World models: which AI lab is ahead?

Last updated: 21 October 2026
Cover of NewMarketPitch's Frontier AI Labs: Who's Ahead report Which labs lead world models, and with what? World Labs, bought by AMD for about $8.2B. Wayve and Yann LeCun's AMI each raised over $1B, while Google and Nvidia ship their own models. Chart from NewMarketPitch's Frontier AI Labs report.Who is fighting on which front of frontier AI? Chat assistants are the most crowded front, with big tech, the US labs and three Chinese rivals. Europe shows up mostly through Mistral, on three fronts. Chart from NewMarketPitch's Frontier AI Labs report.Which frontier AI labs raised $1B+ in their first round? A handful of research labs, most of them only months old and with nothing to sell yet. A typical startup's first round is a few million dollars. Chart from NewMarketPitch's Frontier AI Labs report.Which frontier AI labs raised a fortune with no product? Safe Superintelligence leads, with about $8B and still nothing customers can use. Several other labs raised hundreds of millions, some over $1B, with no product on the market yet. Chart from NewMarketPitch's Frontier AI Labs report.Which other frontier AI labs have raised over $1B? More than a dozen, after OpenAI, Anthropic and xAI. Most are Chinese model makers or research bets that have barely shipped anything yet. Chart from NewMarketPitch's Frontier AI Labs report.Who is fighting whom in frontier AI? OpenAI leads Google on users, Anthropic leads OpenAI on sales, and Alibaba leads DeepSeek on downloads. Alibaba and Kling split video, SSI and Thinking Machines split money and product. Chart from NewMarketPitch's Frontier AI Labs report.Which kinds of frontier AI labs get the money? Mostly chatbots. Recently, more than $9 of every $10 raised by frontier AI labs went to language and reasoning models. Video, voice, images and robots share the rest. Chart from NewMarketPitch's Frontier AI Labs report.How long did each frontier AI lab take to raise its first $1B? From a few months to almost six years. Almost every lab founded since 2023 crossed $1B in under two years, while OpenAI needed more than three. Chart from NewMarketPitch's Frontier AI Labs report.

Every funding round, every investor, and what the labs promised in return, in one document. Have a look ↓

Summary

Google DeepMind is currently ahead in the overall world-model race, thanks to the combination of Genie 3's interactive environments, SIMA 2's autonomous agents, and Gemini Robotics' physical reasoning. NVIDIA is its strongest competitor in practical physical AI, while World Labs, Runway, and Meta lead in different specialties.

Google's advantage comes from research depth. Its Dreamer reinforcement-learning systems, Genie environment generators, SIMA agents, and robotics models address different parts of the same problem: helping AI understand an environment, predict what happens next, and act within it.

NVIDIA may have the stronger platform strategy. Cosmos 3 combines physical reasoning, world generation, and action prediction in downloadable models, giving robotics developers tools they can adapt rather than a research system they can only observe.

Runway has narrowed the visible gap in interactive generation. GWM Worlds 2 delivers 720p video at 24 frames per second, matching Genie 3's headline video specifications, while adding synchronized audio and sessions without a preset duration limit. Neither company has established superior long-term physical consistency through independent testing. Without independent tests, claims of better consistency are hard to trust, so we check who leads on benchmarks and who only claims it in our frontier AI labs report.

World Labs has a different advantage: its Marble and Atlas technologies generate spatially consistent 3D environments that creators can edit, reconstruct, and reuse. AMD's September 2026 agreement to acquire World Labs for approximately $8.2 billion also shows how valuable spatial intelligence has become to computing infrastructure companies.

Meta could eventually challenge the entire generative approach. V-JEPA 2 learns predictive representations from more than one million hours of video and demonstrated robot manipulation after additional training with only 62 hours of robot data. Predicting useful outcomes without generating every future pixel could prove more efficient for physical AI.

The benchmark picture is much less decisive than model announcements suggest. WorldScore, Physics-IQ, WorldModelBench, and robotics evaluations measure different abilities, making it impossible to combine their results into one reliable ranking of all leading labs.

World models already have practical applications, but the commercial opportunities differ. World Labs can sell tools for generating reusable 3D environments, NVIDIA can supply robotics infrastructure, and Runway can explore interactive entertainment. Google's broader simulation ambitions still lack comparable public evidence of commercial deployment.

