Nvidia Corp 2026 Earnings Call

Welcome. Computer graphics is entering a new era, SIGGRAPH remains the place where that future comes into focus. Every generation of computer graphics has had defining moments. This is one of them. For more than 30 years, NVIDIA has been part of the SIGGRAPH community, introducing breakthroughs that have reshaped the industry, from the programmable shading GPU, to RTX, to Omniverse. Today, that legacy continues as graphics, simulation, and AI converge, laying the foundation for the next generation of creative tools, digital twins, and physical AI. Please join me in welcoming NVIDIA to the stage for the 2026 SIGGRAPH Research Keynote.

Hello, SIGGRAPH L.A. Incredible to be with you. 30 years ago, NVIDIA set out to build a new kind of computer, one that could solve problems traditional computers simply couldn't. Our first application was computer graphics, geometry, programmable shading, simulation, physics, materials, lighting. We came to SIGGRAPH, where so many foundational ideas became real. The Utah teapot, 3D geometry processing, Turner Whitted and ray tracing, Pixar, and programmable shading. These ideas changed computer graphics and computing. In 2001, NVIDIA invented the world's first programmable GPU. five years later, we introduced CUDA, turning the GPU into a parallel processor for science and engineering. In 2012, CUDA GPUs made first contact with AI. GPUs, the processor invented for graphics, sparked modern AI. In 2018, right here at SIGGRAPH, we introduced RTX. AI comes back to graphics. Real-time ray tracing, neural rendering, DLSS. Hand-coded algorithms fused with neural networks.

With DLSS 1, one out of every four pixels was inferred. Today, entire patches of the screen are inferred. At SIGGRAPH, we are expanding the canvas of creativity and expression. Artists and designers as far back as Leonardo da Vinci have invented technology in service of creativity. Creators and designers need tools powerful enough to expand their imagination, malleable enough to give them the freedom to shape ideas, and precise enough to realize their vision exactly as intended. We want the power of AI, but grounded in 3D, governed by physics, so that it is shapeable by creators. AI will generate amazing virtual worlds, and in these virtual worlds, the next era of AI, robotics, will begin. Before robots operate in the real world, they will learn in virtual worlds with synthetic experiences. That is why computer graphics matter more than ever.

Whether for games, cinema, robotics, or factory digital twins, the goal is the same: to create virtual worlds that behave with the fidelity and realism of the physical world. There is a complete reset in computer graphics. No community is better able to shape its future than SIGGRAPH. You stand at the intersection of art, science, and engineering. You are the da Vincis of our time. This is your moment. Now I'm delighted to introduce Edward Liu, Director of Applied Deep Learning Research and the technical leader behind the making of DLSS. Thank you, SIGGRAPH. Have a great show.

Good afternoon. Thank you, Jensen, for the introduction. Wow. The history of graphics film is always awesome to watch. I never get tired of watching it. You know what? A lot of it actually started right here as SIGGRAPH paper. Today, I have only one thing for you. I think that could be the next chapter of that film. DLSS 5. We showed just a short glimpse of it at GTC Spring this year. At SIGGRAPH, we get to go deeper. What it is, why we built it the way we did, and why we're so excited about it. DLSS 5 goes back to Turing, which is the first GPU in which we put RT core and Tensor core side by side right on the same chip.

Even back then, we had the belief that graphics and AI would have to advance together in order to achieve real-time photorealism one day. Let me show you what that means. There's a pattern behind that film. Every once a few year, a revolutionary technology come along. The image realism and fidelity of real-time rendering will leapfrog the previous generation. It opens a new era, a new platform. It chart a new curve on this graph. Programmable shader was definitely one of the good examples. Suddenly, all the game artists can write program to shade every pixel. I still vividly remember how big of a jump it is in game fidelity. Fast-forward to today, we're at path tracing with DLSS 4. Fully simulated light at real-time frame rate. Look at the top. We're still not there yet.

We're still not at indistinguishable from reality in real-time rendering. Where does the next curve come from? Let's start with what we know. Traditional graphics build images from first principles. We author the scene, we simulate the light bouncing inside of it using physics. Then we frame it with a camera that we control down to the accuracy of a millimeter. That's rendering. Every pixel is there because somebody wants it to be there. While we're doing that, something else happened. Generative models learn photorealism. Once we feed those model enough of the real world, eventually they learn how skin will scatter light, how the hair will glow if they're backlit, or how the glass will bend the world around them. Those generative models are slow and giant. They're not as easily controllable as rendering can.

