Signals Inbox·September 2, 2026·World Models

World models: who are the top startups today?

World Labs leads the world-model startup race today, with Decart as the strongest immediate challenger and General Intuition rising fastest, while Odyssey, Wayve and Runway form the next group worth watching closely.

Q1What actually counts as a world-model startup today?

A world-model startup today needs to build AI that can represent an environment and predict what happens when something acts inside it.

That definition immediately removes a lot of companies that have started using the term loosely. Generating a realistic-looking video is useful, but a video model can get away with producing something plausible. A useful world model has to deal with consequences. Turn the steering wheel and the road should move correctly. Push an object and the object should respond in a physically sensible way. Walk around a generated room and the room should remain recognizably the same place.

The leading startups approach that problem differently. World Labs' Marble creates persistent, navigable 3D environments from text, images and video. Decart's Oasis 3 continuously generates environments that react to physical control inputs. Odyssey produces interactive simulations frame by frame. General Intuition uses world models as training environments for models that learn to act. Wayve's GAIA-3 generates driving scenarios so autonomous-driving systems can be evaluated in situations that are difficult or dangerous to reproduce on roads.

We therefore consider World Labs, Decart, Odyssey, General Intuition, AMI Labs and Veeda AI the clearest independent world-model startups right now. Runway belongs in the conversation too, although it already has a large generative-video business. Wayve and Waabi are more accurately described as autonomous-driving companies that happen to have built unusually serious world-model technology.

Q2Why are investors suddenly throwing billions at world-model startups?

Investors are pouring money into world models because physical AI still has a training problem that language models never had to solve: robots cannot cheaply make billions of mistakes in the real world.

An LLM can process enormous amounts of text without breaking anything. A warehouse robot, autonomous car or humanoid learning through trial and error can damage equipment, interrupt operations or hurt someone. Simulation offers a way to compress huge amounts of experience into a much safer training environment.

The capital flowing into the field shows how seriously investors now take that idea. During 2026, World Labs announced $1 billion in new funding, AMI Labs raised $1.03 billion, General Intuition announced a $320 million Series A, Odyssey raised $310 million, Decart raised $300 million and the newly launched Veeda AI raised more than $90 million.

Those six relatively pure world-model companies alone account for roughly $3.05 billion of major rounds announced this year. Add Wayve's $1.2 billion Series D, Waabi's $750 million Series C and Runway's $315 million Series E, and the total crosses $5.3 billion before counting milestone-based future commitments.

The concentration is even more striking. World Labs and AMI Labs account for about two-thirds of the $3.05 billion raised by those six pure-play companies. Investors are already financing the strongest teams more like frontier AI labs than normal software startups.

Major world-model startup funding rounds announced in 2026

Q3Can we actually tell which world model is technically best?

There is no credible technical winner in world models today because the leading companies are solving different problems and still lack a common benchmark that captures what we actually care about.

The trade-offs become obvious when we compare their systems. World Labs optimizes for spatial persistence: Marble produces a world that can be explored and exported as 3D assets. Odyssey focuses heavily on interactive generation and has demonstrated models running at around 20 frames per second. Decart says Oasis 3 responds to physical controls with less than 200 milliseconds of end-to-end latency while keeping multiple camera views synchronized. General Intuition's MIRA generates a shared environment for four players at roughly 20 frames per second. Wayve's 15-billion-parameter GAIA-3 is built around controlled evaluation of driving systems.

A company can look exceptional on one of those dimensions and average on another. A perfectly persistent 3D room that takes several minutes to generate solves a different problem from a driving simulator that must react almost instantly. A photorealistic video simulator can still struggle to remember the geometry of a world after a long interaction.

We have also started seeing more honest evidence of those weaknesses. TechCrunch's testing of Decart's Oasis 3 found impressive visual realism and very long continuous simulations, while coherence could deteriorate as the session continued. Odyssey itself has built PROWL-1, an adversarial reinforcement-learning system specifically designed to search its world models for failures in geometry, motion and action following. The companies are effectively telling us that impressive demos still hide plenty of failure modes.

