Co-authored by Antonio Chidiac, MetaProp Investor.

By now, many of us have heard of world models. Every VC, tech founder, and AI Czar has come across something in this realm over the past 6 months. You might have heard about Yann LeCun leaving Meta to start his new world model AI lab or Fei-Fei Li’s work in building World Labs. But, have you ever thought to ask what these things are anyway? Or, where and how do we actually apply them to improve systems in our world?

Below we explore how the next frontier in AI, spatial intelligence models, could help resolve some decades-old challenges that have deeply affected the built environment and its constituents. The use cases are exciting and the early signs are encouraging but there are some real risks. The sooner founders and ecosystem leaders address them, the smoother the road to beneficial deployment. Nevertheless, the prospects for built infrastructure (especially in the climate space) are promising. Let’s take a closer look, shall we?

What are AI world models? And, why should we care?

In its simplest form, an AI world model is an advanced artificial intelligence system that designs simulated representations of the real world to predict the next state of a given environment using a variety of learning mechanisms. These models go beyond passive pattern matching. When someone throws a tennis ball your way, you know the ball will hit you unless you react and catch it or duck. As humans, we do this subconsciously. AI world models internalize this knowledge in a way that enables predictive capabilities in physical world scenarios. So, here’s how a world model builds this complex capability:

• Similarly to other Generative AI like LLMs (large language models), AI world model algorithms are trained using supervised, unsupervised, and reinforcement learning and are fine tuned for improved accuracy across iterations.
• However, whereas LLMs are built to predict the next token in a sequence of words, world models predict the next state of a physical environment or event.
• To do this, these models need to understand a given state’s physics, geometry, cause and effect relationship structure, and how agents interact with it and each other (Rohit Bandaru, 2025).
• While we build common sense through lived experiences, world models develop these common sense-like skills by ingesting complex, unstructured patterns of data via different approaches such as real world simulations, latent state representations (i.e. abstract reflections of thoughts or scenarios), and continuous perception-prediction-action-reflection cycles (Tredence, 2026).

Three key properties distinguish world models from other generative models like AI video production or static 3D reconstruction:

• Generativity. A world model generates simulated world states (frames, scenes, geometry, sensor streams) rather than retrieves them from existing sources.
• Action conditioning. It accepts actions/causes as a first-class input — an agent’s command, a camera move, a robot trajectory, and so on — and the simulated world predicts and responds to subsequent effects accordingly (Google DeepMind, 2025).
• Persistence and consistency. It must retain memory and coherence across time, so that what was on the left of the scene one minute ago is still on the left when the agent looks back. This is the hardest property to achieve and where most current systems still fail because models tend to be stateless and probabilistic, memory modules must be built to resolve this feature (Arxiv, 2026).

World models are tremendously promising precisely because they unlock a whole new realm — the physical world — for machine learning and connected intelligence to tackle and unshackle. Essentially, they are anticipated to unleash for real world problem solving what LLMs unshackled for language-based reasoning. The excitement around them is only growing. For traditional industries, physical asset holders, and heavy industry, the prospects and ramifications could well be gargantuan. Only time will tell.

Where are we in the AI world model development cycle?

World models are still in their infancy — at least in relation to where we need them to be — but the performance is improving, early traction is taking shape, and the wider direction is cementing itself. As is often the case in venture though, capital is flooding the ecosystem faster than commercial use cases are maturing. You’ve seen this story before… first, the labs experiment and yield a major breakthrough, then a hype craze takes the ecosystem by storm, then the market crashes only to see a few resilient ventures survive before, finally, the real longer term winners emerge. At this stage, three signals indicate that we are no longer in the science project phase and the hype cycle is well under way:

1. Labs have shipped the foundations. Google DeepMind released Genie 3 in August 2025. It is the first general-purpose world model that generates navigable 3D environments from a text prompt at 24 frames per second, maintaining coherence for a few minutes at a time (Google DeepMind, 2025). NVIDIA’s Cosmos ecosystem has now been downloaded over 2 million times and is being used by Skild AI, Figure, Lightwheel, and Uber to generate physics-aware synthetic training data (NVIDIA Newsroom, 2026). Wayve, the London-based AV tech company, has open-sourced GAIA-2, a multi-camera generative world model purpose-built for autonomous driving (Wayve, 2025). In 2024, Stanford Institute for Human-Centered Artificial Intelligence’s Fei-Fei Li launched World Labs, a frontier spatial intelligence lab backed by a16z, Radical & NVIDIA (Reuters, 2024).