Wouldn't it be great if there's a way that we can get the best of both worlds? With that question in mind, we set out to seek the next leap in image fidelity for real-time rendering. This is what we ended up doing. The renderer keeps building the world exactly as the game has authored it. The generation becomes the learned stage afterwards to enrich its appearance. Learned generation defines the world. Generation enriches the appearance. That's the core intuition behind DLSS 5. I can describe that all day. This is SIGGRAPH, so let's roll this reel. Cool. Yeah, thank you. Everything you just saw is traditional rendering followed by DLSS 5 generating the final photoreal appearance in real time. I think it looks great. It's a new level of realism. Looking real here is only half the problem.

That is tricky because all the generative models are probabilistics. You can ask it for a portrait of a merchant at sunset. You get a gorgeous merchant. You ask again, you get a different merchant. If you ask for 10 times, you get 10 different merchants. Generally, the field is steering these generative models with various abstract signals, like text prompt, reference images, 3D proxies, or even storyboards. By the way, this is a really fast-moving research area that is really fascinating to follow. They all work. It is just that every one of them still leaves a lot of the room for the model to improvise. That is totally fine for offline generation. For real-time rendering, we don't get to pick.

Whatever the model generate will get presented to the viewer in the right next frame. We need some condition that is more constrained and rigorous. You know what? There is such signal that is hiding in plain sight. We're going to use the frame the game just rendered. It's not a prompt. It's not a reference. It's not a proxy scene. It's a renderer's own output. It's one signal that is actually universally available across all the engines, path tracing, rasterizers, hybrid, whatever. Unlike a prompt, this is actually a very information-dense and pixel-aligned view of the current shot and can really provide the grounding for a lot of things, like character identity, the lighting setup, the mood, the camera composition, and so on. We took the render frame and make it more realistic.

The idea is really simple, but the moment we actually set out to build such things, it turns out to be incredibly challenging. Three fundamental and tough research challenges immediately presented themselves. The first one, preserving the artistic intent. Even when we give the model a render frame, a generative model will change it. That's its whole job. Every change, it comes a chance that it will drift away from the original intent. Say the hero has a scar and the model just don't like it removes it. Yeah, the image now looks cleaner, but it's the wrong character. We drew a hard line and trained the model in a way to respect those intent.

We used some of the renderer's internal buffers, such as albedos, surface normals, and some of the lighting information to make sure the model can understand what it has to stay consistent with. Within those constraint, the model is free to enrich what sells realism. Things like better subsurface scattering, finer material responses, the light transmission across hair and foliage, even better contact shadows, and better environment lighting. The scar can pick up better skin shading, but it should never disappear. DLSS 5 must change the image in a way that it does not change the story. That's the first challenge. Challenge number two, temporal streaming. Most video generators work in chunks. They reason across multiple frame and they generate a batch of frame at once.

With real-time render, we don't get to do that again because it has to respond to the user input immediately and present and render frame by frame. Luckily, when we're working within the game engine, we do have access to things like motion vectors and all the DLSS have been using them from day one. They can just tell the model how things are moving across adjacent frames, and the model does not need to infer across pixels and frames to figure out what the motion is. Thanks to that and other things, we managed to make DLSS 5 run causally and frame by frame, one frame in, one frame out, and it does not need to peek ahead, and it remains temporally stable and coherent while doing so. No shimmer, no drifting, no swimming. Challenge number 3, speed. A 4K frame is made of about 8 million pixels, and when we're rendering at larger than 60 FPS, we get really less than 16 milliseconds, and most of that is used by the game itself.

Like I said, the foundational model that can generate those photoreal appearance are large and slow, and they can generate anything in seconds. What real-time rendering want is something more specific. It should be smaller and more efficient and more focused. We distilled. We took what those general model that knows about the world appearance and compressed it into a much smaller one-step pixel space diffusion transformer model that does only one thing, make real-time rendering look more real.

It's still not a regular filter because it inherited all the broad world knowledge from those giant foundational models, and it really understand the image that it's looking at, understand characters, understand lighting and context, but what it has lost is generality. It just cannot generate funny cat videos anymore. Because it only does this one thing, we've managed to make it compact and efficient enough to really run side by side in the game and still make the whole experience feel like a game. Narrative intent, frame by frame coherence and streaming, and frame budget. Solve all three, we can have generation live inside a renderer. The render frame is really only one layer of control. The problem is the mapping between a render frame and the more realistic version of it is not one-to-one.