For now, we would judge the models across persistence, controllability, latency, physical consistency, multi-agent capability, developer access and transfer to real machines. Turning all of that into one score would create false precision. Too neat, basically.

Where leading world models look strongest today

Q4Has venture capital already picked the world-model winners?

Venture capital has created a clear first tier of world-model startups, but the money is still moving too quickly for funding alone to tell us who is winning.

The clearest example is General Intuition. The company announced a $320 million Series A at a $2.3 billion valuation in June 2026. Only weeks later, TechCrunch reported that Valor Equity Partners, Point72 Ventures, Seven Seven Six and existing backers were discussing another round at a $6 billion pre-money valuation. The reported financing was still being finalized, so we should treat $6 billion as a proposed valuation rather than a completed one.

Decart has seen a different form of validation. The company raised $300 million in May, taking disclosed funding above $450 million. More recently, Reuters and Bloomberg reported that Anthropic had discussed acquiring Decart for around $6 billion. The talks could still disappear, but a frontier AI company even considering a deal of that size says something about the strategic value of Decart's combination of real-time models and compute-efficiency software.

Odyssey reached a $1.45 billion valuation with its $310 million Series B. AMI Labs raised its first $1.03 billion at a $3.5 billion pre-money valuation. Wayve reached an $8.6 billion post-money valuation after raising $1.2 billion.

Investors are placing several different bets simultaneously: persistent 3D worlds, interactive video simulation, action models, efficient inference and domain-specific autonomy. Capital has narrowed the field. It has not settled the technical argument.

Q5Is World Labs still the world-model startup to beat?

World Labs remains our number-one pure world-model startup today because it has combined elite research talent, huge funding and an actual product better than any other independent company in the category.

Fei-Fei Li's company first emerged with $230 million and later announced another $1 billion round backed by investors including Nvidia, AMD, Fidelity and Autodesk. Autodesk alone invested $200 million, giving World Labs an unusually relevant strategic partner in industries that already depend on 3D environments.

The more convincing part is Marble. Users can create persistent worlds from text, images, panoramas, multiple views or video. World Labs also opened a public World API, so developers can generate those environments inside their own software. The API exposes Gaussian splats and collider meshes, which makes the output useful beyond simply viewing a generated scene.

World Labs has kept improving the product as well. Marble 1.1 and Marble 1.1 Plus expanded the available model lineup, with the Plus version designed to cover larger spaces. That gives us several layers of evidence at once: research, financing, a consumer-facing product, developer infrastructure and downloadable spatial assets.

The biggest gap is physical deployment. World Labs talks about robotics as a major future application, yet companies such as Wayve, Waabi and even General Intuition currently show more direct evidence of models interacting with real machines.

Still, if someone asks which independent startup currently looks most capable of turning world models into a broad platform, World Labs is the best answer we have.

Q6Is Decart now the most dangerous challenger to World Labs?

Decart is currently the strongest challenger to World Labs because its world models already run interactively, developers can access them, and its infrastructure business attacks one of the category's ugliest problems: inference cost.

Oasis began as a continuously generated Minecraft-style experiment. Oasis 3 is much more serious. Decart built it for physical AI, starting with autonomous driving, and says the system can generate synchronized multi-camera environments that respond to real control inputs with under 200 milliseconds of end-to-end feedback.

The company also owns DOS, its optimization stack for AI training and inference. That gives Decart an interesting position. World models need to generate new states continuously, often many times per second. If the category scales, computational efficiency becomes part of the product rather than an invisible backend concern.

Decart has lately been pushing that speed advantage across several products. Lucy 2.5, its live video model, runs at 30 frames per second, while the company has continued improving DOS and Oasis. Nvidia joined its latest financing as an investor, alongside a group that includes Benchmark and Radical Ventures.

The reported Anthropic acquisition talks add another clue. Reuters said Decart's team could join Anthropic's inference and performance organization if a deal happens. That suggests Decart's value goes well beyond a flashy world-model demo; its infrastructure may be just as strategically important.

We still see memory and long-session coherence as the technical issue to watch. If Decart can improve those while keeping its latency advantage, World Labs will have a much harder time holding the top spot.

Q7Is General Intuition becoming the hottest world-model startup right now?

