turning Marble into a commercial platform. Odyssey and General Intuition have also raised hundreds of millions despite remaining early in developer adoption. These rounds are bets on research talent, proprietary data and the cost of training frontier models rather than conventional software traction.

• The funding is closely connected to the infrastructure and distribution stack. Chipmakers, cloud providers and major software platforms repeatedly participate in these rounds. Their involvement can provide access to compute or future distribution in addition to capital. This is particularly important for world models because training and serving interactive visual environments require considerably more infrastructure than a conventional SaaS product.

• Two different financing models are emerging. Runway and Luma are using large rounds to expand from an existing creative product into broader model research and enterprise distribution. AMI Labs and several newer entrants are starting from the opposite direction, raising around a frontier research thesis before defining a mature product. Smaller companies such as SpAItial show that a focused spatial product can still enter with a more conventional seed round, but they face competitors whose financing allows them to fund research, infrastructure and global distribution at the same time.

Category: Robotics & Embodied Intelligence

What is this category about?

• Building models of the world from the perspective of a robot. General spatial models try to represent environments that people or agents can explore. Companies in this category focus more specifically on what a robot sees and how the environment changes when it moves, grasps or manipulates something. The model must connect sensory observations with physical actions closely enough to support decisions on real hardware.

• Learning the parts of physics that are difficult to program manually. Conventional simulators work well when the objects and rules of an environment can be modelled precisely. They become harder to construct for clothing, cables, cluttered spaces or contact rich manipulation. Companies in this category are betting that models trained on real robot data can learn some of these interactions directly, reducing the need to recreate every object and physical property by hand.

• Giving robots an internal space in which to anticipate actions. A world model can generate several possible futures from the same starting observation, each conditioned on a different action. The robot can then estimate which action is most likely to succeed before executing it. In the more ambitious version of this idea, the model is not only a simulator used during development but part of the robot’s intelligence while it operates.

• Pursuing embodied intelligence rather than a general model of every environment. These models are grounded in a particular robot, sensor configuration or class of physical tasks. This narrower scope allows them to learn the dynamics that matter for action, but it also means that a model trained for one embodiment may not immediately transfer to another. The category therefore ranges from custom simulators built for individual robotics teams to vertically integrated systems trained alongside a company’s own humanoid hardware.

What do products in this category do?

• The most immediate approach is a software replica of the robot’s working environment. Fern and One Robot build world models from recordings collected by a customer’s own robots, then let teams run their policies inside those learned environments. The point is not to reproduce an entire factory perfectly. It is to reproduce the parts of reality that determine whether a specific robot succeeds or fails, so engineers can test far more policy versions without repeatedly returning to the physical machine.

• Once the model can reproduce failures, it can also help create the experience needed to fix them. Fern is already using its simulators for reinforcement learning, while One Robot plans to generate more examples around situations where a policy performs poorly. This changes the role of simulation: instead of being a static test environment, it becomes part of an improvement loop that discovers weaknesses, creates targeted practice and sends a stronger policy back to the robot.

• Other companies place the world model inside the robot rather than around the development process. 1X uses a model to imagine how a task could unfold before translating that imagined sequence into movements for NEO. The model therefore acts less like an external simulator and more like a form of visual planning. This approach reflects a broader bet that a robot may learn many tasks more easily by predicting what successful action should look like than by programming each movement separately.

• Not every embodied world model tries to generate a visually detailed future. Stanhope AI focuses on helping a machine maintain and update a compact understanding of its surroundings as new information arrives. Its system is designed for drones and other autonomous machines that must decide what to do under uncertainty, often with limited computing power. Here, the value of the world model lies in adapting the plan when reality differs from what the machine expected.

• The category therefore spans two distinct product positions. Fern and One Robot sell infrastructure to other robotics teams, while 1X develops the model alongside its own hardware and can use every deployment to improve both. Saturn Dynamics is pursuing a broader platform between these positions, but remains too early to show whether its physics aware models will be sold as simulation infrastructure, embedded intelligence or a combination of both.

What are the use cases?

• Evaluating whether a new robot policy is actually better. A learned simulator can place several policy versions in the same initial situation and run far more trials than would be practical on hardware. This makes evaluation more reproducible and can expose failures that would remain hidden in a small set of manually observed demonstrations.

• Finding edge cases before a robot reaches production. Teams can vary object positions, actions or environmental conditions inside the model and observe where the policy stops behaving correctly. One Robot describes using these failures to determine which additional data should be collected, while 1X uses world-model predictions to estimate the likely success of NEO’s actions before wider deployment.