There doesn't exist a reference ground truth for us to match against with, and multiple outcome could make sense. Which look do you want? Where do you want it? How strong do you want to apply? Those are all artistic calls, and they belong to the people making the game, not the model. The best person to show you what that means is an artist. Let's invite Gaff, who's our creative artist, and he will direct DLSS 5 for you guys. Welcome, Gaff. Thanks, Ed. Thanks, everyone.

All right. What I'm going to show you today is how we can direct DLSS 5 from the developer standpoint and art direct it exactly as we want, so the final frame looks exactly as we want it to be. Okay. Before I show you that. Sorry. I think it's working now. Okay. What I'm going to show you today is how we can control DLSS 5 for everything you see on the screen, characters, elements and props, as well as the full environment. Before I do that, I'm going back to what Ed has mentioned, that DLSS 5 is basically enhancing the frame and it's respecting the render frame. It is not changing the geometry in any way, and it's uplifting the image.

If we take a look at this example here, you can see that when we turn DLSS 5 on this character, for example, it's gaining better contrast, better ambient occlusion, better contact shadows. I want you to pay attention to the ears, for example, how they gain subsurface scattering that's previously missing. Look at the ambient occlusion between the top of the head and the piece of cloth on the top of his head, and the eyes gain better reflections. As you can see, DLSS 5 is fully respecting the render frame and is just enhancing it. Now that we have seen that, I'm going to show you the controls that we're building in DLSS 5 to art direct the frame.

I'm going to start from broad level of controls, and I'm going to show you how we can get very granular with how we can control what DLSS 5 is doing to the frame. The first set of controls is actually the models itself. DLSS 5 has different models. As you can see in this frame here, we have different models, model A, B, C. They have different output because they're trained on different parameters. When a developer picks a model, you're not locked into one. The theme that you're going to see in all my presentation here is that you can sort of mix and match whatever you want. Picking a specific model does not lock your whole game, for example, into one. You can pick a model for a scene, a model for a different scene, a model for a cut scene.

You can pick whatever suits you. Some more global controls. We have structure intensity as well as tone intensity. Structure intensity is modulating the high-frequency detail of the frame, and tone intensity is modulating the low-frequency detail of the frame. A simpler way to think about it is structure intensity is basically adding things like better ambient occlusion, better subsurface scattering, and better reflections. Whereas tone intensity is a bit more expressive and is impacting the lighting, is enhancing the lighting. As you can see, if we move the tone slider, it's changing the colors because it's effectively changing the subsurface scattering and the lighting. These are global controls that, again, enhance the whole frame. Now I'm going to show you some controls that we have. There's two different level of controls.

There's controls that come from the engine side and controls that come from the model side. Because DLSS 5 is a smart model, it understands what's in the frame. This, for example, is a model auto-masking. The model understands the character, the base character itself, and we can use the slider to control the intensity at which DLSS 5 is being applied just on the base character, as you can see. This is coming entirely from the model side. I'm going to show you one more example of that. In this example, DLSS 5 is off on the environment. We're just turning it on again just for the character because the model understands which one is the character. This is coming entirely from the model side because the model understands the semantics of the scene.

What I'm going to show you here is an example of the same model with different structure intensity. You can see that structure intensity 25% is obviously closer to the render frame. We have a midpoint, and then we have 95%, which is closer to the maximum uplift that DLSS can provide. Again, you can see the developer has the option to pick whatever intensity they prefer for their final frame. Next, I'm going to show you sort of the opposite of that. I'm going to show you the same structure intensity, but with the different models. Again, different models give different outputs because of the training parameters being different. Now that we've taken a look at characters, I'm going to show you how DLSS 5 is also enhancing the rest of the scene.

In this example here, let's say we want to, again, enhance the whole scene. Actually, we don't want to do that. Let's say we want to enhance specific items in the scene. These are controls that come from the engine side. Here we've masked, for instance, the pitcher, the grapes, and the bottles. These are implemented on the engine side, and you can mask as you see fit. You can have as many masks as you want. As you can see, we can control the sliders independently. Each item can be adjusted in intensity and tone completely independently from each other. You can completely art direct whatever item in the scene you basically want DLSS 5 to enhance or not enhance. One more example from one of our scenes. Here, we have grouped the bottom part.

