World models, the AI systems designed to simulate physical reality for autonomous vehicles and robotics, have attracted over $2 billion in funding in recent months, yet researchers have discovered fundamental mathematical boundaries that limit how well these systems work in practice. The contradiction reveals a widening gap between investor enthusiasm and the technical reality of building machines that can accurately predict the world over long periods of time.

What Exactly Is a World Model, and Why Does It Matter?

A world model is an artificial intelligence system trained to build an internal representation of how the physical world works, the way your brain maintains an intuitive understanding of physics, gravity, and cause-and-effect relationships. Instead of just predicting the next frame of a video or the next word in a sentence, a world model learns to simulate entire scenarios: a car swerving around an obstacle, a robot arm reaching for an object, a construction zone with contradictory lane markings. For autonomous vehicles and robotics companies, world models promise a way to generate millions of edge-case training scenarios that would be impossible, dangerous, or prohibitively expensive to capture in the real world.

The category became serious money almost overnight. World Labs raised $1 billion in February at a reported valuation near $5 billion, and weeks later, AMI Labs raised $1.03 billion in what was described as the largest seed round in European history, at a $3.5 billion pre-money valuation. Neither company had mature products generating revenue. Both were funded on the conviction that world models represent the next frontier of artificial intelligence after large language models.

Why Is the Math Suddenly a Problem?

On May 25, three researchers published a proof that drew a hard boundary around how most world models actually work. The researchers, including Yann LeCun, the executive chairman of AMI Labs, demonstrated that the leading statistical approach to building world models recovers the true structure of reality only under specific conditions: when the variables being modeled follow a bell curve distribution and drift in a particular, gentle way. Most physical systems of interest violate both conditions. A child darting into traffic, a truck jackknifing across lanes, a sudden weather shift, a sensor malfunction, these are the edge cases that matter most for safety-critical applications, and they are precisely the scenarios that break the mathematical assumptions underlying the dominant approach.

The timing created an unusual situation: the same person whose startup raised a billion dollars on world models put his name on the proof that identifies a fundamental limitation in how most of them work. The capital says the future is here. The math says the future, in its current form, has a boundary the capital has not priced.

How Do World Models Actually Fail in Practice?

A benchmark posted on May 20 revealed the practical consequences of these mathematical limits. One leading world model planned correctly about half the time in clean, controlled conditions, then dropped to approximately 12 percent accuracy when the agent changed color and to about 6 percent when the background changed. This brittleness matters enormously because a model that forecasts the next frame beautifully can still steer an autonomous vehicle into a wall ten steps later. The problem is not just theoretical; it is a concrete enterprise risk that appears in plain sight once you know where to look.

The field has conflated two different things: demo fidelity and planning reliability. A world model that generates photorealistic images of the next frame looks impressive in a presentation. A world model that maintains accuracy over long rollouts under distribution shift, the kind of accuracy you would need to trust it with a vehicle or a robot, is something else entirely. Investors and executives have been scoring the wrong metric.

What Are the Different Types of World Models, and Where Does Each One Break?

The term "world model" now covers three fundamentally different kinds of systems, each with different failure modes and different applications:

• Synthetic-data generators: Systems like NVIDIA Cosmos that produce training data for robotics and autonomous vehicles. These are production-grade today and work well for generating diverse scenarios at scale, but they are not designed to reason about long-horizon planning or physical constraints.
• Generative 3D systems: Models like Marble and Genie 3 that create photorealistic content and simulation environments. These excel at visual fidelity and are ready for content creation and game development, but they do not necessarily maintain physical consistency over extended rollouts.
• Long-horizon reasoning systems: Models trained to predict and plan over many steps into the future while maintaining physical accuracy. This is the category that has attracted the most capital and the most hype, and it is also the category where the math has drawn the hardest boundary. A model you trust to reason about your physical plant and act on it over long horizons is not yet a thing you can buy at production grade.

How Should Companies Actually Evaluate World Models?

The gap between hype and reality creates a practical dilemma for companies considering world models for their operations. The benchmarks and demos that vendors showcase are not reliable predictors of whether the system will work in your environment, under your conditions, with your sensors and your edge cases.

Before committing to a world model vendor, companies should demand specific information and testing protocols:

• Long-horizon consistency metrics: Ask any vendor for the accuracy maintained over extended rollouts, measured under the kind of distribution shift your actual environment will throw at it. Do not accept next-frame accuracy as a proxy for planning accuracy.
• Retraining costs: Statistical world models generally have to be retrained to absorb a new constraint or a new physical law. Ask a vendor what it costs, in time and money, to teach their model something it did not learn during initial training. This is a hidden line in the total cost of ownership.
• Testing under your own distribution: Run a pilot under your conditions before you sign off on a rollout. A color shift, a lighting change, a sensor your supplier did not train on, any of these can move a model out of the regime where it works. Brittleness is an enterprise risk in plain sight.

The smart money is about to start asking how far a model stays accurate before it drifts. Founders and vendors who get caught reporting next-frame accuracy as if it were planning accuracy will lose the technical diligence with investors who actually understand the field.

Where Is the Real Market Opportunity?

The general-purpose race is already crowded and well capitalized. AMI Labs, World Labs, Google DeepMind, and NVIDIA own the general-purpose lane and have the capital to keep it. Waymo unveiled its own generative simulation architecture in February, built on top of Google DeepMind's Genie 3, and positioned it as a core training tool for the Waymo Driver. Tesla, Nvidia, and Wayve are building comparable systems in-house.

The opening for smaller companies is domain-specific reliability: the place where a general model drifts, whereas a focused one need not. Decart, an Israeli startup, launched Oasis 3 on June 10, a real-time world model that renders photorealistic driving environments on demand via public API, priced at $0.02 per second. The company is betting it can own the simulation layer for physical AI before the big players build one themselves. Decart's real market is everyone below the top tier: the dozens of autonomous vehicle programs, robotics labs, and drone startups that want frontier-grade simulation without a DeepMind-sized research budget attached.

"Decart wants to be the infrastructure layer for physical AI the way OpenAI became the infrastructure layer for language," according to the company's framing of the Oasis 3 launch.

Decart CEO Dean Leitersdorf, as reported in Source 2

Oasis 3 generates three synchronized camera feeds, one front-facing and two side-facing, which covers the perception setup of most camera-first autonomous vehicle stacks but stops short of the lidar output that Waymo's system produces. The $0.02 per second rate is the public on-ramp. Enterprise pricing scales with the use case, and that is where the real revenue will sit if the strategy works.

What Should Investors and Executives Know Right Now?

The capital is ahead of the math, and that gap is your timing. A billion dollars moved on a thesis the same season a proof drew a line under part of it. The companies that win this race will be the ones whose technical story survives the proof and the ones that can demonstrate long-horizon consistency under perturbation, the way you would stress-test a credit book.

The term "world model" is about to lose its meaning. On the record, AMI's own CEO, Alexandre LeBrun, told TechCrunch that within six months every company will call itself a world model to raise funding. He runs one of the two best-funded examples in the world, and he is warning investors. Diligence the architecture under the label, because the label will be worthless by the fall.

The next twelve months will show whether renting simulation becomes the default for physical AI or stays a niche for teams that cannot build their own. If mid-tier autonomous vehicle companies and robotics startups start shipping products trained substantially on Oasis-generated environments or similar systems, then someone owns a layer of the stack nobody else is selling at retail. If the big players' internal models keep pulling ahead, two cents a second buys impressive demos and not much else. The real signal will be in the customer announcements, not the benchmarks.