Think You Know AI? 25 Startups Prove You Wrong

Welcome, welcome, everybody. Should be a fun one. Let me start with a question. When I say AI, what's the first thing that comes to your mind? For most of you in the audience, it's probably your chat bot, or your favourite AI coding assistant, or one of the many AI agents you have working on your productivity. These days, it might even be OpenClaw or NemoClaw. And these tools are great, right? The sheer pace at which these tools have become part of our everyday lives has been absolutely incredible. They're always in the news and rightly so. But today, in this session, we're going to talk about something else. We're going to talk about AI in the physical world. Thanks, Joshi. So, you know, the trend of the 20th century, right, was just this massive expansion of the world, right? And in the 21st century, as we see AI start to bring us together, start to connect more of that world, we've seen its ability to interact with a huge array of problems, right, to solve things that we didn't think were possible without it. And so one of the things that we're gonna take you through today is some major areas of the human experience, right? Major areas of industry, of commerce, of health, of other very deeply personal human aspects of the world. And so I think what you'll see as we go through here is that AI has not stopped in all of those areas, right? And that hopefully this is some inspiration for what is possible with AI and what startups are doing for real in the world. So Chris just gave you a map of what we're gonna talk about today. Let's make it more concrete. We're gonna talk about 25 startups across six teams. These startups are from all over the world. They're at various funding stages, different founder backgrounds, but all of them are solving hard problems in the physical world. All of them are building on NVIDIA's tech stack, whether it's at the edge or in the cloud. Now, we have 25 startups to cover in roughly 30-ish minutes, so things are going to move fast, no doubt. You can think of this more as a guided tour of what these startups are working on to build awareness than a proper deep dive. So with that, let's get started. Let's talk about agriculture. Agriculture is one of the most important and oldest human industries. In the 20th century, We produced more food than ever before. As our population grew enormously, agriculture expanded too. But that expansion came with challenges. Soil depletion, heavy chemical use, water wastage. And now, in the 21st century, as our population has grown to 8 billion, we deal with the same challenges. In fact, The World Food Programme says that AI and ML offer unprecedented opportunities to accelerate our progress towards achieving zero hunger. So in the 21st century, as our population grows even more, how do we feed 10 billion people while also solving world hunger? And at the same time, mitigating the challenges I just talked about. Let's see how AI can help us do that. So I love this one. Look, you know, we talk about deep learning all the time, right, to the foundation of modern AI. This is gonna be shallow learning today because that's all the time we got for each of these startups. I love this kind of the mission of this startup, right? I didn't think about, before meeting this mission here, I hadn't really thought, frankly, about what it takes to make sure that pollinators can have healthy lives, right, healthy colonies. Some of you have maybe heard of the big challenges in colony collapse among pollinators in agriculture, right? And so it turns out that one solution to that is to give bees luxury apartments. And that's barely even a joke, right? This is, these boxes that you can see here are fundamentally the kind of automated system for holistic health management for bees. Using edge computing, computer vision, using all kinds of sensors to determine if the bees are happy, if they're healthy, if they've got parasites. And being able to have that kind of always-on monitoring enables a much larger, to your point, much larger coverage of the agriculture than if we had to, you know, humans drive out and inspect every one individually. So you can have that kind of persistent measurement of what's going on there. So happy bees for our agriculture. You're going to see in the next two slides here a couple of different ways that robots are playing a major role in advanced agriculture. So this company in particular, by doing edge computing, we spend a lot of time these days, like you said, Joshi, thinking about cloud and really massive, that was a big subject That's the concept of Jensen's keynote. is how big of scale we can go. We're also kind of thinking at NVIDIA about how small can we go, right? How much can we put at the edge with embedded compute? And so by having the compute at the edge flexible enough, right? Not just single function chips, but flexible enough because you have GPUs to be able to then adjust to different kinds of model depending on the need. Make sense? So that captures the traditional methods for agriculture, weeding and spraying, right? But also, again, thinking of everything as a sensor, right? Thinking of everything as a way to continually improve and grow the foundations of agriculture. And in a particular kind of sense, right, one of the challenges, like we said, is how do we provide nutrition for the