As you can see, the whole cutting board and the props at the top. Again, completely masked. The rest of the environment is not being affected. Here we can control. You can see DLSS 5 is adding better reflections. As you can see, better contact shadows onto the ground. For the bottles, you'll see much better translucency in the bottle, better metal reflections. Those are just the elements that we're uplifting without touching the rest of the scene. Again, tone intensity, structure, everything can be tweaked as the developer sees fit. Last but not least, what I want to show you is how DLSS 5 can enhance the whole environment, in this case, the foliage scene. As you can see, this scene looks very nice.

It is a ray traced scene. It still looks a little bit plasticky. When we turn DLSS 5 on, as you can see, the plants come to life and they look really photoreal. That's because DLSS 5 is adding subsurface scattering. That's something that, as you know, traditionally would be incredibly expensive to do in traditional rendering. DLSS 5 is taking the image to the next level. Again, here we could control the intensity that we wish fully. What I hope I showed you today, this is just a small glimpse of the control we have. What I hope you get away with is the fact that DLSS 5 is fully controllable from the developer. We have a lot of controls.

We are building more. We're working with our partners to incorporate their feedback so that DLSS 5 becomes a tool for the artist, the art directors, and the creative directors to basically able to have their vision be reflected into the final frame. Thank you. Thank you, Gaff, for demoing for us.

You just saw the whole idea. The frame really defined what has to remain. Gaff, the artist, directed what could change. Control and artistic direction is a really exciting research for us. This is something we will continue iterate and work on based on the developer feedback that we receive. AI has been applied in real-time rendering for a few years now. There's broadly 2 major categories. The first one is reconstruction. The job of reconstruction is to recover the ground truth reference rendering that was assuming if the render has no resource limit, the perfect render, the perfect image. Current DLSS suite are some of the best small real-time models for that. Then we also have function approximation. It's technologies like Neural Radiance Cache, Neural Materials.

They use tiny neural networks to approximate some far more expensive computation. DLSS 5 is adding a third. It's generation. Right now the model actually changes and generate the final appearance in the image. Artistic direction now becomes the new ground truth. Remember the chart that I showed you earlier? The one with the missing curve? Here it is. Historically, our lever for scaling up the image quality is to scale up the render. More rays, higher resolution textures, more triangles. Those still matter. A better rendering will give DLSS 5 a better grounding to start with. DLSS 5 adds a second lever now. Now if we scale up the model, the image realism and fidelity will also improve.

The more exciting thing here is that even if we don't scale up the model size, improvement on the offline training side, such as better training data, better loss functions, just improvement in the training dynamics, will also translate to a direct improvement in image fidelity at runtime without any additional cost. It's also worth calling out that even though we're at the curve of path tracing right now, every one of the curve previous to that is still alive. We're still writing all of them. DLSS 5 is the same. It extends existing pipeline. It does not replace them. I also want to be clear that we're not claiming we have solved real-time photorealism. True photorealism is still ahead of us.

Right now, the animation does not get uplifted the same degree as appearance, eventually, we want the model to be able to generate more and more realistic images. DLSS 5 really sits at one end of a much wider generative spectrums. It's optimized for games, it's really tightly grounded, it's real-time, it's causal, on the other side, the interactive video foundational models, it's a whole different story. Out there, it gets even trickier on what must stay true. We don't have every answer, but what we do know is that the more the model can offer and can generate, the more control it will have to offer for it to be something useful and valuable. That is wide open research opportunity for the SIGGRAPH community. This curve really is just getting started. We're still on the steep part.

That's DLSS 5, a new level of realism. Simulation defines the world, generation enriches its appearance, artists direct the outcome. Not AI replacing graphics, AI extending graphics the same way programmable shaders and ray tracing has extended graphics before. Every pixel in a game is there still because somebody wants it to be there. DLSS 5, it was built to keep it that way. We're shipping it this fall. All right, I'm going to wrap up. Next, my colleague Neil, who flew all the way here from England, is going to show you more interesting way that AI can help us run simulation, AI physics. Let's welcome Neil Ashton. Cheers.

Great. Thank you. Thank you, Edward. As Jensen said, Leonardo da Vinci was more than just an artist. He was a scientist. He was an engineer that hundreds of years ago tried to understand the fundamental nature of the world around us. Anybody who's seen his illustrations will see how he tried to understand the movement of water and fluids. That's what I wanted to talk about today was about physics-based simulation, about realistic simulations and how we can use AI technologies to enhance and augment them. Where else to start than looking at the world around us? This is one of the largest and most complex simulations of the Earth and the climate ever achieved.