• Training policies without keeping real robots continuously in the loop. Reinforcement learning requires many repeated attempts, including actions that fail. Running all of them on hardware consumes operator time and can damage equipment. Fern has demonstrated policies trained inside its learned simulator and subsequently deployed on real robots, while its roadmap includes reusable environments for both offline and online reinforcement learning.

• Modelling difficult manipulation tasks. Clothing, ropes, boxes and loosely arranged objects are harder to reproduce with hand-authored physics than rigid and predictable factory parts. One Robot focuses on tasks such as textiles and box folding, while Fern has used bimanual rope manipulation and rice scooping to demonstrate that its model can respond to the actions of a real production policy.

• Supporting autonomous missions in unpredictable environments. Stanhope AI has demonstrated its world model through drone missions and has worked on tests involving delivery drones and defence applications. Here the use case is not to generate large volumes of photorealistic training video. It is to let an autonomous machine update its beliefs, search for missing information and change its plan when the environment no longer matches what it expected.

What are the major trends shaping this category?

• The demand for more capable robots is increasingly driven by structural labour constraints rather than technological novelty alone. Manufacturers, logistics operators and service businesses are using automation to protect productivity as skilled workers become harder to recruit and labour costs increase. World models matter within this broader shift because they could extend automation beyond repetitive tasks in carefully controlled environments and allow robots to cope with more variation.

• Physical AI has become a major strategic investment theme, but commercial deployment remains far behind the level of attention and capital. Humanoid robots are now being tested by industrial companies and backed by some of the largest technology groups, yet most projects remain pilots in controlled settings. Gartner expects fewer than 20 companies to reach production deployment in manufacturing and supply chains by 2028, suggesting that the coming years will be defined less by a sudden robot takeover than by a gradual search for the first repeatable applications.

• The competitive bottleneck is shifting from building an impressive robot to operating one reliably and economically. Progress now depends on whether companies can collect useful deployment data, integrate robots with existing operations and demonstrate that they are safe enough to run around people and valuable equipment. This should favour companies that treat the model, hardware and deployment process as one learning system, while standalone world-model providers will need to prove that they can improve performance across different robots rather than inside one controlled demonstration.

How does the funding environment look?

• Funding is concentrated in the vertically integrated robotics company rather than the specialist world model layer. 1X raised a $100 million Series B in January 2024 after its earlier $23.5 million Series A. The capital was intended to bring NEO to market, increase embodied AI data collection and support existing enterprise deployments

• The specialist software companies remain much earlier. One Robot entered Y Combinator’s Winter 2026 batch and is also backed by Accel, but still operates with a two-person founding team. Fern and Saturn Dynamics have not disclosed substantial institutional rounds publicly. This suggests that investors have not yet assigned the same capital intensity to learned simulation software as they have to humanoid platforms, partly because the standalone market and eventual pricing model remain less proven.

• Stanhope AI represents a third financing pattern built around differentiated research and strategic applications. The company raised £2.3 million in 2024 and an additional $8 million in February 2026 to develop its active inference model and expand field trials. Its investor group and early work in defence, drones and autonomous systems indicate that capital can also follow a scientifically distinct architecture tied to high stakes applications, even before broad commercial adoption is visible.

\

Category: Earth, Energy & Industrial Systems

What is this category about?

• Building models of systems that evolve continuously over time. Unlike spatial world models, these companies are not primarily trying to generate an environment that can be explored. They model processes such as the atmosphere, an electricity market or an industrial operation, where many variables interact and the value comes from predicting how the system will change.

• Betting that learned models can complement or partially replace hand built simulation. Weather forecasts and industrial models have traditionally relied on physical equations, engineering assumptions or manually configured rules. Companies in this category believe that models trained on large amounts of observational data can learn some of these dynamics directly, producing useful predictions more quickly and at a lower computational cost.

• Trying to understand the hidden state behind many separate signals. A turbine, construction site or weather system cannot be understood from one measurement alone. The relevant condition emerges from relationships between sensor readings, visual observations and external variables. These world models attempt to combine those partial signals into a representation of what is happening inside the larger system.

• Treating uncertainty and rare events as part of the product. These systems are often most valuable when conditions are changing or when an unusual event is approaching. A model must therefore do more than predict the average outcome. It needs to show the range of plausible futures or identify a transition that has rarely appeared in its training data.

• Connecting prediction with an operational decision. The end goal is not simply to produce a more accurate forecast. It is to help an energy trader adjust a position, an operator investigate a machine or a city respond to changing conditions. This distinguishes the category from more general world-model research: the representation of the system is built around a concrete decision loop.

What do products in this category do?

• Some products begin with a large physical system and turn its evolution into a decision tool. Jua models the atmosphere, but the commercial product is not simply a better weather map. Its forecasts are packaged around the questions energy traders actually face, such as how wind and solar production may change before a market closes. Athena takes this further by interpreting forecasts alongside prices and portfolio exposure, moving the product from prediction toward advice.