The biggest technical uncertainty is whether simulated experience improves real-world performance. Creating a convincing environment or completing tasks inside it does not establish that an agent will handle unfamiliar physical objects, friction, collisions, and unpredictable conditions more successfully.

Our conclusion is that Google deserves the overall lead today, NVIDIA has the clearest open physical AI infrastructure advantage, and World Labs is particularly strong in usable 3D environments. The ranking could change quickly if another lab demonstrates substantially better physical prediction or consistent improvements in real robot task completion under comparable testing conditions.

Written by the New Market Pitch teamWe track new markets so founders and investors can move faster. About us

Which labs lead world models, and with what? World Labs, bought by AMD for about $8.2B. Wayve and Yann LeCun's AMI each raised over $1B, while Google and Nvidia ship their own models. Chart from NewMarketPitch's Frontier AI Labs report.
Frontier AI Labs reportThis page and 70+ morePDF Get the report

This is just a preview page. Don’t worry if a number looks old: the report gets refreshed, and you’ll always get the latest version.

Why are Google, NVIDIA, and other AI labs racing to build world models?

World models have become one of the biggest AI research races because Google DeepMind, NVIDIA, Meta, Runway, and World Labs are now building systems that can simulate environments, predict physical events, and help AI agents learn how to act.

The race accelerated after Google DeepMind introduced Genie 1 in February 2024, Genie 2 in December 2024, and Genie 3 in August 2025. NVIDIA launched Cosmos in January 2025 and followed with Cosmos 3 in May 2026. Meta released V-JEPA 2 in June 2025, while World Labs moved from Marble's public release in November 2025 to Atlas in September 2026. Runway introduced GWM-1 in December 2025 and GWM Worlds 2 nine months later. Several labs shipping world models within a year is a lot of change, and we go through how much has moved in our report on frontier AI labs.

These companies are chasing different goals. Google and Runway want AI to generate environments that people and agents can explore. World Labs focuses on building consistent 3D spaces. NVIDIA is connecting world simulation to robot actions, while Meta is developing models that learn to predict events by watching videos.

The question is which approach will produce AI systems that can understand and navigate unfamiliar situations. A model that accurately predicts what happens when a robot pushes, lifts, or drops an object could ultimately be much more valuable than one that creates convincing virtual worlds.

What makes a world model different from an AI video generator?

A world model goes beyond generating convincing videos because it attempts to represent how an environment works and predict what happens when something changes.

Google DeepMind's Genie 3 generates navigable environments that respond to user movements. NVIDIA's Cosmos 3 combines video generation with physical reasoning and action prediction. Meta's V-JEPA 2 takes a different approach, learning predictive representations from video without generating every pixel of the future scene.

World Labs focuses heavily on spatial consistency. Its models create three-dimensional environments in which camera positions and viewpoints correspond to an underlying representation of the same space.

The differences become clear when we imagine a glass falling from a table. A video model might generate a realistic-looking sequence of the glass hitting the floor. A world model should predict how the glass moves, where it lands, and what happens when the conditions change.

We therefore need to compare whether an environment remains consistent, whether users can interact with it, and whether its predictions are accurate enough to help an AI agent make decisions.

Is Google DeepMind currently winning the world-model race?

Google DeepMind is currently our overall leader in world models because it has connected interactive environment generation, autonomous agent research, and robotics more convincingly than any other AI lab.

Genie 3 generates photorealistic environments at 720p and approximately 24 frames per second. Users can navigate those environments, revisit locations, and introduce events that change the simulated scene. Google reports that the model can maintain visual consistency across several minutes of interaction.

The advantage becomes more interesting when we connect Genie to Google's other research programs. DeepMind has demonstrated SIMA agents navigating unfamiliar virtual environments, while Gemini Robotics develops AI systems capable of reasoning about and controlling physical robots.

Google also has years of relevant research behind these systems, including Dreamer, a family of reinforcement-learning agents that learn predictive models of their environments. DreamerV3, published in 2023, demonstrated strong performance across more than 150 tasks using a single configuration.

That gives Google's world-model research considerably more depth than Genie alone suggests. NVIDIA has stronger open physical AI infrastructure, however, World Labs has impressive spatial reconstruction capabilities, and Runway now competes directly in interactive environment generation.