billions of people that we have to? And if we were stuck waiting on agriculture that is appropriate to the climate of everywhere individually in the world, that would be a much bigger challenge. Here we're seeing with greenhouses, and particularly with automatic greenhouses, robot-assisted greenhouses, we can grow healthier, more robust food In places that it might not be able to otherwise. And so that's a huge change, kind of a step change, in our ability to provide a variety of nutrition all the time for populations around the world. And then let's zoom in on a really specific part of the production chain here, right? So sustainability, like you said, Joshi, is a huge part of how we're gonna feed 10 billion people. And so thinking beyond just kind of the normal weeding and spraying, thinking to how do we do that without So, on the left you can see here a company that's using mechanical hose, slamming this kind of metal implement into the dirt and pulling out the weeds as it's going by by using vision at the edge, right, to detect when there needs to be a weed literally physically pulled out of the ground. Solar-powered, actually even wind-powered and using all of that technology to become kind of autonomous at pulling out these weeds. On the other hand, on the right side here, you can see what happens when we put lasers on similar machines, right? When you can have, at the edge, micro-focused lasers killing weeds with energy, right, instead of with chemicals. Pretty cool to see how that all enables scale in a way that wasn't possible before. Yeah, same problem, different methods. Absolutely. We just talked about farming and agriculture. Let's talk about protecting communities in the world. Think about the last wildfire you saw in the news. By the time it made the headlines and people found out, it was too late. A lot of damage had already been done. Neighborhoods were wiped out. And think about the world we live in today, right? The population is increasing, our cities are getting more dense and complex, the weather's unpredictable, and so we're at risk more so than ever for these threats. And it's not just wildfires, it's storms, floods, other urban emergencies, right? These are not rare events anymore, right? In fact, the UNEP says that extreme wildfires are projected to increase by up to 50% by the end of the century. Think about that, up to 50%. So these problems are gonna get harder, not easier. Now, in most cases, The early signs are there. We just don't detect them early enough to do something about them. And therein lies the challenge, right? How do we shorten the gap between when one of these threat occurs and we can do something about it? Let's see how AI helps us do that. So it turns out that there are signals of potential wildfires in lots of different places. In some cases, the signal is very much at the ground level, especially the micro signals, being able to see kind of individual patches of a forest burning. Or starting to burn, right? And so that's what we're seeing on the left here, is how do we put ground-based sensors empowered by compute at the edge, empowered by models that are driven by training, not at the edge, right? And then use that to shorten the time between occurrence of some threat, right? And detection of that threat. Be able to get out ahead of it. And it turns out that the other direction here is that one, if you think of one sensor as looking horizontally, like looking out over a forest, over can be a very different word if we say over in a satellite sense. And so it turns out you can put GPUs on satellites. It's a pretty good place to do your advanced computing so you're not beaming down all of your images at scale, these huge images that if you want to cover huge tracts of land. It's much more efficient if you can process it on board the satellite and beam down just the detections or the zoomed in parts of what you need to be looking at. So having these two very different approaches to what is fundamentally a major similar problem. Is a huge part of conceptually how we should think about what startups are able to bring to AI, right? That it's not about one solution to a problem, right? It's kind of, it's about a thousand solutions to the problem. And then figuring out which of them is most successful in doing the thing that we need it to do. There's a kind of interesting follow-up to that, right? If you think of the left side here, maybe even the right side as adding new sensors, right, into the world. I love when startups, you know, have a major part of their business being collecting more data about the physical world. And so that's where, you know, it's not enough just to have kind of a sensor, right, or to rely on legacy sensors that weren't, we should remember, weren't laid out for AI processing, right? Traditional weather sensors weren't laid out for how do I maximize the fidelity of my AI models. But if you start from the premise that I need to make as good of a detection chance as I can have for a particular kind of AI, then putting up radars might make sense for you. That's what we're seeing on the left with climate vision here, literally putting up a radar network kind of across the south-southeastern United States that then helps augment places where otherwise traditional weather forecasting methods may have had gaps or been less able to predict. The kinds of