Why The Next AI Investment Cycle May Begin Inside Virtual Worlds
The current artificial intelligence boom was built on language. Models absorbed vast quantities of text, learned how words relate to one another and turned that training into systems capable of writing, coding, summarising and answering questions. The next stage may require a different kind of data: not descriptions of the world, but records of what happens when something moves through it.
This is why videogames are beginning to attract attention from AI investors for reasons that have little to do with entertainment. A game contains space, movement, objectives, obstacles and consequences. A player opens a door, changes direction, operates a machine or reacts to another character, and the environment responds. These interactions create data about actions unfolding over time—the type of information machines will need if they are expected to navigate warehouses, control robots, assist drivers or operate independently in physical environments.
The emerging field is often described through “world models”: systems designed to develop an internal representation of how an environment works, predict what may happen next and choose an appropriate action. Their investment significance lies in the possibility that AI’s next large market may extend beyond software that interprets language and into machines that understand physical reality.
Videogames could provide part of the training ground. The more important investment question is who will own the computing infrastructure, simulation tools, specialist data and commercial applications built around that transition.
Language Models Describe the World but Do Not Necessarily Understand It
Large language models are effective because the internet contains an enormous written record of human knowledge. They can identify patterns across books, websites, code repositories, reports and conversations, then produce plausible responses based on those relationships.
Yet language is an abstraction. A written explanation of how to climb a staircase is not the same as balancing, adjusting weight and reacting when a foot lands incorrectly. A manual can describe how a warehouse operates without capturing every spatial decision made by a worker moving among people, shelves and machinery.
Many human capabilities develop through observation and interaction rather than formal instruction. Children learn to move by trying, failing and adjusting. Animals understand terrain without reading descriptions of it. Skilled workers often react to physical signals they would struggle to express fully in words.
AI systems intended to operate in the physical world face a similar constraint. They need to understand where objects are, how they move, what changes after an action and which consequences are likely. Text can contribute context, but it cannot provide the entire training environment.
This helps explain the interest in world models. Instead of predicting only the next word, these systems attempt to predict the next state of an environment. Given a scene and a proposed action, the model estimates what should happen next.
The commercial implications extend far beyond gaming. A model capable of understanding actions in space and time could support robotics, autonomous vehicles, industrial automation, defence systems, logistics and advanced digital assistants. It may also allow AI agents to complete tasks that require continuous interaction rather than a single written answer.
Why Games Offer Something the Internet Cannot
Real-world interaction data is difficult and expensive to collect. A company training a robot must operate physical machines, capture their movements, label outcomes and manage the safety consequences of failure. An autonomous-driving system needs enormous quantities of road footage, sensor information and edge cases that may occur only rarely.
Videogames and simulators offer a more controlled alternative. They provide complex environments in which actions can be repeated, modified and observed at scale. An AI system can attempt a task thousands of times without damaging a vehicle, injuring a person or interrupting a factory.
A flight simulator contains information about navigation, instruments and changing conditions. A driving game provides roads, traffic patterns and vehicle behaviour. A cleaning or maintenance simulator records sequences of physical actions and their results. Even fictional games can teach more general relationships involving distance, orientation, objects, goals and cause and effect.
Some developers are therefore training models across thousands of different games rather than concentrating on one environment. The purpose is not to create an AI that becomes exceptionally good at a particular title. It is to identify transferable patterns across varied situations.
This distinction matters. Earlier gaming-related AI milestones often involved systems mastering a single game under narrowly defined rules. World-model developers are pursuing something broader: intelligence that can adapt when the environment, objective or available action changes.
Games also produce unusually rich behavioural data. They show not only what is visible on screen, but what a player chose to do, which sequence of actions followed and whether the decision succeeded. That relationship between observation, action and outcome is valuable for training systems expected to act independently.
The Opportunity Is Larger Than the Gaming Industry
The most obvious conclusion would be that videogame publishers are likely to benefit. Some may. Large publishers control valuable interactive worlds, proprietary engines and vast quantities of player behaviour. Their assets could become useful to AI developers, particularly where licensing arrangements allow game data to be used for training.
The direct investment case is less straightforward. Owning a successful game does not automatically give a company the technical capability or legal freedom to transform player interactions into commercially valuable AI training data. Privacy, copyright and contractual restrictions could limit how that information is used. Publishers may also prefer to improve their own products rather than become data suppliers to robotics or industrial companies.