General Intuition is probably the hottest world-model startup among investors these days, and its unusual advantage comes from training on what humans did rather than only what humans saw.

General Intuition was spun out of Medal, the gameplay-clipping platform. Medal gives the company access to billions of gameplay clips from around 17 million monthly active users. Crucially, those clips can include action labels showing which controls players used and when they used them.

That creates a richer dataset than ordinary internet video. A video can show a character jumping. General Intuition can potentially know that a player pressed a specific button, the character jumped and the environment changed in response. We get a much cleaner view of action and consequence.

The company demonstrated that approach with MIRA, a 5-billion-parameter multiplayer world model developed with Kyutai and Epic Games. MIRA was trained on 10,000 hours of four-player Rocket League gameplay and generates the perspectives of all four participants while reacting to their controls.

General Intuition has also started testing whether its learned behavior can leave games. TechCrunch visited the company and reported a quadruped robot navigating an office after a model was fine-tuned with only eight minutes of additional real-world robotics data. That is still an early demonstration, but it is exactly the experiment that could validate the company's entire thesis.

The market has reacted aggressively. As pointed out above, General Intuition's latest reported financing discussions would move its valuation from $2.3 billion to around $6 billion within a very short period.

We would rank General Intuition behind World Labs and Decart for now because external developers have much less access to its core technology. Technically, however, it may have the most interesting proprietary training asset in the field.

Q8Is Odyssey actually keeping up with World Labs and Decart?

Odyssey is keeping up better than its lower profile suggests, and its recent work shows one of the broadest attempts to turn world models into general interactive simulators.

Odyssey-2 generates environments interactively at roughly 20 frames per second. The company has also moved beyond single-user demos. Agora-1 allows multiple human or AI participants to act inside the same generated world at once, which opens a different set of uses in games, training and multi-agent AI.

Its PROWL-1 research is arguably even more interesting. Odyssey lets a reinforcement-learning agent deliberately explore environments looking for weaknesses in the world model. The agent receives rewards for finding failures such as broken geometry, bad motion or incorrect responses to actions. Those failures can then become new training material.

That creates a potentially useful loop: generate a world, let an agent attack the model, discover where the simulation breaks, then train on those failures.

Odyssey also has enough capital to keep pushing. Its $310 million Series B brought Amazon, AMD Ventures, GV, EQT and In-Q-Tel into the shareholder base at a $1.45 billion valuation. AWS became Odyssey's preferred cloud provider, and the companies plan to work together around Amazon's Trainium hardware.

The question around Odyssey is increasingly commercial. We can see rapid technical output and serious financial backing. We still do not know how many developers will pay enough for continuous world simulation to support the huge compute bill behind it.

Q9Does AMI Labs deserve to rank near the top before it ships a product?

AMI Labs belongs among the most important world-model research companies, but we would not put it near the top of the startup ranking until outsiders can actually test what Yann LeCun's team has built.

AMI has an exceptional starting position. Yann LeCun spent years arguing that next-token prediction would eventually hit fundamental limits and that intelligent systems need internal models of the world, memory, planning and the ability to reason over abstract representations.

Investors gave AMI more than $1 billion to pursue that idea. Its $1.03 billion financing came at a $3.5 billion pre-money valuation with backers including Nvidia, Temasek, Samsung and Bezos Expeditions.

Yet the gap between the financing and the public evidence remains huge. World Labs has Marble. Decart has Oasis. Odyssey has Odyssey-2, Agora-1 and PROWL-1. General Intuition has shown MIRA and early robot transfer. AMI's public story still revolves mostly around its research direction and the reputation of its founders.

That may change quickly. AMI arguably has more freedom than almost anyone on this list to pursue long-horizon research without worrying about immediate revenue.

Today, however, we rank observable progress higher than reputation. AMI could eventually become the most technically important company here. It has not shown enough yet for us to call it one of the current leaders.

Q10Is Runway really a world-model company now?

Runway is genuinely becoming a world-model company, and its existing video business gives it a commercial head start that the younger labs do not have.

The company now describes its broader mission as building AI to simulate the world. GWM-1 comes in versions for explorable worlds, virtual characters and robotics. The robotics work can generate possible future outcomes conditioned on robot actions, evaluate policies offline and create synthetic training data.