Who is fighting on which front of frontier AI? Chat assistants are the most crowded front, with big tech, the US labs and three Chinese rivals. Europe shows up mostly through Mistral, on three fronts. Chart from NewMarketPitch's Frontier AI Labs report.
Frontier AI Labs reportThis page and 70+ morePDF Get the report

Just a preview. If a number here looks old, no worries: your report comes with the latest data.

Has Runway caught up with Google's Genie 3?

Runway has caught up with Google DeepMind on several visible world-generation specifications, although GWM Worlds 2 has not demonstrated that it is a better general-purpose world model than Genie 3.

Runway introduced GWM Worlds 2 in September 2026, demonstrating continuous interactive video at 720p and 24 frames per second, with synchronized 48 kHz audio. Those video specifications match the resolution and approximate frame rate Google published for Genie 3 in August 2025.

Runway also offers a different approach to interaction. Users can move the camera and issue text instructions that affect characters, objects, weather, or the environment. Multiple actions can overlap, allowing the scene to change while the user continues navigating. A different take on interaction leaves room for newcomers; we map where the openings are in our frontier AI labs market report.

One major distinction is session length. Google describes Genie 3 as supporting interactions lasting several minutes, while Runway designed GWM Worlds 2 without a predetermined generation limit.

But unlimited generation isn't the same as unlimited memory. Runway acknowledges that difficult camera movements can degrade visual details and that long-term memory remains imperfect. Independent testing has yet to establish how well either system maintains physical relationships and object permanence over extended interactions.

Capability Google Genie 3 Runway GWM Worlds 2
Main focus Interactive world simulation Interactive audiovisual worlds
Published resolution 720p 720p
Published frame rate Around 24 fps 24 fps
Continuous interaction Several minutes No preset duration limit
Camera control Interactive navigation Interactive navigation
World modification Promptable events Text-directed actions
Generated audio Not a headline Genie 3 capability 48 kHz audio
Agent applications Demonstrated through SIMA Experimental agent-control demonstrations
Current availability Experimental access Research preview
Biggest unanswered question Long-term physical consistency Long-term physical consistency

Is World Labs better than Google at creating realistic 3D worlds?

World Labs is one of the strongest AI labs for generating persistent, editable 3D environments, and its Atlas model has capabilities that Google's Genie 3 does not emphasize.

World Labs released Marble publicly in November 2025. Users can generate navigable 3D environments from text, images, video, and spatial layouts, then edit or export the results into other 3D workflows.

In September 2026, the company introduced Atlas, which builds on this direction through a model trained across text, images, video, and three-dimensional information.

Atlas can generate camera-controlled video at up to 1440p for one minute and reconstruct environments from a small number of reference images. World Labs reports that two or three photographs can often produce recognizable reconstructions, with larger image sets improving the results.

These capabilities are especially useful for architecture, game development, virtual production, and other applications where creators need to control exact camera positions or reuse the same virtual environment.

Genie 3 takes a different approach. It generates interactive experiences that can respond continuously as users move through them, making the system particularly interesting for simulated agent training.

World Labs has also made its technology more accessible for developers. Marble supports the export of generated 3D assets, and the World API, introduced in January 2026, allows external applications to create environments programmatically.

There's also a major corporate development. On September 28, 2026, AMD announced an agreement to acquire World Labs in an all-stock transaction valued at approximately $8.2 billion. The deal is expected to close by the end of 2026, potentially giving World Labs far greater hardware and infrastructure resources.

For editable 3D content, we would currently favor World Labs. For general-purpose interactive simulation and autonomous agents, Google has the stronger research position.

Which frontier AI labs raised $1B+ in their first round? A handful of research labs, most of them only months old and with nothing to sell yet. A typical startup's first round is a few million dollars. Chart from NewMarketPitch's Frontier AI Labs report.
Frontier AI Labs reportThis page and 70+ morePDF Get the report

This is a preview page. Spotted an outdated number? That’s normal here. The report you get is always the freshest version.

Is NVIDIA ahead of Google in world models for robots?

NVIDIA has the strongest claim to leadership in open physical AI world-model infrastructure, while Google remains ahead in our overall assessment of general-purpose world-model research.

NVIDIA introduced Cosmos 3 in May 2026, combining physical reasoning, video generation, and action prediction. The platform includes Cosmos 3 Super with 64 billion parameters and Cosmos 3 Nano with 16 billion parameters.