disasters that threaten particularly kind of heavy population centers, right? The kind of things that, you know, have caused so much damage as extreme climate has proliferated as you described from the UN. Another way to think of this, though, is to go from sensor processing. How do we, instead of just taking images or from traditional sensors, how do we really get a detailed picture? In some sense, this is about how do we shape our understanding of the world? The other side of this, on my right here, is how do we, instead of going from traffic cameras to a 3D understanding of what's happening in our cities? To be able to say, hey, this problem isn't just about counting cars but about being able to track flows of people and of things through cities so that we can treat them as integrated systems rather than as individual points on a map. And so it's really cool to think about it from, again, two very different angles on a similar problem, right? If we have challenges affecting our cities, adding new sensors and interpreting those sensors in different ways really provides some powerful leverage to adapt to those problems. So we just talked about protecting communities, and the key there was early detection, right? The theme was urgency. Let's talk about the environment where problems build up gradually, slowly, right? When you think about waste management systems, recycling, water systems, forests, the problems in these areas are slow and gradual, and we don't often find out about them until it's too late, right? Let's go to the board. But the problem is that if we cannot see these problems, we cannot measure them, right? We don't do a good job measuring them. And as the UN says, 68% of environment-related sustainable development goals lack sufficient data to assess progress. Let's think about that. We've set our goals, for 68% of them, we don't even have the data to measure and see if we're making progress, right? So we're just flying blind. And so the opportunity here is how can we use AI to get more visibility into the system? So one, we can start measuring them, and then two, we can start tracking and see if we're making progress against our goals. Let's look at a few examples. So look, it is very easy to get excited about the shiniest version of AI, right? The version of AI that talks about building the sci-fi glass and steel version of the future, right? But I'll be honest, I get just as excited about the really low level micro, you know, it's there and we try not to think about it, parts of our world. And that's what we're seeing here with these companies. They're dealing with our waste streams. In some sense, the challenges that we're solving by having denser cities, smaller transportation costs, more interconnected communities, they come at a different kind of cost. Now we have to deal with the waste or the recycling that we need to do from those human populations. And so if we can instead apply AI to Analyzing, getting data on, as you can see on the left here, scanning to understand what's in our waste flows. So that instead of just saying there is a lot of it, we can do something with that. We can turn this into the kind of optimizable cycle that we're so used to in tech and bring that to industries that have lots and lots of potential for really applying 21st century tools to these problems. So once you have the data, just like you were saying, once we have the data, then we have the ability to make progress on these goals. And so that lets us then, for instance, add robots into the process to help us sort and pick from those waste streams, to bring out of those waste streams things that don't have to be there, again, at scale. Right, not just, it's not just kind of another dirty job. This is something that now we can kind of solve problems at city scales, which is not trivial, but it's exciting that, you know, startups are taking that on. And it's really exciting to me personally, that the market has appreciated that, right? That these startups are able to access the kinds of people and technologies that they need to make this meaningful contribution to the world. If we go kind of hard the other direction, you think about, okay, that was, that's thinking about kind of urban environments, right? If we're reading the signals of the planet on a more geophysical scale, right? Looking at systems that again, you know, on this theme of what sensors do we need? To tell if we are making progress, and if we are, how much progress towards our macro goals as a society. Putting up weather balloons, again, it's not the kind of thing that's like, you know, making the shiniest possible headlines. It's super important. Right, otherwise we can do all of the things that we want in the world, but we don't know if they're having the real impact. And I care a great deal about that. The other side of this, if we think of collection as the left side, we can think of on the right, how do we make the data that we collect accessible, a common operating picture of the planet? But that's not easy. There have been lots of cases where people have started trying, but bringing that together in a way that we can have, you know, that people will have a common operating picture of their business, right, or an industry, if we take that same mindset and we're able to say, hey, it's just as important to have a common understanding of the planet, that