Runway is also building actual robotics tooling around those models. Its platform now includes policy inference, offline policy evaluation and data augmentation rather than stopping at a research paper or video demo.

What makes Runway unusual is the business underneath all of this. Millions of people already know the company for video generation, and Runway has relationships with entertainment and enterprise customers. TechCrunch reported in May that the company had added about $40 million of annual recurring revenue during the second quarter alone, according to one of its founders.

The company raised another $315 million earlier in 2026, with General Atlantic leading and Nvidia, Adobe Ventures, AMD Ventures and Fidelity participating. Runway explicitly said that the capital would fund the next generation of world models.

We would still classify Runway as a broader AI company rather than a pure world-model startup. That distinction may become less meaningful if GWM becomes central to its product line. Runway already has something the younger labs are still searching for: paying customers and distribution.

Q11Are Wayve and Waabi already ahead where world models actually matter?

Wayve and Waabi currently provide the strongest evidence that world models can help develop AI systems meant to operate expensive machines in the real world.

Wayve's GAIA-3 has 15 billion parameters and was trained on roughly ten times more data than GAIA-2. The model generates controlled driving situations that can be repeated, modified and used to evaluate autonomous-driving systems, including rare events that would be unsafe to recreate deliberately on public roads.

That world-model work sits inside a much larger commercial push. Wayve raised $1.2 billion at an $8.6 billion post-money valuation and brought Mercedes-Benz, Nissan and Stellantis into the investor group alongside Microsoft, Nvidia and Uber. The company says its AI has driven zero-shot across more than 500 cities in Europe, North America and Japan.

Waabi takes an even more simulation-heavy approach. Waabi World recreates situations from real sensor data, generates new scenarios and tests the Waabi Driver against them in closed loop. Waabi published a 99.7% simulator realism score based on paired simulation-versus-real-world tests. Since the methodology comes from Waabi itself, we should treat the exact percentage cautiously, but the attempt to quantify simulator fidelity is useful.

Waabi then raised a $750 million Series C and secured an additional milestone-based future commitment from Uber tied to robotaxi deployment.

Both companies have narrower goals than World Labs or Odyssey. That specialization has an advantage: we can judge whether their models actually improve a system with a concrete job to do. General-purpose world-model startups have a much harder proof ahead of them.

Q12Who has the strongest data moat in world models?

General Intuition currently has the most distinctive general-purpose data moat, while Wayve and Waabi have deeper proprietary loops inside autonomous driving.

Medal gives General Intuition an unusual stream of human behavior at enormous scale. The company says it can train on billions of clips tied to the actions players actually took. Collecting an equivalent dataset through robots would be painfully expensive, and normal internet video usually lacks those control labels.

Wayve has something harder to reproduce in driving. Its models learn from real vehicles operating across many geographies, vehicle types and traffic patterns. GAIA-3 covers data from nine countries across three continents, and Wayve's commercial relationships could keep increasing that diversity.

Waabi has a slightly different loop. Real-world sensor data feeds Waabi World, simulation generates additional situations, the Waabi Driver encounters them, and failures can feed back into development. That closed loop is highly valuable inside autonomous vehicles even if it has limited use outside transportation.

World Labs and Odyssey have invested heavily in spatial and video data, but we have not seen evidence of a similarly privileged proprietary stream. Decart's more obvious advantage sits in real-time infrastructure and inference efficiency.

The big unanswered question is whether General Intuition's gaming data transfers far enough. If it does, billions of action-labelled gameplay clips may be much more valuable than they first appear. If physical reality demands substantially different training data, Wayve and Waabi's narrower datasets could prove more useful.

Q13Which world-model business model looks strongest today?

The strongest world-model businesses currently attach simulation to an expensive existing workflow where better prediction can save real money.

Autonomous driving is the easiest example. Wayve and Waabi can justify enormous spending on simulation if it reduces road testing, catches failures earlier or speeds up deployment. One successful autonomy platform can also be worth far more per customer than a consumer world-generation subscription.