NVIDIA reports strong performance across several physical AI evaluations, including Physics-IQ, PAI-Bench, R-Bench, VANTAGE-Bench, and RoboLab. These benchmarks examine different aspects of physical understanding and robotic behavior. NVIDIA reporting strong physical AI results is a claim worth weighing; we sort out who is ahead in our report on who's ahead in frontier AI labs.

The bigger advantage is how NVIDIA distributes its technology. Developers can download Cosmos models, customize them, and combine them with NVIDIA's simulation and robotics tools.

Its ecosystem also includes robotics companies and researchers developing real systems. NVIDIA has worked with organizations such as Agile Robots, Generalist, and Skild AI, alongside developers of autonomous vehicles and industrial robotics.

Google's Genie technology has been much more restricted. Developers cannot generally download Genie 3 and modify it for their own robots in the same way they can work with available Cosmos checkpoints.

That difference could have major consequences. If robotics companies increasingly build on Cosmos, NVIDIA could become essential infrastructure for physical AI even without producing the single best-performing world model.

For now, NVIDIA is our preferred leader in accessible robotics-oriented world models. Google's broader research program still gives it the overall edge.

Has Meta fallen behind Google and NVIDIA in world models?

Meta remains a serious contender in predictive world models, although V-JEPA 2 has a narrower visible set of applications than Google's Genie research or NVIDIA's Cosmos ecosystem.

Meta released V-JEPA 2 in June 2025 after pretraining on more than one million hours of internet video. The model learns representations that help predict events and understand actions without generating detailed video frames for every possible future.

In its published research, Meta reported 77.3% top-1 accuracy on Something-Something v2, a benchmark focused on recognizing actions and object interactions. The model also achieved 39.7 recall-at-five on EPIC-KITCHENS-100 action anticipation.

Meta then adapted its pretrained model using only 62 hours of robot interaction data from DROID. The resulting system demonstrated reaching, grasping, and pick-and-place tasks in unfamiliar configurations.

The contrast with generative models is important. Meta can focus computing resources on predicting information useful for decisions rather than reconstructing every visual detail that a person might want to see.

This approach could be especially efficient for robots, where correctly anticipating an object's movement is often more important than producing a beautiful video of the event.

Meta has already demonstrated that large-scale video learning can support physical prediction with relatively little robot-specific data. What remains unclear is whether V-JEPA's approach can match more integrated systems on complex, long-term planning tasks.

Which AI lab is best at understanding real-world physics?

NVIDIA and Meta provide some of the strongest published evidence for physical understanding, but no AI lab has convincingly demonstrated that its world model is best across all types of physical reasoning.

NVIDIA's Cosmos 3 research reports strong results on Physics-IQ, PAI-Bench, and other physical AI evaluations. Meta's V-JEPA 2 performs well on action recognition and anticipation, while Google's Gemini Robotics research connects physical reasoning to robotic manipulation.

The tasks are more different than they initially appear. Recognizing that somebody is opening a drawer is quite different from predicting how much force a robot must apply to open it, or whether an object inside will slide when the drawer moves.

A world model may perform well on video-based questions while struggling with contact forces, friction, collisions, or interactions involving objects it has never encountered.

NVIDIA is trying to combine physical prediction with action generation. Meta concentrates on learning useful predictive representations. Google develops separate but connected models for world generation, reasoning, and robot control.

We would give NVIDIA a narrow advantage in publicly accessible, physically oriented generative modeling, while Meta deserves particular attention for efficient predictive learning.

Which frontier AI labs raised a fortune with no product? Safe Superintelligence leads, with about $8B and still nothing customers can use. Several other labs raised hundreds of millions, some over $1B, with no product on the market yet. Chart from NewMarketPitch's Frontier AI Labs report.
Frontier AI Labs reportThis page and 70+ morePDF Get the report

Preview only. Some numbers may have changed since we made this page, but your report will be up to date.

Can world-model benchmarks actually tell us which AI lab is best?

Current world-model benchmarks can identify strengths and weaknesses, but they still cannot produce one trustworthy ranking covering Google, NVIDIA, Meta, World Labs, and Runway.