we bring data together, make that data available to... Consumers of it downstream in a way that has not necessarily been available and to use AI to both accelerate and improve the precision of that kind of processing so that then it is available at planet scale. Let's talk about something that everything else depends on, the power grid. Now, most of us don't think about the power grid until our lights go out, right? But it is something that everything else depends on, hospitals, transportation systems, communication, and of course, data centers. But the grid was built decades ago. It was built for another world. It was built for a world where power flew in one direction, from the utility plant to people's homes and businesses. Think about how that's changed now. We now have solar rooftops pushing electricity back into the grid. We have EVs charging and discharging rapidly at unpredictable times. Now all of this, plus the fact that one, the grid is old, and two, it has to deal with extreme weather means that there is more strain than ever on our grid. And inevitably, when the grid fails, not everyone is impacted equally. The Council on Foreign Relations says that when severe storms strike, lower income communities wait the longest for power to return. Just think about that. The people who need power the most have to wait the longest to get it back. Now, one solution is to just rebuild the grid from scratch, but that's not practical. So how can AI help us make our grid more resilient, flexible, and smart? Let's look at a few examples. Yeah, absolutely. You know, look, one of the challenges of living in an environment where we are sharing the planet, thankfully, with lots of plants is that they don't always have a purely positive impact on your access to electricity, right? If you've had a tree or branch fall on power lines, you've certainly experienced what I'm talking about. And so it turns out that there are, of course, startups dealing with those problems too. Here, the trick is satellite imagery, right? Being able to look down at a fine-grained understanding of the local environments, figure out that there is foliage putting at risk parts of the grid. And then be able to proactively address that in a way that so far has very much been, like you said, unevenly distributed, you know, human-based patrols looking for those kinds of risks. Being able to do that again at scale all the time on kind of a persistent defense strategy is a huge part of how we make grids more resilient with the systems that we already have, right, without doing a wholesale kind of ground-up revision of the grids. If one version of this is my usual story, hopefully you don't get too tired of we need more sensors, more data processing, and then more action based on the data, that's kind of the theme of how AI works. Here's a specific kind, a flavor that we haven't talked about yet but is very cool and is something that NVIDIA has played a significant role in, digital twins. Being able to have in the computer a kind of physically driven, sensor-informed model of how the grid works, being able to use that model to simulate different kinds of stress on it, gives us the ability to, again, proactively kind of move forward the action point and not just be waiting for it to fail and responding to that, right? Because if we do the latter, again, you get these kind of inequities in how... in how those resources are applied, where if we can get out ahead of things, we have the ability to prevent rather than react. And then, you know, like Joshi said, right, if we say, hey, the grid was for 100 years, 150 years, has been electricity coming from generation sources and flowing out to people, right, if we invert that paradigm and we think, hey, there are lots of generation sources now, there are lots of, we kind of have to do this two-way street, and that just, it doesn't just, you know, increase the complexity, right, it kind of squares the complexity. Because now everybody is a consumer and a producer in this simplified model of the world. And so now you've got all of the flows going from everywhere to everywhere. It's much more complicated. And so it doesn't really work with the old models. And so then having particularly sensors at the edge that let us have an always-on understanding of the whole back and forth of the grid. As we're preparing for data centers and the impact, as you described, that they're having on the environment, to use AI to mitigate the impact of AI. And that's what we're seeing on the right side here, that the only, in some sense, the only solution that we're going to have for how do we create a sustainable growth pattern in these industrial areas is going to be the kind of modeling, the kind of data analytics, the kind of forecasting, the kind of simulation, and then building around that that is made possible by AI. For the new grid, for a grid that is people-centric, you have to really change the origins of where you're thinking. All right, we're halfway there. So far, we've been talking about big systems, like planet-level stuff. We've talked about farms, we've talked about cities, talked about the environment, we've talked about the power grid, of course. Let's zoom all the way in down to an individual. Now, some of the most important moments in people's lives also happen to be the most private and quiet. Let me give you