Running on more than 20,000 GPUs at a resolution of one kilometer, a 50 billion cell grid, able to capture the atmosphere, a coupled simulation of the atmosphere, of the land, of the oceans, the most accurate ever achieved, which won the Gordon Bell Prize. It is the epitome of the most accurate physics-based simulation you can get of the Earth. As you well know, these physics-based simulations are incredibly computationally expensive. Whilst we have made great progress in accelerating them, going to the next stage of an even finer resolution, of even more accuracy requires significantly more compute. People know if you have a smaller grid size, that increases the compute. If you want to keep the same numerical stability, the temporal scheme means you have to have a smaller time step.

All of this means that whilst it is absolutely the right thing to try and achieve greater simulation accuracy, there is a logical question which is, can AI somehow help? It is not feasible to be running, with today's technology, a 20,000 GPU run every single day for all of these. Let's look at maybe how AI could augment this. NVIDIA actually has helped to pioneer with many a research institutions, this idea of using AI models trained on simulation data. Just as a large language model will take text and images and tokenize it Since the ChatGPT moment, they increasingly thought, "Well, how can we use this same AI technology for science and for engineering?" It turns out that actually some of the biggest breakthroughs have been in the weather and climate domain.

What this allows you to do is you take this sort of simulation, you can take 40 years of data, this is the ERA5 dataset, 40 years of simulation that has been corrected with observations, train a machine learning model, and then be able to autoregressively predict what is the next hour or day of the weather, but do it in seconds or minutes compared to hours or days. What's remarkable about this, which initially was met with some skepticism, which has been completely now overcome, is this is as accurate or more accurate than traditional simulation approaches. In fact, weather centers around the world use these AI models in production today.

Interestingly, this gap between synthetic data and real data is also something that has been developed, which is that, of course, for the weather, we know that a lot of it comes from satellite data and local observations. This is an example of StormScope, which is one of the most advanced AI models trained on satellite and observation data that can predict storms, how they merge and propagate together using an AI model which runs orders of magnitude faster than simulations and give you more flexibility to train and to test what if scenarios of the weather. I guess, having been in the previous presentation where you see all about the complex geometries, as a fluid mechanics person myself, I say a little bit cheekily that the weather community has it easy because it's a fixed geometry. It's the Earth. It is in some ways much less challenging than trying to model the airflow realistically over an aircraft, or a car, or a data center.

This is the challenge that we have to translate the success of using these AI models in weather and climate into the world that we live in, to be able to design the planes and cars and data centers of the future. What you're seeing here is one of the most accurate and largest simulations of an aircraft ever achieved. This is capturing the fundamental nature of the turbulent flow that Leonardo da Vinci saw in nature through the motion of water around rocks, but couldn't fully understand.

I guess, to put it in simple terms, the challenge that you have is if you're designing an aircraft, and probably some of you came and flew here today like myself, you sometimes look outside, I always like the example of the strake on the side of the engine nacelle. What is that even for? If you couldn't see the fundamental nature of the flow, how could you design it? In fact, that is really the challenge is that air is invisible, isn't it? You can't see what it's doing. You only see its effects. You see it through smoke, maybe. You see it through your breath on a cold morning.

Doing these high-fidelity numerical simulations gives the ability for these aircraft designers to see that, in this case, designing the strake here creates a vortex that propagates downstream, hits the wing, and actually helps its performance when it's taking off and landing. Something that would be very difficult to achieve without this sort of numerical simulation. As you probably see, and particularly those who are working in the movie industry, you obviously want the realism. You want there to be this capability because more and more you want coupled simulations with greater accuracy. The question is, where can AI play a role in that? Just as we've seen the success of weather and climate, can we bring it into your world? Can we bring it into also the science and engineering world?

One of the things that NVIDIA is really focused on is helping to achieve that vision. The first thing to say is that these machine learning models, the data is the key thing here. The data is what you need to generate. You cannot just scrape it off the internet. You need to generate this synthetic data, this physically accurate simulation, to be able to train these models. So NVIDIA has this stack of not just the compute, but things like PhysicsNeMo and Earth-2 that gives you these open source frameworks to develop these models. To, I guess, put it in some numbers, because you may be thinking it sounds a little bit like magic, that this AI model can take this and achieve predictions in seconds that would take hours or days. How can this possibly be done?