• Causal Labs starts from a similar system but is making a broader technical bet. Weather provides a large, continuously observed environment in which the company can train a model to understand how physical variables affect one another. The initial use case is forecasting, but the longer term ambition is to build models that can reason about causes and possible interventions across other physical systems.

• Archetype AI approaches the category from the opposite direction: it begins with the sensors already distributed across an operation. Newton combines measurements, video and language to infer what is happening inside a machine, construction site or infrastructure network. Its agents then turn that shared understanding into specific products for monitoring equipment, identifying unusual behaviour or checking whether work was completed correctly. The model is general, but the product becomes useful through narrowly defined operational questions.

• kausable focuses on the moments when a system stops behaving as it normally does. Its models try to discover the causal relationships inside complex systems and predict transitions that may have very few historical examples. The intended value is not simply improving the average forecast, but recognizing when an energy grid or another dynamic system may be approaching an unstable state. This remains an early product direction, but it gives the company a distinct position around rare and consequential events.

What are the use cases?

• Forecasting renewable energy production and supporting power trading. Wind and solar generation depend heavily on local weather conditions, while trading decisions must often be made before the latest conventional forecast becomes available. Jua’s models refresh frequently and provide ensemble forecasts for European markets, while Athena turns changes in weather, prices and portfolio exposure into briefings, scenarios and backtests.

• Understanding industrial operations from sensor and video data. Archetype AI worked with Kajima to search thousands of construction videos and measure how equipment was being used across different tasks. Its agents can apply the same model architecture to machine monitoring, anomaly discovery and process verification, allowing operators to investigate conditions that are not visible in one sensor stream alone.

• Monitoring infrastructure and public environments in real time. Archetype has also demonstrated agents for road safety and city infrastructure, where data must be interpreted close to where it is generated. These deployments show how a physical world model can operate continuously at the edge rather than sending every sensor stream to a centralized cloud application.

• Anticipating rare or dangerous system changes. kausable presents TipPFN as a way to identify critical transitions such as instability in an energy grid before the failure occurs. Causal Labs has a related but longer term ambition: use weather as a densely observed physical system from which a model can learn not only to predict future conditions, but eventually to identify actions that could influence them. These remain research led use cases rather than proven operational products.

What are the major trends shaping this category?

• ´AI forecasting is moving from an experimental technology into public infrastructure. NOAA and ECMWF now operate AI weather models alongside conventional forecasting systems, showing that learned models can become part of the trusted machinery behind real world decisions. As the underlying forecasting capability becomes more available, startups will find it harder to differentiate through model accuracy alone and will increasingly need to own proprietary data, specialized forecasts or the workflow through which a prediction becomes an action. 

How does the funding environment look?

• Funding is following companies that connect a technical model to an operational market. Archetype AI has raised $48 million across its seed and Series A rounds while packaging Newton into deployable industrial agents. Jua has raised $27 million while narrowing its initial Earth model ambition around energy traders and utilities. In both cases, investors are funding a foundation-model thesis, but the commercial story is attached to an identifiable workflow and buyer.

• Research first companies are also attracting meaningful seed rounds. kausable raised €12 million before showing broad customer deployment, largely on the strength of its causal-model research and the potential to predict critical transitions across domains. Causal Labs raised $6 million around a similarly ambitious Large Physics Model thesis, although its initial product and distribution strategy remain less developed.

• The capital requirements remain below those of general world model laboratories. None of these companies has raised the billion-dollar rounds seen in general spatial intelligence. Their models can be trained around narrower datasets and monetized through enterprise applications earlier. The trade-off is that the market may remain fragmented, with different models emerging for weather, industrial sensing and rare-event prediction rather than one system becoming the universal model of the physical world.

Category: Biology

What is this category about?

• Building models of living systems rather than individual biological objects. Earlier biological AI models often focused on one bounded problem, such as predicting a protein’s structure from its sequence. Companies in this category have a broader ambition: represent how molecules, cells and tissues interact, then predict how the system may change when a gene, drug or environmental condition is altered.

• Betting that biology can be understood across multiple scales. A mutation may change a protein, which affects a cell, alters its surrounding tissue and eventually influences whether a patient develops or responds to a disease. The companies in this category are trying to connect these layers rather than analyse each one in isolation. This explains why they describe their systems as virtual cells, digital organisms or world models of biology.

• Treating biological context as part of the prediction. A cell does not behave the same way in every patient or tissue. Its response depends on its type, internal state and neighbouring cells. These companies therefore combine molecular measurements with spatial information, pathology images and clinical outcomes to represent not only what biological components are present, but the environment in which they operate.