an example. There's probably an elderly person right now that might have a fall. But because they live alone, it might be hours before someone finds out, right? Because these moments are private and quiet, not everybody's watching, right? Not everyone has access to a caregiver. Not everyone has access to a specialist. Not everyone has someone or something watching them at those important moments when they need it the most. And I think this is where AI can help, right? The ITU Focus Group says that AI-based technologies hold great potential in improving the accessibility, quality, and value of healthcare outcomes. In this case, this is as much an accessibility problem as much as a tech problem, right? How can AI help us reach more people in the moments they need it the most? Let's look at a few examples. One of the biggest challenges worldwide in health is vision impairment. There are low cost interventions, right? Glasses that still need to be proliferated around the world. There are sophisticated interventions, right? Dogs as companions for visually impaired folks, but those are frankly quite expensive and very unevenly distributed. And so if we can bring down the cost, if we can bring in the learning from people around the world as to what they need to navigate their environments, we have the ability now to build technologies that, like this one, give freedom back to people who need it, right? This headset here has sensors, onboard processing, again, that... Let you see without sight, right? They'll let you see in the same way that cars do and so you can you can see the the kind of demonstration here is being Able to reach out and grab an apple With no visual data going to your eyes. That's, it's remarkable. And again, I think the thing that I'd hopefully leave you with here is that the, making that available, making that accessible, making that widely distributed, right, is in some sense the thing that is most important to me about how these technologies are changing lives. It's less directly about how cool it is, and it is very cool, but it's really about how this can be replicated more easily than many of the alternatives have been historically. If we think, you know, I kind of heard the other direction, right, if this is let's make everybody's day to day more accessible, right, in the worst and the hardest and the most stressful moments, like you're saying, we also need people watching our backs. And emergency first responders need people watching their backs as they try to provide care under, again, some of the hardest times that you can experience. And the last thing we want to do is have people, you know, making their best judgments under, you know, significant stress and having to monitor lots of different things all at the same time. And so one of the things that AI can do, it turns out, is understand how, what people are hearing on emergency dispatcher calls, right, 911 calls here in the U.S., how people's voice patterns and stresses and other signals can tell us about their individual situation. Right, and take off some of that cognitive load from the dispatchers and the clinicians who are being asked to provide diagnoses and response while still keeping them in charge, right, acting as kind of a supplement to their expert judgment rather than something else. We also, as Josie said, in kind of the quietest moments, in the moments where you least want people constantly monitoring, to have AI be able to provide that impersonal privacy-protecting support. Being able to say, hey, this senior needs help, without having to have a person monitoring that. To be able to understand early in your vocal patterns, if there's something cognitively that is impairing or an early sign of something happening in your body, incredibly important as we're going to deal with the challenges of aging populations and in equal access to healthcare. And so again, in all of these cases, the AI is serving the role of spreading and of making broader the kinds of health care that otherwise would be expensive and limited. I'm excited for that world, and I know a lot of people very close to me who are benefiting that way. Right, in all the sections we've covered so far, the team has been seeing things early and more clearly. That's true in this one as well. Let's talk about science and medicine, where some of the most important decisions also happen at the very beginning. Now, we have the most amazing people working in science, but science itself has gotten too complex. 200 years ago, you could argue that one smart individual could probably keep track with everything going on in the field of science. But that's just not true anymore. We've made so much progress, and there's so much information out there, right? It's impossible for one person to keep track. We now have thousands of teams working on tens of thousands of projects, all solving small parts of the big puzzle that is science. Now, if you think about it, before a trial or a treatment begins, these smart people have to make some early but consequential decisions. Which path to take for a given trial? Which direction to take for a treatment? And these early decisions ends up shaping years of research after. The challenge is that these folks are making these decisions often without having the full picture. Therein lies the opportunity, right? AI is enabling the analysis of extensive