It is actually remarkable in that in some ways it's like an amazing data compression. This is an example, actually, of an open dataset that has around 2,000 aircraft simulations just like you saw then, at different geometries, different boundary conditions. And if you run those simulations with more than a billion degrees of freedom, it generates around 200 terabytes of data. Now, what's amazing is the way that the modern architectures work, many of those that have been developed by NVIDIA Research in partnership with probably many institutions here today. Actually, when you take them through these transformers and use some very clever techniques to tokenize this information, you find that the model checkpoint, and this isn't a typo on the slide, and I trained this model myself, so I can guarantee you this is true, is only around 200 megabytes.

It is more than 10,000 times faster to do the prediction of an unseen geometry that the model has not seen, or a boundary condition and achieve accuracy within 1% or 2%. It almost seems like magic, but it is true. And so you have around a million times compression of that 200 terabytes down to 200 megabytes. And in fact, what it allows you to do is this sort of demo that actually looks in some ways like an animation or a movie, but it is physically correct and consistent in that what it is showing is that you can give this trained model, and then visualized using Omniverse libraries, a completely different plane angle or design, and then within a second, only on an RTX 6000, so a relatively modest GPU, it will give you the flow.

Not just that looks pretty, but is physically representative with only around a couple of percent accuracy. The question that I guess I give to this community is, I think you actually have pioneered many of the methods that have enabled this to be true. A lot of the differentiable physics, a lot of the geometry modeling has come out of this community. And what we would love to see is a collaboration across these different communities to apply and try to develop these AI methods that can accelerate the next era of science and engineering. Just as you are experts in modeling the water, smoke, cloth, can you bring that and help us to do this for electromagnetics, for structural simulations of cars and planes?

Can you do this across, which would help to speed up and bring innovation across science and engineering and discover potentially brand-new designs that could never be achieved before. In fact, actually, for me, what I find most interesting is how this converges and leads very nicely to the next topic of how can we understand how these objects, these planes, these cars, interact with the world around us. Can we build a world foundation model that can learn this and allow us to be able to train the next generation, for example, of autonomous vehicles? Who better to talk about this than Ming Liu, the VP of the Cosmos Lab at NVIDIA, who's going to come and tell you all about it. Please welcome him on stage.

Thank you. Hello, SIGGRAPH. Glad to be back. Physical AI has a data problem. Robots need to learn from the real world, but the real world only run in real-time. You cannot fast-forward the reality. Let's look at different data sources for physical AI. We have tele-op data give you real demonstration. They are highly accurate, but painfully slow to collect. Classical simulation help you scale, but only the scenarios you can hand-build. The internet see the world from human eyes, not robots. That's why world foundation model change the equation. If you have a model that can generate the world based on robot's perspective, data stop being something you collect. It becomes something you compute. Every GPU cycle turn compute into new scenarios, new variation, new test cases that no fleet can collect. That's why compute is data.

That's why world foundation model is the data engine of physical AI. With world foundation models, robot capability scale with compute. Cosmos is NVIDIA's world foundation model. We built Cosmos to help physical AI developers build more generalizable physical AI systems. We built Cosmos to help developers in three ways: better data, better environments, and better starting points. Better data. Physics-correct simulation generation and video analytics. Better environments. A 1,000 kitchens, a 1,000 warehouses, a 1,000 weather conditions before your robot hit the real world. Better starting points. Cosmos are strong backbone to build policy models. You don't start from zero. Data, environments, starting points, all in one model. Let me show you the Cosmos video.

The real world is infinite and unpredictable. Physical AI needs data, but real-world data is impossible to scale. For physical AI, compute is data. This is NVIDIA Cosmos, the frontier of physical AI. An open omnimodel built on a new mixture of Transformers architecture. Pixels, action, sound, and language flow into the auto-regressive Transformer, which reasons, plans, and instructs the Diffusion Transformer, which generates what comes next. Developers post-train Cosmos across embodiments and use cases. As a VLM, Cosmos watches the physical world, understands what's happening, describing scenes, and flagging what matters. As a world model, Cosmos generates controllable, physics-aware synthetic video from text, images, video, and sound. When action-conditioned, Cosmos is a simulator, inferring actions and generating the next plausible futures, closing the loop for policy training and evaluation. Post-train Cosmos it becomes a world action model.