2. Capital is in flow state. In 2026, World Labs launched its flagship commercial world model, Marble, which generates persistent, editable 3D environments exportable as real-time renderings or 3D wireframes (TechCrunch, 2025). Just three months later, the company raised $1 billion in fresh funding from AMD, Autodesk, NVIDIA, Fidelity, and Emerson Collective (The AI Insider, 2026). Around the same time, Yann LeCun, often credited as one of AI’s academic pioneers, left Meta after 12 years to launch AMI Labs, a Paris-based startup focused exclusively on world models — and raised over $1 billion at a $3.5 billion pre-money valuation before launching a single product (WIRED, 2026).

3. Use cases are evermore present. Bedrock Robotics — founded by ex-Waymo leaders and backed by CapitalG, NVentures, Tishman Speyer, and MIT, is using world model-trained perception to retrofit existing heavy equipment with autonomous capabilities. TerraFirma — a venture-backed startup built by former SpaceX engineers — is building a continuously improving general contractor of the future, embedding elementary world models into its plug-and-play construction equipment technologies. They have already operated revenue-generating and profitable excavation projects for Starbucks, Zachary Cole Foundation, and other institutional developers. Causal Dynamics Lab — founded by ex-Microsoft, Uber, and VMware engineers — is a niche AI lab focused on a unique application of world models. Its platform uses virtual spatial intelligence to predict how proposed code amendments will behave in the production environment before such changes are shipped. By building maps of the code base, Causal’s flagship model outperforms Anthropic & OpenAI in code localization.

Overall, however, the technology is still demonstrably immature. Genie 3 only sustains coherent worlds for a few minutes; long-horizon reasoning remains an open research problem (Google DeepMind, 2025). Recent academic work shows that current world models routinely produce “physical hallucinations” — objects passing through walls, fluids behaving non-physically, glass shattering before impact (Cornell University, 2025). In essence, we are still likely a few years away from production-grade reliability. That said, beyond the foundation labs, those poised to win in this space tomorrow are those best positioned today. And, those are the ventures capturing, orchestrating, and applying real-world data in the most ingeniously useful and productive ways.

What the built environment stands to benefit from world model innovation

The built environment has overseen digitization via a sequential deployment over the past 25 years: BIM (geometric information), GIS (geographic information), IoT (sensor data), digital twins (real-time remote operations and site auditing), and most recently agentic AI (workflow automation) — the latter of which is still in its nascent stages with significant adoption and productivity gains cemented only by a few. The problem with sequential implementation is that each layer adds new data points without much contextual understanding.

For instance, a digital twin can tell you that an industrial kitchen’s commercial refrigerator is running at 75% capacity. It cannot tell you what happens to food and beverage quality, staff comfort, and electricity cost divergence if you reduce capacity to 50% over the two hours as a heat wave is ramping up and a facility manager is running a team meeting in the warehouse next door. That kind of question requires simulating the industrial facility, the people, the goods stored and equipment used, the electricity grid, and the weather as a single, multi-pronged dynamic system.

Well, guess what AI world models are precisely designed to do.

We see four built-world capabilities being unlocked as the technology matures:

• Synthetic cause-effect interaction data in droves. The biggest bottleneck in physical AI for built environments has been the cost of collecting real-world action-outcome data: What happens when an excavator engages this soil? What happens when a sprinkler activates in this configuration? What happens when 10,000 people evacuate a stadium? World models break that bottleneck.
• Multi-variate operations scenario analysis. World models could potentially run thousands of “what if” scenarios at near-zero marginal cost: What if a pipe bursts on floor 14? What if a tenant adds a 20 pax team next quarter? This enables a transition from static monitoring to dynamic scenario optimization.
• Long-horizon development planning. Construction sequencing, capital deployment, retrofit prioritization, and tenant fit-outs all involve long-term decisions with physical consequences. World models and the planning algorithms sitting on top are the natural compute substrate for such decisions.
• Future-ready risk evaluation. Catastrophe planning models built on historical data are deceptive tools for a non-stationary climate. World models, applied at the scale of an asset or a portfolio, allow asset-specific, danger-specific, and situation-specific simulation of how a particular structure or environment could respond to a given event. Such a foundation is poised to play a critical role in the establishment or evolution of next generation insurance underwriting, credit assessment, and capital-stack pricing venture products.