World Labs has a different path that already looks commercially sensible. Marble sells directly to creators, while the World API lets developers put generated 3D environments into games, design tools and simulations. Autodesk's investment gives World Labs a natural route into industries that already spend heavily creating 3D assets.

Runway has another advantage: it can introduce world-model technology to customers already paying for generative media. The company does not need the entire new category to become commercially mature before it can monetize parts of the technology.

Decart may have the cleverest hedge. It can sell Oasis for physical simulation, Lucy for live visual generation and DOS for the compute layer underneath them. Even if the eventual winner uses a different world model, companies will still care about making expensive models run faster.

Odyssey and General Intuition have exciting technology but less visible proof of what customers will pay for today. AMI and Veeda remain earlier.

This makes us skeptical of one giant horizontal “world-model API” swallowing the market soon. The first big businesses are more likely to appear inside robotics, autonomous vehicles, 3D creation, entertainment and specialized simulation.

Q14Who looks best positioned to win world models for robotics?

No startup has won world models for robotics yet, although General Intuition, Decart, Runway and the newly launched Veeda AI are now making very different bets on how robots should learn.

General Intuition wants to start with human behavior. Its models learn actions from games and then transfer that intuition into physical embodiments. The quadruped experiment reported by TechCrunch gives us an early example of that transfer working with very little additional robot data.

Decart starts from simulation. Oasis 3 creates interactive environments where autonomous systems can act, fail and encounter new scenarios continuously. The first major target is driving, but Decart has discussed extending the approach toward drones, off-road machines and other robotic systems.

Runway is building a more accessible toolkit for robotics developers. Its GWM Robotics work includes simulated rollouts, policy evaluation and synthetic-data generation. That could make Runway useful even to robotics companies that train their own control models.

Veeda AI is the newest and perhaps purest version of the simulation thesis. Sanja Fidler, formerly Nvidia's vice president of AI research, launched the company with Žan Gojčič and Huan Ling. Veeda raised more than $90 million shortly after formation and says its sole mission is to build simulated reality where physical AI systems can learn through repeated interaction.

Then we have Wayve and Waabi, which provide the reality check. Both already use simulation to improve autonomous machines in a specific domain. General robotics may eventually need a broad foundation model, but specialized systems are currently much further along in proving that the loop works.

We would therefore give General Intuition the most interesting upside, Decart the strongest real-time simulation position and Wayve the best current evidence of world-model ideas reaching real machines.

Q15So who are the top world-model startups today?

World Labs is still our number-one world-model startup today, with Decart and General Intuition close enough behind that the ranking could change quickly.

World Labs earns first place because we can already use its product, call its API, generate persistent 3D worlds and export useful spatial assets. It has also attracted enough capital and strategic support to keep competing at frontier-model scale.

Decart comes second. Its real-time performance is exceptional, Oasis has moved into physical-AI simulation, and DOS gives the company valuable technology underneath the models themselves. The recent Anthropic acquisition discussions make Decart look increasingly strategic across the broader AI stack.

General Intuition is third and currently rising fastest. Its action-labelled Medal dataset could become one of the category's strongest moats, its early robotics experiments are encouraging, and investors are already discussing a valuation far above the company's previous round. We want more public access and more physical-world evidence before moving it above Decart.

Odyssey takes fourth. The company continues shipping meaningful research rather than repeating the same demo, with multi-agent worlds through Agora-1 and adversarial world-model training through PROWL-1.

Wayve ranks fifth because its world models already sit inside a serious autonomy program connected to major automakers. Runway follows closely thanks to GWM-1, its robotics tooling and an existing commercial business.

AMI Labs has enormous potential but too little public evidence. Waabi has exceptional simulation technology inside a narrower autonomous-driving business. Veeda AI has one of the strongest new founding teams we have seen lately, though it is simply too early to put the company above startups with working products.

The field remains unusually open. We have plenty of impressive simulations now. What we still have not seen is a general world model that repeatedly gives robots or agents a huge, measurable advantage across several unrelated environments.

The company that demonstrates that will probably stop this ranking from being difficult. For now, World Labs leads, Decart looks like the strongest immediate challenger, and General Intuition is the startup moving up the board fastest.

Top world-model startups today