• Aiming to make biological experimentation more selective rather than replacing it. Living systems are too variable and poorly understood for the models to become self contained simulations today. Their near term value lies in narrowing the search space: identifying the most promising targets, experiments, biomarkers or patient groups before committing time and capital to physical studies.

What do products in this category do?

• One product approach tries to connect previously separate biological models into a common system. GenBio AI’s AIDO includes models for DNA, RNA, proteins, cells and tissues, while Bioptimus is expanding from its H-Optimus pathology models into the multimodal M-Optimus platform. The reason for building this broader architecture is that a useful model of biology cannot stop at one data type: it needs to relate molecular information to what eventually appears inside cells, tissues and patients.

• A second approach concentrates on the cell and its immediate environment. Noetik’s OCTO-vc places virtual cells inside spatial representations of real tumour tissue, while GenBio AI’s AIDO.Tissue models cells through their surrounding neighbourhoods. These products exist because cellular behaviour is not determined by gene expression alone. A cancer cell’s interaction with immune cells, fibroblasts and nearby tissue can be as important as its own internal state.

• Some companies enter the market through a narrower model that is already useful on its own. Bioptimus began with H-Optimus for pathology images, while GenBio AI releases separate modules for areas such as protein design and spatial biology. This approach gives researchers a practical tool before the complete biological world model exists, while also generating usage, feedback and integrations that can support the broader platform over time.

• Another product type combines the model with a proprietary biological data engine. Noetik generates multimodal tumour data and uses Perturb map to conduct functional experiments, while Bioptimus launched STELA to create clinically linked datasets spanning pathology, spatial biology, molecular profiles and patient outcomes. The underlying logic is that biological models cannot improve through public data and computing power alone: companies also need access to standardized experiments and matched patient data that reveal how biological states change.

What are the use cases?

• Predicting which patients are most likely to respond to a treatment. Noetik’s TARIO-2 uses routinely collected pathology images to identify patient groups associated with better response and survival outcomes. Bioptimus similarly positions M-Optimus around biomarker discovery and clinical trial stratification. The goal is to replace broad patient selection criteria with a richer representation of the biology driving treatment response.

• Finding therapeutic targets that are more likely to work in humans. Researchers can use virtual cell models to estimate how cells or tissues respond when a gene is removed or a pathway is altered, then prioritize the interventions that appear most promising for laboratory testing. Noetik combines this modelling with Perturb-map experiments, while GenBio AI is developing models for predicting cellular responses to genetic and chemical changes.

• Designing new biological structures and selecting better experiments. GenBio AI extends its platform into generative protein design, while its broader virtual cell strategy aims to help researchers identify which physical experiments would provide the most useful new information. The model becomes a tool for deciding what to build or test rather than only interpreting data after an experiment has occurred.

What are the major trends shaping this category?

• The virtual cell is becoming a recognized strategic objective across the research ecosystem. Arc Institute launched a public challenge focused on predicting the effects of genetic perturbations, backed by NVIDIA, 10x Genomics and Ultima Genomics. Scientific publications and research initiatives increasingly frame virtual cells as a major next step after protein-level models, suggesting that the category is becoming a shared research direction rather than the narrative of a few startups.

• Biological AI is moving from large collections of one dimensional data toward multimodal and spatial representations. Gene expression data can describe what is active inside a cell, but it often loses information about where that cell sits and how it interacts with surrounding tissue. The field is therefore combining genomics, proteomics, imaging, spatial profiling and clinical records to produce more complete models of cellular states and disease.

• The competitive focus is shifting from model size toward proprietary data and experimental validation. Independent evaluations have found that some large single-cell foundation models do not consistently outperform simpler approaches, making ambitious “virtual cell” claims insufficient on their own. Companies will increasingly be judged by whether their predictions survive laboratory tests, reproduce across institutions and improve real drug development decisions. This favours businesses that can close the loop between model predictions, biological experiments and clinically linked data.

How does the funding environment look?

• Investors are willing to fund biology foundation models at unusually large early stage levels. Bioptimus raised $35 million at seed and another $41 million in its Series A, while Noetik raised $14 million at seed and $40 million in its Series A. These rounds are substantially larger than typical software seed financings because the companies need to fund AI research, biological data generation, scientific talent and partnerships with laboratories or healthcare institutions at the same time.

• Data ownership is becoming part of what investors are financing. Noetik has built a large proprietary collection of multimodal tumour data, while Bioptimus is using part of its capital and partnerships to construct STELA. These assets may become more defensible than the model architecture itself because they require access to physical samples, clinical outcomes, institutional relationships and standardized data-generation processes.