What we at MetaProp believe

Frontier world models are being developed by well-funded labs like NVIDIA, Google DeepMind, Meta FAIR, World Labs, and Wayve. However, these aforementioned players may not be best placed to build the vertical orchestration and real-world application layers for construction, building operations, energy system coordination, or real asset management. This is precisely where venture-grade opportunities lie and where MetaProp possesses distribution advantages that could help founders fly.

Leaning on our deeply entrenched knowledge, operational capabilities, and strategic LP base — representing over 20 billion square feet in managed RE assets — our right to play in this realm sits where world models help redefine the end-to-end construction, or how buildings are monitored and managed remotely, or how capital markets allocate finances based on risk-reward models, and so on. With many more on the horizon, we identified three fitting and intriguing areas to explore as the ecosystem develops:

• Spatial AI middleware and design-to-fabrication automation in buildings
• Infrastructure-level energy system coordination & decarbonization
• Physical risk underwriting and real estate climate adaptation

1. Spatial AI middleware & design-to-fabrication automation in buildings

What to do about buildings? How about starting with construction, often considered one of the most laggard and economically inefficient sectors in the global economy. While manufacturing, finance, and the overall economy have benefited tremendously from innovation over the past 50 years, construction saw Total Factor Productivity (value add) decline by over 40% over the same period (The University of Chicago: Becker Friedman Institute, 2023). Several reasons have contributed to this perplexing trend among which are increasing regulatory burdens, low/slow technology adoption, increased project complexity in modern buildings, reduction in project standardization, and shifts towards smaller scale projects for mid-sized contractors. The combination of the above makes economies of scale harder to make use of and productivity benefits more elusive.

Spatial intelligence brings about some exciting capabilities. They make it financially viable to (a) maintain a live, physics-aware model of a project, (b) generate synthetic training data for semi-autonomous equipment across complex site conditions, and (c) reason about consequences of construction sequence decisions before execution. Some of these applications are already seeing the light of day, with pilot testing and early adoption in motion with tech-enabled contractors and on-site tech providers. Some of these early-stage ventures have had the foresight to see that data capture and orchestration is critical to spatial intel building and deployment in the future.

Another area of interest in buildings lies in commercial real estate. That is investing in ventures developing the middleware where the spatial-AI layer converts a particular multi-modal real world simulation data into usable, action-conditioned input for an interactive workflow. Just imagine a coordination function serving as the connected tissue for voice AI to serve tenants and visitors. Commercial and residential building infrastructure and how agents interact with them stand to be revamped entirely. Imagine where today’s leading voice labs and vertical orchestration pioneers land once safe, auditable, and reliable spatial intelligence models are embedded into their AI stacks.

Early birds scratching the surface:

2. Infrastructure-level energy system coordination & decarbonization

Buildings are simultaneously the largest energy demand source, the largest near-term decarbonization opportunity, and the most underused source of grid flexibility. Buildings account for roughly 30% of global final energy consumption and around 26% of energy-related emissions (IEA, 2024). With deep strains on the grid due to rapidly growing demand and aging infrastructure, the consequences on energy systems downstream are compounded further. In fact, we put together a thought leadership on this back in March, arguing that now might be the time for venture capital to invest in solutions to these long-standing and now exacerbated challenges.

World models are unusually well-suited to tackling these problems for a few reasons. First, the underlying physics (e.g. thermodynamics, electrical flow, computational fluid dynamics, building envelope behavior) has been comprehensively researched for decades and is well-understood. Second, the promise of the technology is that it could couple this know-how with weather patterns, market signals, occupant behavior, and equipment wear turning simulations into dynamic and dimensional operating tools.

We envisage the proliferation of spatial intelligence-enabled digital twin products and OS solutions that help manage and navigate systems not only at the physical asset level (e.g. data center rack, chiller, building) but also at the tangential environment-level (i.e. neighborhood, township, gated community, district). A few ventures are already in execution mode with today’s nascent technology. If they can build a moat around data capture for model development and integration or value realization via orchestration, their success odds would improve dramatically.