The broader opportunity may lie with businesses that build the infrastructure connecting simulation and physical-world AI.
Game-engine companies are an obvious part of this ecosystem. Their platforms already create three-dimensional environments, model physical interactions and support real-time simulation. The same tools can be adapted for training robots, testing autonomous systems or constructing digital replicas of factories and cities.
Chipmakers and cloud providers remain central because world models are computationally demanding. Training systems across video, movement and simulated environments may require more processing than text-based models alone. Inference could also become more intensive if robots or vehicles must interpret their surroundings and update decisions continuously.
Sensor manufacturers, robotics platforms and industrial-software companies provide another layer. World models become commercially important only when they connect to machines capable of acting. A model that predicts how an object will move needs cameras, lidar or other sensors to perceive it and control systems to respond.
The investment landscape therefore resembles a stack rather than a single sector. It includes simulation, data, processors, cloud infrastructure, sensors, robotics and application-specific software. The companies creating the most recognisable AI models may capture only part of the eventual value.
Robotics Could Become the First Major Test
Robotics is one of the clearest commercial applications for world models. Conventional industrial robots perform highly repetitive tasks in controlled environments. They are efficient when every object arrives in the expected place and every movement can be programmed in advance.
More adaptable robots need a different form of intelligence. A warehouse robot may have to navigate around people and unexpected obstacles. A machine in a factory may need to handle objects that vary slightly in shape or position. A domestic robot would face environments far less predictable than an assembly line.
World models could help these machines anticipate the consequences of their actions. Before moving an arm or changing direction, the system could estimate what is likely to happen, compare possible outcomes and select the safer or more effective option.
Simulation is particularly valuable here because physical training is slow. Robots are expensive, mechanical components wear out and mistakes can damage equipment. A virtual environment allows developers to generate far more experience before transferring the model to a real machine.
The transfer is not seamless. Simulations simplify reality, while physical systems encounter friction, imperfect sensors, lighting changes and unpredictable human behaviour. The gap between simulated and real environments remains one of the field’s central technical challenges.
Even so, progress does not require a perfectly general robot. Commercial value can emerge from narrower systems that become reliable in logistics centres, factories, mines, farms or hospitals. Investors should therefore watch for companies that combine advanced models with specific operating environments and access to proprietary real-world data.
Autonomous Systems Need More Than Better Vision
Self-driving vehicles have already demonstrated the difficulty of moving from perception to action. Recognising a pedestrian or road sign is only part of the task. The vehicle must understand how the surrounding scene may evolve and how its own decisions will affect the outcome.
A world model could help predict whether another vehicle is likely to change lanes, whether a cyclist may enter the road or how traffic will respond to sudden braking. Similar capabilities could apply to drones, delivery vehicles and industrial machinery.
This does not eliminate the need for conventional engineering. Safety-critical systems require redundancy, validation and clear operating limits. An AI model cannot be treated as a substitute for the entire control architecture.
The investment opportunity may therefore favour companies that integrate world models into established systems rather than those promising a single universal intelligence. Automotive suppliers, industrial automation groups and defence contractors already understand certification, hardware integration and operational risk. Their advantage may lie in turning experimental AI into products that can be deployed under demanding conditions.
At the same time, incumbents face the risk that new software companies capture the most valuable layer of the system. A vehicle or robot manufacturer may provide the hardware while an external platform controls the intelligence, data and ongoing software revenues.
The balance of power will depend on who owns the training data, interfaces and client relationships.
World Models Will Increase the Value of Simulation
Simulation software has traditionally been used to design products, test engineering choices and train human operators. World models could expand its role from a design tool into a core AI-development environment.
A manufacturer might build a digital version of a factory and allow AI agents to test production sequences before changing real operations. A logistics company could simulate different warehouse layouts and train robots to handle congestion. A city could model traffic, infrastructure failures or emergency responses before deploying autonomous systems.
This creates opportunities for companies already strong in computer-aided design, industrial software, digital twins and engineering simulation. Their advantage is not merely the ability to generate visually convincing environments. They understand the physical rules, constraints and technical data required to make a simulation useful.
Accuracy will matter more than graphical realism. An environment designed for entertainment can tolerate simplified physics if the experience remains engaging. A model training a robot or aircraft system needs much closer alignment with real operating conditions.
The strongest platforms may combine synthetic environments with real-world feedback. A system trains in simulation, operates in the physical world, records errors and uses that experience to improve the virtual model. This creates a data loop that becomes more valuable over time.