Perceiving, reasoning, planning, generating actions for any embodiment in any environment for everything that moves. A new kind of data, a new kind of teacher, generated by compute. NVIDIA Cosmos, the open foundation model for developers of the age of physical AI.

What makes a foundation model a foundation model is its capability to consume enormous amounts of diverse data. When the model is trained with data from different tasks, something powerful happens. It learns the synergy. We train Cosmos with data from diverse physical AI tasks, including world understanding, prediction, simulation, and action. We use one backbone to solve all the different tasks. As neural networks learn internal representation to solve the task, you might wonder why one representation can work for all the different tasks. The answer is simple. We live in one world governed by the same physics. The data for different physical AI tasks come from the same physics. That's why one shared representation works. What's even better, with this shared representation, one capability helps others stronger. One world, one shared representation. How do we learn the shared representation?

We use a mixture of transformer architecture. As shown in the slide, the architecture has two towers. On the left is the autoregressive tower. It takes a vision token, task token in, and predicts task tokens one at a time. It does the causal attention. The autoregressive tower is where the model reasons. It thinks in language to explain what it sees. On the right-hand side is the diffusion tower. This is where the model generates. It generates vision tokens, audio tokens, and action tokens. Every token in the diffusion tower attends to all other tokens in the diffusion tower, as well as the tokens in the autoregressive tower. This gives global understanding so that the generation output is globally coherent. You have beautiful simulation results, physically correct. Each tower has its own layer norm and MLP.

The two towers share something critical in common, which is the multimodal attention. It is through the multimodal attention we align language, vision, audio, and action. Two towers, one shared representation. Action is the first-class modality in Cosmos. Whose action? A car's steer, an arm reach, a hand grasp. Every embodiment speaks a different language. Our solution? To build a common vocabulary. We build a common vocabulary using a set of geometric primitives. Ego pose, where the body is, where the head, to what. Effector poses, where the hands are, the grippers are. Grasp and gripper, what the hands and gripper are doing. With this, every embodiment becomes a combination of these primitives. Self-driving car, ego pose alone. Single-arm robot, one effector, one gripper. Humanoid robot, one ego pose, two effectors, two graspers. Even human egocentric video falls in the same scheme.

One ego pose, two effectors, two hands. This turns every human egocentric video into valuable training data for robot learning. These primitives are purely geometric. They are grounded by the visual structure of the world, connecting pixels to motion, space, and control. This is why one world model can work for all the robots. Different bodies, one action language. From here, what the model learns transfers across the embodiment. Cosmos models action and world states jointly. Action falls in two directions. A Cosmos model out of the box supports three important physical AI modelings. Forward dynamics, the action in what the future states out. If the agent takes an action, what will be the result? The model plays the consequences before they happen. Inverse dynamics. Visual transition in and action out. Which action caused what I just saw?

This model is important because it turns every video into action labeled training data valuable for robot learning. Policy. Visual state in, what should the action be and what will be the outcome? What should the robot do to complete a task, and what will be the visual consequences? This is a key of physical AI. Forward dynamic effect. Inverse dynamic cause and policy. Three important physical AI modelings in one model. Cosmos models are shipped with frontier world knowledge. With Cosmos, you never start from zero, you start from Cosmos. You post-train Cosmos in two directions. The first one is accuracy. Post-train Cosmos with your own data from your own embodiment. The graph is the story. The more custom data you put in, the better the data model. This is not just a theory.

We actually pre-train Cosmos on DROID dataset and achieve state-of-the-art pick-and-place policy result. The other direction is speed. Inference costs matter. We pre-train Cosmos 3 Super with distribution matching to reduce the number of denoising steps to four. This leads to 25 times speed-up for text-to-image and image-to-video generation results, and they are still state of the art according to Artificial Analysis. You can make generation 25 faster with distribution matching. One more important is that we are releasing all this post-training script, the one we used to build custom model, as a ready model so you can build your own. Both accuracy improvement post-training and speed improvement post-training can be done with a DGX station. What I presented you comes with a family of open models of different sizes for different deployment use cases. Last month in NVIDIA GTC Taipei, we released Cosmos 3 Nano and Cosmos 3 Super.