A survey of world-model evaluation research finalized in September 2026 identified 102 representative benchmarks published between 2018 and 2026. Of those, 85 covered more than one evaluation category. The lack of a common test for world models is one of the biggest unsolved problems we look at in our frontier AI labs report.

Researchers now test visual quality, spatial consistency, physical plausibility, long-term memory, controllability, and practical usefulness. The problem is that models often perform very different tasks, and some leading systems cannot be evaluated under the same conditions.

WorldScore examines camera-controlled world generation. Physics-IQ focuses on physical behavior. WorldModelBench tests instruction following and physical plausibility using 350 evaluation instances calibrated against approximately 67,000 human judgments.

Robot evaluations add a different requirement: the model must help a physical or simulated agent complete a task successfully.

The lack of common testing is one reason we favor Google overall without claiming it leads every individual capability.

Benchmark Main focus What it cannot prove alone
WorldScore Consistent and controllable world generation Real-world physical accuracy
WorldModelBench Instructions and physical plausibility Long-term agent reliability
Physics-IQ Physical behavior General 3D reconstruction quality
VBench-2.0 Video quality and related dimensions Reliable interactive simulation
RoboArena Evaluated robot policy performance Universal world-model superiority
RoboLab Robotics-related model evaluation Broad interactive-world quality
Something-Something v2 Understanding object interactions Accurate future-state simulation
EPIC-KITCHENS-100 Anticipating human actions Real robot task success
NAVSIM Driving policy evaluation General-purpose world understanding
SIMA agent tasks Acting in unfamiliar virtual environments Reliable real-world robot control

Can Google's Genie 3 actually train better AI agents?

Google has shown that Genie 3 can provide interactive training environments for AI agents, but there is still limited public evidence that this produces better performance on real-world tasks.

DeepMind has demonstrated SIMA agents operating within Genie-generated environments. The agents receive instructions, navigate unfamiliar settings, and react as the simulated world changes.

The company strengthened this connection with SIMA 2 in November 2025, integrating Gemini-based reasoning into agents designed to handle unfamiliar virtual worlds.

Earlier SIMA research had already covered more than 600 instruction-following skills across several commercial video games. SIMA 2 extended the work toward more flexible reasoning and learning from generated experience.

However, the critical result would be measurable improvement outside the environments used for training. An agent might learn to navigate simulated rooms successfully while struggling with real objects, unpredictable surfaces, or unfamiliar physical conditions.

We would need controlled comparisons showing that training with Genie improves performance against agents trained without it, ideally across real robotics or independently designed tasks.

Is World Labs already turning world models into useful products?

World Labs has one of the clearest near-term commercial applications for world models because its technology already produces editable 3D environments that developers and creators can reuse.

Marble became publicly available in November 2025, followed by the World API in January 2026. The API allows developers to generate navigable environments from text, images, panoramas, and video.

World Labs also developed Spark, an open-source renderer for Gaussian splats, helping developers display 3D environments in web applications.

Atlas expanded those capabilities through improved reconstruction and camera-controlled generation. Together, these releases show a consistent direction toward tools that fit into existing creative and software workflows.

The opportunity is particularly tangible for game development, architecture, virtual production, and immersive applications. These industries already work with 3D assets and can benefit from faster environment creation without waiting for fully autonomous AI systems.

NVIDIA has a different commercial advantage. Cosmos fits into its existing robotics, simulation, and computing ecosystem, allowing the company to supply infrastructure to developers of autonomous vehicles, industrial robots, and physical AI systems.

Runway could bring its world-model research into interactive entertainment and creative applications, building on its established generative-media business. However, GWM Worlds 2 remains a research preview, and its commercial adoption has not yet been demonstrated at scale.

Google's Genie research is more ambitious in general simulation and autonomous agent training, although publicly demonstrated research does not automatically translate into paying customers. Meta's V-JEPA approach could also become useful infrastructure for AI developers, but its commercial opportunity is less directly visible.

We still lack public financial data showing how much revenue World Labs generates from these products. Its agreed acquisition by AMD establishes substantial strategic value, but the approximately $8.2 billion transaction price should not be confused with recurring product revenue.

Which other frontier AI labs have raised over $1B? More than a dozen, after OpenAI, Anthropic and xAI. Most are Chinese model makers or research bets that have barely shipped anything yet. Chart from NewMarketPitch's Frontier AI Labs report.
Frontier AI Labs reportThis page and 70+ morePDF Get the report

This is just a preview page. Don’t worry if a number looks old: the report gets refreshed, and you’ll always get the latest version.