Companies controlling such loops could develop durable competitive advantages. The software improves because it is used, while clients become increasingly dependent on the platform’s accumulated data and integrations.
Data Rights Could Become a Strategic Asset
The early AI investment cycle rewarded businesses with access to large text and image datasets. The next cycle may place a premium on action data: records showing how people or machines behave in response to changing environments.
Videogame publishers, simulation platforms, robotics companies and autonomous-system developers may all possess valuable forms of this data. The challenge is determining whether it can be used legally and commercially.
Player behaviour can reveal personal information and may be covered by privacy rules or platform agreements. Game assets and environments are protected intellectual property. Developers training models across multiple titles may face questions over whether licences permit the extraction and reuse of interaction data.
Synthetic data may reduce some of these constraints. Developers can generate new environments and action sequences without relying entirely on recorded users. However, synthetic data is valuable only when it reflects reality closely enough to improve performance outside the simulation.
Companies capable of combining legally secure data, realistic environments and measurable physical-world results may become particularly attractive. Their data advantage would be difficult to reproduce through computing expenditure alone.
Investors should be cautious about companies that describe any large collection of gaming data as a defensible asset. Quantity matters, but relevance, rights and transferability matter more.
The Public-Market Exposure Is Still Indirect
Many of the companies working most directly on world models remain privately held. This makes the theme difficult to access through a simple listed-equity strategy.
Public-market investors may instead gain exposure through enabling technologies. Semiconductor companies provide processors for model training and real-time inference. Cloud platforms supply computing capacity and development tools. Industrial-software groups build simulation environments, while automation and robotics companies bring the systems into factories and warehouses.
Game-engine and videogame companies may offer more direct thematic exposure, though investors need to distinguish between participation in the idea and material earnings potential. A company can be frequently mentioned in connection with world models without generating significant revenue from them.
The same caution applies to established technology groups. Their size gives them the resources to invest in new AI architectures, but world models may represent only a small part of a much broader business. Success may not meaningfully change the investment case for years.
Private-market valuations present another risk. Capital is already moving towards companies promising the next AI breakthrough, often before technical feasibility or commercial demand is established. Large funding rounds can create the impression of validation when they may primarily reflect investor competition for scarce exposure.
World models could become an important computing architecture and still produce disappointing returns for investors who enter at excessive valuations.
The Winners May Differ From the First AI Boom
The first generative-AI cycle concentrated attention on foundation-model developers, cloud platforms and advanced chips. Those businesses will remain important, but physical-world AI introduces a different set of requirements.
It needs simulation, spatial data, real-time decision-making, sensors and hardware integration. It must operate under physical constraints and, in many cases, safety regulation. Distribution will depend on factories, vehicles, warehouses and machines rather than consumer access to a chatbot.
This may broaden the group of companies able to capture value. Industrial incumbents with deep domain knowledge could become more relevant, while software developers that lack access to physical systems may struggle to commercialise impressive models.
Gaming’s role is therefore best understood as an enabling one. Videogames offer environments in which AI can begin learning relationships that language alone cannot provide. They may supply data, simulation methods and development talent for a much larger market.
The investment thesis should not rest on the assumption that the next AI giant will emerge from the videogame industry. A more plausible conclusion is that technologies refined in games will move into robotics, transport and industrial systems, creating opportunities across several layers of the computing economy.
A Promising Theme With Unusually High Execution Risk
World models address a genuine limitation in current AI systems. Machines that can write convincingly are not automatically capable of navigating a room, handling an unfamiliar object or adapting safely when a physical environment changes.
Games and simulations provide a practical way to generate experience at a scale that would be impossible in the real world. The approach has a clear logic and significant commercial potential.
It also carries considerable uncertainty. Models trained in virtual environments may fail to transfer reliably to physical settings. Development costs could remain high, regulation may slow deployment and many applications will require years of integration before they generate meaningful revenue.
Investors should therefore separate the technological thesis from the individual company case. The field may advance rapidly without today’s most highly valued start-ups becoming its eventual winners. Enabling platforms may capture more value than model developers, while established industrial groups could prove better positioned than venture-backed specialists to deploy the technology at scale.
The current AI cycle taught machines to work with language. The next one may teach them how actions change the world around them. Videogames are unlikely to be the final destination, but they may provide one of the most useful places to begin.