Early birds scratching the surface:

3. Physical risk underwriting & real estate climate adaptation

At this point, anyone remotely connected to the real estate ecosystem via financial services, technology, venture, municipal politics, or civic engagement knows about the insurance exodus from climate risk-prone regions. Where insurers haven’t totally dropped policy provisions, real estate has been repriced at practically every layer of the capital stack. Simultaneously, credit providers are quietly tightening LTV ratios on exposed assets. In some cases, they too have declined to underwrite new mortgages.

The main culprit — besides human-induced climate change itself — is the antiquated underwriting models themselves, particularly catastrophe-modeling which is largely reliant on historical loss data (SwissRe, 2021). Such a hindsight approach cannot keep up with rapidly evolving, non-linear climate-related consequences. Another major limitation of traditional models is that they fail to account for the non-stationary environment and long-tail events which are heavily influenced by warming temperatures (ScienceDirect, 2024). Guess what directly addresses these two constraints? That’s right, world models do exactly that — in theory, at least.

World models, applied at the scale of an asset or a portfolio, enable forward-looking, physics-rooted simulations of how a specific building responds to a given peril such as a flood or wildfire. Such applications of a spatial intelligence layer could help in two ways that are deeply relevant to the build environment: (a) dramatically enhance underwriting models to bring back and better price policies in affected areas; and, (b) incentivize climate resilience and adaptation via retrofits, significantly helping avoid otherwise unaccounted for future costs. In our view, these use cases serve as the cleanest and clearest fit between world models and our fund-level thesis. From there, value propositions only get stronger as vertical or horizontal knowledge is added to the spatial intelligence stack and is extrapolated to navigate even more complex situations.

Early birds scratching the surface:

One eye on risks and limitations

Overall, we’re long on world models as an applied AI spatial intelligence navigation layer for built world use cases. That said, we’re also clear-eyed about the fact that world models are maturing and their trajectory isn’t exactly a fixed certainty. A serious investment thesis should first acknowledge and, later, deeply understand limitations. There are two major types of risks to highlight: a) General development and regulatory risks and, b) Product-specific implementation and adoption risks. The list of general risks is quite cumbersome, so here is our attempt at surmising the key points below:

• Physical hallucination. As mentioned in the development cycle section, current models produce stochastic violations of basic physics. This “realism gap” between photorealistic output and physically reliable simulation is a major risk to proliferation (Cornell University, 2025).
• Long-term memory coherence. Also mentioned earlier is Genie 3’s ability to maintain memory for limited time periods. Ultimately, built-world applications need week, year, or decade-long horizons. Hierarchical planning and consistent memory remain open research problems for now (Google DeepMind, 2025).
• Hefty compute economics. Spatial and geophysics-based simulations are substantially more compute-intensive per inference than text generation. While Moore’s Law may resolve this, short-term deployment where margins are thin and per-asset value is moderate will face challenging hurdles or outright barriers.
• Training data dilemma. Foundation models for built environments need domain-specific video from construction sites, building interiors, and more. Most of said material is owned by contractors, owners, or vendors. The scarce and proprietary nature of such data makes licensing increasingly litigious and foundational data hard to acquiesce.
• Base capabilities absorbed. NVIDIA Cosmos is open-licensed, Marble has an API, and Genie will eventually open. The marginal cost of base capability is trending toward commodity and the foundation labs have a clear advantage here. Defensibility for the rest of the market sits in areas such as vertical data capture and interpretation, customer integration, and workflow ownership.
• The interpretability issue. As is the case with LLMs and neural networks, many world models operate as “black boxes,” meaning interpreting how they function is often not totally understood even if they bring about intended results. For critical sectors, this may not be enough, particularly when implications are significant or potentially catastrophic.
• Cybersecurity & misuse risks. Amplified by the sheer scale of their reach and how they could be embedded in robotics, smart building management, or electricity grids, the risk of cybersecurity exploitation for malicious usage should be taken very seriously before widespread adoption.
• Autonomy risks. While operating as agents, world models may run simulations and tasks autonomously. Without proper guardrails and failsafe settings, humans may be too late to react or control system malfunctions and failures. In a high stakes setting, such scenarios could be unforgiving (International AI Safety Report, 2026).