Could NVIDIA's open models help it overtake Google?

NVIDIA's open-model approach could eventually give Cosmos a larger developer ecosystem than Google's Genie, especially among robotics companies that need to customize their models.

Cosmos 3 includes downloadable models, development tools, and training resources that external organizations can adapt to their needs. Downloadable models against tightly controlled access is a business-model choice, and we explore which business models are winning in our report on frontier AI labs.

Google has taken a more controlled approach to Genie 3. Access has largely been limited to experimental products and research environments rather than broadly available model weights.

For robotics teams, that distinction is practical. Developers often need to train models on their own cameras, robot bodies, operating environments, and task-specific data.

Meta also offers open research models through V-JEPA 2, while World Labs provides API access and exportable 3D assets. These options serve different needs, from modifying model weights to integrating generated environments into applications.

NVIDIA has an additional advantage because Cosmos fits alongside its existing simulation, robotics, and computing ecosystem. A company already using NVIDIA tools can potentially adopt world-model components without replacing its entire development stack.

NVIDIA's position resembles a platform strategy: it can benefit from many organizations building physical AI applications, even when those organizations develop their own specialized models.

That could become more valuable than winning individual research benchmarks. However, no public data yet establishes that Cosmos has become the dominant world-model foundation across the robotics industry.

Are world models actually making robots better today?

World models are already helping robotics researchers test planning and control methods, but we still lack enough independent evidence to say how much they improve robots across real commercial tasks.

Meta's V-JEPA 2 demonstrated how video pretraining could support robot manipulation after additional training with 62 hours of interaction data. NVIDIA's Cosmos research has focused on predicting both future observations and robot actions, while Runway has explored action-conditioned video prediction through GWM Robotics.

World Labs has also investigated reconstructing environments so robot policies can train in simulations based on real-world spaces.

Collecting real-world robot data involves hardware, supervision, maintenance, and potential failures. Simulation can produce additional training experiences more cheaply once the underlying system has been developed.

However, the benefits depend on how accurately the simulated environment behaves. Incorrect predictions about weight, friction, collisions, or object movement can teach robots behaviors that fail when deployed.

The next meaningful evidence would be consistent improvements in physical task completion rates, ideally across different robots and environments.

Could Wayve beat Google and NVIDIA in driving world models?

Wayve is one of the strongest specialists in driving world models, although its work is too focused on autonomous vehicles to establish an overall lead against Google DeepMind or NVIDIA.

Wayve introduced GAIA-1 in 2023 and GAIA-2 in March 2025. The models generate driving scenarios to support autonomous vehicle research, synthetic data creation, and evaluation.

Driving places demanding requirements on world simulation. Models must represent road layouts, pedestrians, vehicles, lighting, and unusual traffic situations while keeping events consistent as the simulated vehicle moves.

Wayve benefits from concentrating its research on a specific problem. Google and NVIDIA are developing technologies intended to work across a much wider range of environments and robot systems.

NVIDIA's Cosmos ecosystem also serves autonomous vehicle developers, so the two companies compete in some applications despite following different strategies.

A fair comparison would require testing both systems on common driving scenarios and measuring whether their generated data improves autonomous driving performance.

Wayve could be ahead on selected driving tasks without being ahead in general world modeling. We consider it a serious specialist rather than a direct challenger for the overall title.

Who is fighting whom in frontier AI? OpenAI leads Google on users, Anthropic leads OpenAI on sales, and Alibaba leads DeepSeek on downloads. Alibaba and Kling split video, SSI and Thinking Machines split money and product. Chart from NewMarketPitch's Frontier AI Labs report.
Frontier AI Labs reportThis page and 70+ morePDF Get the report

Just a preview. If a number here looks old, no worries: your report comes with the latest data.

Why are world-model researchers building such different types of AI?

Google, NVIDIA, Meta, World Labs, and Runway are developing different world models because the best way to generate a convincing environment may be very different from the best way to predict actions.

Google's Genie and Runway's GWM Worlds generate interactive visual experiences. Their approach makes sense when humans or autonomous agents need to explore and interact with simulated environments.