data sets and also helping uncover hidden patterns. So if you think about it, if we can augment our amazing scientists and researchers with the complete information, if we can augment their intuition with the full picture, right at the beginning when those decisions really matter, then we can accelerate the pace of discovery. Let's look at a couple of AI startups that are helping us do that. You know, like you said Joshi, it's really challenging, right, when you have to pick paths without being able to, for very good, in this case, privacy protective reasons, right, that aggregating kind of all of the healthcare data of patients with a particular condition or a particular family of research, that would be enormously helpful to the model, but a significant risk. And so, one of the things that is, I think, underappreciated about the modern kind of deep learning training solutions is what we, you'll forgive the jargon here, but federated learning, where I can take a model and move it to the hospital, right? Move it to where the patient data is stored, rather than moving the data to where the model training is happening, right? And so that means that every medical facility, all of these protected enclaves of data, the model comes in, gets a little bit of training, and then moves on. And then if you do that, turns out you're training on the data. You are still building the model. You're just doing that in a way that then doesn't aggregate everybody's medical and health information to one place. So that's really exciting, right? You can facilitate, you can power a whole new generation of models without having the kind of risks and concerns that you might have had otherwise. But then you can build on top of that, right? You can say, okay, we're going to simulate We're going to, just like we were with the kind of digital twins for the power grid, we can simulate people's biology and, as the caption here observes, test drive, right? Try out a treatment in this model, in silicon, right? In a way that then gets to that high leverage point at the very beginning of treatment. Or if we're designing new ways to test, as you can see on the right here, design new ways to test and pipeline through interventions without having to limit ourselves to traditional ways of assessing the data that you're getting, right? Be able to test them on realistic surrogates, analogs, for the actual place that you want to deploy a therapy. And so being able to, again, create from the ground up these very novel places to make the decisions, assisted by AI, that's an incredibly powerful way to do medicine. I'm excited for that. And then here I'll give you kind of the In some sense, if I've got drugs, pharmaceuticals, and I've got my genes, the hope is that I can have some kind of personalized therapy that goes along with that. If you think of going from drugs to outcomes, being able to inject the response that my particular genetic makeup will have to those therapies is incredibly powerful. And it's something that, when we talk about personalized medicine, right, there are lots of ways to personalize medicine, but one of the most fundamental is to understand how I might personally respond to something, to a particular treatment. Again, being able to intervene at that high leverage point early on and not have to wait until I'm seeing the reaction as a patient. All right, so everything you saw here today is about how AI is helping solve hard problems in the physical world. We just covered a lot. We just covered 25 startups across six teams. Now, I don't wanna recap everything, but I did wanna leave you with a few, right? You saw AI in farming, how it is helping farmers treat at the individual plant level instead of the farm level. We saw satellites detecting wildfires from space in a matter of minutes. We saw robots helping sort recyclables so less of it ends up in landfill. We saw how AI can help us create a digital twin of our power grid so we can spot weaknesses and be better prepared for emergencies. We saw a pair of glasses that can help the visually impaired navigate the world independently. We saw AI can help us create a digital tumor to help doctors test treatments or test which treatment would work for a patient before the actual treatment began. And like I said before, all of these companies are solving hard problems and all of them are building an NVIDIA stack. The other thing that's common for all 25 is they're all part of NVIDIA's inception program. It's how we support these startups by giving them access to free compute, technical resources, go-to-market support, connecting them with the broader NVIDIA ecosystem. We help them from zero to one and then one to 100. We are there throughout the way. Best of all, it's free. We work with startups in all industries, Including the non-obvious one, as you saw today. And we have startups from over 120 countries that are in Inception. In fact, some of the startups that you saw today are here at GTC. If you wanna learn more about Inception or wanna learn more about the startups that we talked about today, please come talk to us at the Inception info desk in the expo hall. And then lastly, if you're building a startup or you know someone that's building a startup, please join Inception. The link is right there. We'd love to work with you and be supportive of your journey. Thank you. Thank you so much.