Regulatory scrutiny is likely to eventually oversee some AI applications and the aforementioned risks suggest world models won’t be spared, perhaps reasonably so. World models used in safety-critical contexts (e.g. power distribution, autonomous equipment, life safety, insurance pricing, etc.) will face a similar regulatory development arc as autonomous vehicles. Therefore, ventures that have the foresight to bake in interpretability, validation, and safety early on stand to best navigate such an environment and take the market by storm.

While of second order importance in magnitude, understanding the friction points specific in deploying venture-backed products across our three primary target investment areas is perhaps more critical as it is more controllable and influenceable. As MetaProp, this is where we have the ability to carve out a distribution and go-to-market path and a support function for portfolio companies. Below are a few we identified:

• Reality & spatial intel is sensitive and proprietary. Clients will be skeptical of who gets to ingest, own, and profit off of site data. Founders need credible answers on data governance from the very onset.
• Sales cycles in real estate and infrastructure development are lengthy and tedious. Even with proven ROI or payback period guarantees, operators adopt new technology slowly. MetaProp’s distribution clout can certainly help but is unlikely to overhaul procurement in the near future.
• Downstream distribution hurdles. Venture models powered by spatial intel may plug into industrial incumbent OEM equipment built by Caterpillar, Komatsu, Trane, Carrier, and more. Integrating physical-AI cross-over capabilities will be challenging at scale unless challengers purchase/lease equipment as inventory or acquire OEMs upfront.
• Data and licensing regulation varies across jurisdictions. Future-oriented models first gain ground in California, Texas, and in EU hubs, each of which has their own regulatory peculiarities. Founders need to embed this in their GTM roadmaps.
• Liabilities for model misuse should be taken seriously. Inaccurate spatial models that adversely affect outcomes in an underwriting model or operational decision are litigiously exposed. A well-composed user policy incl. historical validation, six sigma+ confidence interval limits, AI model cards, and stress tests are necessary.

Venture increasingly on the prowl

The world-model investor landscape has bifurcated. At one end, frontier-model investors are writing very large checks into a small number of foundation labs — a16z, Sequoia, Bessemer, Lux, Khosla, Founders Fund, and co. as well as major sovereigns are all engaged at the foundation layer (Bessemer, 2026). At the other end, application-layer investors are deploying more typical checks into vertical use cases. That is where MetaProp sits. Sooner rather than later, we envisage strategics, CVCs, and sector specialists, such as Camber Creek, Brick & Mortar, Energy Impact Partners, and JLL building conviction within the rapidly developing spatial intelligence realm.

There is a competitive dynamic worth flagging: built environment and infrastructure venture funds do not yet have a coordinated world model thesis. And, sure, one can argue that the horizontal AI investment firms backing the foundational labs do. However, what these funds don’t have is access to the LP base and distribution network that enable use case application, early adoption, and real world deployment at scale, particularly within a traditionally laggard and high-friction user base. Serving as trusted advisors to our LPs, we can help them pilot the right technologies and ensure use cases realize important value propositions on the ground. This is our right to play.

Closing argument

Conceptually, world models have been brought to life by classic novels and, later, by the big screen — some might recall HAL 9000’s spatial navigation capabilities in 2001: A Space Odyssey or autonomous vehicle navigation simulations in Total Recall or F.R.I.D.A.Y’s spatial analysis and heroic functions in the MCU’s Iron Man series. There’s so much to be excited about with spatial intelligence and the world models founders will be able to leverage to build exciting, useful, and beneficial products. Building smart cities coherently, creating affordable and well connected transportation systems while reducing social and environmental trade-offs, and navigating climate risks such as wildfires and flooding could well become within reach.

Nevertheless, the risks are real and may have more serious ramifications than LLMs. Sci-fi media is far from short of downside risk examples. This is precisely why coming in early is of utmost importance: to help dictate the direction of this fledgling ecosystem by ensuring protocols and safety standards are built in the foundational models and designing applied use cases that understand the contextual challenges and demands that the built world itself and its stakeholders have. We’ll back the spatial intel ventures that make the built environment and its core infrastructure more efficient, more adaptable, interpretable, and safer. Founders focused on this stand to win big in the process.