World Labs gives more importance to explicit three-dimensional structure, allowing environments to remain consistent across different camera positions and to be reused in creative workflows. Interactive video and explicit 3D are two competing designs; we follow which technologies are gaining traction in our frontier AI labs report.

Meta's V-JEPA 2 avoids reconstructing every visual detail. Instead, it learns predictive representations that focus on information useful for understanding events and planning.

NVIDIA's Cosmos 3 combines several approaches, including physical reasoning, video generation, and action prediction, within a broader architecture for physical AI.

We can also see different release strategies. Google introduced three Genie generations between February 2024 and August 2025, while Runway progressed from GWM-1 in December 2025 to GWM Worlds 2 in September 2026. World Labs moved from Marble's public release to Atlas over roughly ten months.

These release intervals show rapid development across the field, although they cannot establish which company's underlying technology is improving fastest.

The technical differences may prove more important than the speed of releases. A model optimized to produce beautiful interactive environments can consume considerable computing power on visual details that a robot does not need.

Conversely, an efficient predictive model may understand important aspects of movement without providing the detailed environment generation required for immersive simulations.

It is quite possible that future world models will combine several of these techniques rather than follow one architecture exclusively.

Does Google have a bigger data advantage than NVIDIA and Meta?

Google has enormous computing and research resources, but there is no reliable public evidence that it possesses the largest or best training dataset specifically for world models.

Google's advantage comes from its broader research infrastructure and experience with video, simulation, robotics, and autonomous agents. Its Genie, SIMA, and Gemini Robotics programs also create opportunities to connect different types of data and experiments.

NVIDIA has disclosed more concrete information about parts of its physical AI training resources. A Cosmos technical publication described approximately 767 million images, 348 million dynamics videos, and eight million action samples.

Meta's V-JEPA 2 research reported more than one million hours of video pretraining. This shows the scale of observation-based learning, although video hours and individual image or action samples are different measurements and cannot be compared directly.

World Labs also has substantial financial backing. The company raised $1 billion in February 2026, with investors including AMD, Autodesk, NVIDIA, and Fidelity. AMD subsequently agreed to acquire the company for approximately $8.2 billion in September, with closing expected by the end of 2026.

However, funding and computing resources only explain part of the competitive position. A million hours of internet video may help a model understand what events look like, while a much smaller collection of carefully recorded robot interactions can provide information about actions and their consequences.

Google and NVIDIA appear exceptionally well positioned to scale this research, with Meta also possessing significant resources. We cannot establish a clear winner in world-model training data from the numbers companies have published.

Which kinds of frontier AI labs get the money? Mostly chatbots. Recently, more than $9 of every $10 raised by frontier AI labs went to language and reasoning models. Video, voice, images and robots share the rest. Chart from NewMarketPitch's Frontier AI Labs report.
Frontier AI Labs reportThis page and 70+ morePDF Get the report

This is a preview page. Spotted an outdated number? That’s normal here. The report you get is always the freshest version.

What would prove that one world model is really better than the others?

The clearest proof of world-model superiority would be better predictions and better agent performance on the same unfamiliar tasks, especially when the results transfer to the real world.

For interactive environments, researchers should test whether a model remembers objects after they leave the camera's view, preserves spatial relationships, and responds consistently when the same action occurs under comparable conditions.

Physical reasoning requires additional tests involving collisions, friction, falling objects, object manipulation, and interactions between multiple agents.

Long-term consistency is also essential. A model may appear convincing during a short demonstration but accumulate errors after hundreds of actions.

Robotics provides a particularly useful way to evaluate practical value. Researchers could train identical robot policies with and without a particular world model, then compare task completion rates on physical robots in unfamiliar environments.

The most convincing results would show reliable improvements across several robot types and tasks, along with a measurable reduction in real-world training data requirements.

Until leading laboratories publish more comparable evaluations, we should treat impressive demonstrations as evidence of individual capabilities rather than definitive proof of overall superiority.

World models: Which AI lab is actually ahead?

Google DeepMind is currently ahead in the overall world-model race, with NVIDIA leading in open physical AI infrastructure, World Labs standing out in editable 3D environments, and Runway rapidly closing the gap in interactive generation.

Our ranking puts Google first because its research connects Genie 3's environment generation, SIMA 2's autonomous agents, and Gemini Robotics' physical reasoning and control. The combination covers more of the capabilities needed for general-purpose world modeling than any other laboratory has demonstrated as clearly.