Nano is a 16 billion parameter model. It's a great starting point for every physical AI team. It can be easily post-trained for your downstream application. Super comes with frontier world fidelity. When accuracy is the primary concern, you start with Super. Today we are releasing the last member of this family, Cosmos 3 Edge. Cosmos 3 Edge is a four billion parameter model. We built it for a single purpose: run real-time on devices. That robot policy and video analytics can happen in Jetson Thor, in RTX, in DGX Spark. One model lives where robot lives. No round trip to data center. Edge for the robot, Nano for building, and Super for frontier. Let me show you a demo we did with Cosmos 3 Edge policy. We connect a robot arm and a camera to Jetson Thor, the gray box on the table.

The box runs Cosmos 3 Edge policy. The video feeds in, the policy model analyzes the video and predicts the action, and the action controls the robot arm to achieve the task. It's fast. It's fast enough for real-time control. This is what physical AI on device looks like. Cosmos is not a one-time act. It's a platform that has been compounding. We introduced Cosmos to the world in January 2025, 18 months ago. From the arc, you can see we have been building component by component. We have Cosmos Predict for world generation. We have Cosmos Reason for world understanding. We have Cosmos Policy for action. In Cosmos 3, we unify all these Cosmos models into a single omni-modal world model. It can do everything the previous version can do in one body. This is where we are today, and we are not stopping here.

We have started building Cosmos 3.5, which is going to bring in multi-view and other capabilities. Cosmos is an open platform. In addition to the open models, we also have open datasets, open libraries, and the open framework. We provide support script to help people customize Cosmos model, as well as build your own world models. Cosmos is the frontier of physical AI. Not just we say it. Across eight different benchmarks, spanning generation, reasoning, and policy, Cosmos model rank number one. One more announcement. Building on top OmniDreams research, we are announcing Cosmos Dreams. Cosmos Dreams are neural closed-loop simulators. The first one we built is for autonomous vehicle. As a diagram shown here, you have a policy model, Optane model here. Take an action. Cosmos Dreams will add the action and generate what the sensors will see. Photorealistic, frame by frame, condition on the actions.

It's fast. It's at interactive speed. Your car basically drive in the imaginations of Cosmos Dreams. Let us show you a live demo. Andy, are you ready? I'm ready, Ming-Yu.

Thank you so much. I am so excited to show the audience our demo here. Cosmos Dreams here is simulating an entire world from a single frame. For this demo experience, we're using a relatively sunny day, suburban neighborhood that you see everywhere, with few cars, some lanes, things like that. You'll also see a bird's eye view on the bottom right-hand side, that helps with the simulation to be accurate. After the first frame, everything is completely generated from what you expect from the world, and also the inputs from my PS5 controller right here. What is incredible is this demo here, and this generation, is being done on a single RTX 6000 Ada Generation workstation GPU. When we first started creating this demo, however, we needed to use 64 GB300 GPUs to bring this real-time demo to life.

We were able to drive this real-time simulation with 16 Vera Rubin GPUs. What an improvement. However, now, with the new Cosmos model that Ming-Yu was showing earlier, it is on a single RTX 6000 Ada Generation workstation GPU that you can see right here on my table.

Fantastic. Now every autonomous vehicle developers can do everything simulation. They can use Cosmos Dreams to do policy verification. They can know their accuracy before deploy the model in the real fleet. They can also use Cosmos Dreams to train their models, generating scenarios difficult to craft in the real world. Is that correct? Absolutely. This accelerates the development velocity and saves a bunch on their operational costs.

Awesome. Thank you, Andy, and the team. Remember our promises. We built Cosmos to help physical AI developers in three ways: better data, better environment, and better starting points. It's already happening. As a media, Metropolis VSS, Video Summarizing Search, leverage Cosmos for video analytics. NVIDIA Isaac leverage Cosmos for realism enhancement and variation creation for robot policy training. Cosmos Dreams, closed-loop simulator built on top of Cosmos. Optane model, NVIDIA's flagship robot taxi driving policy, built on top of Cosmos. GR00T, NVIDIA's flagship humanoid policy, built on top of Cosmos. This is the most important slide of my part. We are building Cosmos platform with Cosmos Coalition members. They are contributing research, data, and evaluation to help us advance Cosmos platform. Cosmos have been adopted by physical AI builders, as shown in the second row, and there are more. It has reached millions of developers.

Join us to accelerate the physical AI revolution. We have a amazing program lined up this week at SIGGRAPH, covering topics like USD, simulation, and AI. Here is the QR code. Hope to see you there.