NVIDIA is the most consequential alternative. Cosmos 3 combines physical reasoning and action prediction with downloadable models and a growing ecosystem of robotics developers. Its commercial and distribution advantages could become increasingly important as world models move from research into industrial applications.

World Labs deserves a separate distinction for its explicit spatial representations and practical 3D creation tools. Atlas and Marble make the company especially competitive where developers need persistent, editable environments. Its pending acquisition by AMD could give this work substantially greater computing and engineering resources.

Runway has become a much stronger contender with GWM Worlds 2, matching Genie 3's published video resolution and frame-rate class while supporting continuous audiovisual interaction without a preset duration limit.

Meta remains important through V-JEPA 2, which demonstrates a different route toward useful physical prediction using extensive video pretraining.

The biggest unresolved question is whether any of these systems can consistently improve an AI agent's performance on unfamiliar, real-world tasks. The public evidence is still too limited to establish a universal technical winner.

For now, Google DeepMind earns first place based on the breadth of its research and its demonstrated connection between simulated worlds and autonomous agents. NVIDIA is the closest competitor for practical physical AI, while World Labs and Runway are particularly strong in their respective specialties.

We would change that ranking if a competitor demonstrated substantially better physical prediction or reliable gains in real-world robot performance under common testing conditions. That is the evidence that could eventually turn today's fragmented world-model race into a clear technical victory.

How long did each frontier AI lab take to raise its first $1B? From a few months to almost six years. Almost every lab founded since 2023 crossed $1B in under two years, while OpenAI needed more than three. Chart from NewMarketPitch's Frontier AI Labs report.
Frontier AI Labs reportThis page and 70+ morePDF Get the report

Preview only. Some numbers may have changed since we made this page, but your report will be up to date.

Sources and methodology

This analysis compares the leading AI laboratories developing world models using eight research and application dimensions: world generation, spatial consistency, physical reasoning, interactive control, autonomous agents, robotics applications, technical accessibility, and commercial deployment. We examined Google DeepMind, NVIDIA, Meta, World Labs, Runway, and specialist developers including Wayve.

We reviewed original model announcements, technical publications, benchmark research, and documented demonstrations. We separated published specifications from independently measured results, research capabilities from commercial deployments, and specialized achievements from progress toward general-purpose world modeling. We also distinguished model access, downloadable weights, developer APIs, and exportable assets rather than treating them as equivalent forms of availability.

As explained above, no single benchmark measures every capability relevant to world models. We therefore evaluated the evidence within its respective domain rather than combining unrelated results into an artificial score. We gave particular weight to the connection between simulation, physical prediction, and autonomous action, while recognizing that successful performance in generated environments does not automatically establish real-world reliability.

The overall ranking reflects the breadth, relevance, and strength of demonstrated capabilities, not investment levels, model parameters, release frequency, or visual quality alone. Where the evidence was incomplete or systems could not be tested under comparable conditions, we treated leadership as an analytical judgment rather than a definitive experimental result.

Key Google DeepMind sources include Genie 3's interactive world-generation research, SIMA 2's autonomous-agent research, Gemini Robotics, and Gemini Robotics 1.5. For NVIDIA, we examined the Cosmos 3 technical overview, the original Cosmos foundation-model platform, and its synthetic-data and physical-reasoning research.

Meta's predictive modeling was assessed through its V-JEPA 2 announcement and benchmark findings and the V-JEPA 2 research paper. World Labs sources include Atlas, Marble, the World API, its February 2026 financing announcement, and AMD's September 2026 acquisition announcement. The proposed acquisition was treated as an announced transaction, not a completed integration or evidence of product revenue.

For interactive generation, we used Runway's original publications on GWM Worlds 2 and GWM-1. Wayve's GAIA-2 research provided the specialist autonomous-driving comparison. Evaluation frameworks included WorldScore, WorldModelBench, Physics-IQ, VBench-2.0, and the 2026 survey of 102 world-model benchmarks.

We prioritized original laboratory publications and primary benchmark research that provided specific, checkable evidence. Our conclusions aggregate those findings across capabilities, with particular attention to what each laboratory has demonstrated, which claims remain dependent on company-reported results, and what additional evaluations would justify changing the ranking.