NVIDIA GTC 2026 Live

Live from the SAP Center in San Jose, this is NVIDIA GTC Live. Over the coming days, the world's leading developers, engineers, scientists, and founders will gather here at the center of the AI revolution. For the next 3 hours, we'll hear from the pioneers, the architects, and the builders transforming every industry. From robotics and research to global AI platforms, every layer of the computing world is being reinvented. All leading up to the moment we've all been waiting for. NVIDIA founder and CEO Jensen Huang takes the stage to define what comes next.

Welcome to GTC. So settle in. You're not going to want to miss a moment. Nvidia GTC Live starts now.

Welcome to GTC Live. We are kicking off Nvidia's 4-day GTC conference where developers, researchers, startups, and industry leaders gather to explore the latest advances in artificial intelligence and accelerated computing. Everyone's starting to arrive for Jensen Huang's keynote address. But first, we are here to bring you the pregame show.

We have the key builders of the age of AI here with us today. Get ready for a lively and in-depth conversation about the technology shaping the AI era. Creating intelligent agents, robotics, autonomous systems, and the infrastructure around the world to power it all.

I'm Sarah Guo, founder and managing partner of Conviction, an AI native venture firm where we partner with founders building the next generation of intelligent software companies.

And I'm Gavin Baker, founder and managing partner at Atreides Management, a crossover fund investing in high growth tech and consumer companies with a focus on AI semis and other emerging technologies.

AI isn't just a feature of the global economy. It's the engine reshaping how we build, explore, discover, and manufacture on a planetary scale.

AI has moved from experimental to essential. It's rewriting the rules of industry, redefining security, accelerating science, and retooling manufacturing, all in real time.

This show marks the one-year anniversary of the first GTC pregame show. Last year, the pregame show was a small production tucked into a corner of a balcony here at the SAP Center. But since then, GTC has traveled the world, Paris, Taipei, Washington DC, and back again today to San Jose, where it's literally bigger than ever. Visionary CEOs, founders, and scientists are here to build the future, connect with each other and you. And there's one common thread. It's all accelerated by NVIDIA.

The GTC pregame will bring you five panel discussions with tech's biggest names covering accelerated computing, how AI is becoming essential infrastructure, how open models power the ecosystem, how Agentic AI is at an inflection point, and finally, how AI is entering the physical world. No question, this is the place to be if you want to know where artificial intelligence is headed.

Just how huge has the GTC pregame become. What better way to show you than from the sky?

And how awesome is it that we are outside this year? Sarah,

It's amazing. And you know, this stadium usually houses tens of thousands of Sharks fans screaming about hockey. Now they're all equally at the same level of energy about token throughput, compute access. I love it.

Same level of energy. I mean, this is, I feel like this is next level. All due respect to the San Jose Sharks.

I'm sorry. Better level of energy.

Yeah, we have an arena filled with token maxers.

And what an amazing production. To help us cover it all is our special GTC correspondent, Tiffany Jansen, who's out there in arena green. Tiffany.

Thanks, Sarah. Hi, everyone. I am here in the pregame lounge, and can we just take a moment to appreciate this? I mean, first of all, look at all these awesome couches, chairs, coffee behind me. Not necessarily needed, though, because the energy here already is incredible. Now, I got here, I think it was around 7:00 a.m. this morning, and there was already a line of people waiting to get in, which is incredible. And now you can see the line behind me. It's even bigger and it's going to continue to just get bigger and bigger before the keynote, which speaking of attendees, you know, they come from all over the world. I go to a lot of different conferences and events, but I can tell you with certainty, this is the biggest. People come from everywhere. When you go into a room, the energy is just electric. You hear people talking about different business deals, future of technology, starting companies. I mean, it's you can't help but feel so inspired. Now, today I'm going to be taking you along for some really incredible conversations with some of the biggest names in tech. We are going to be talking about everything from agentic AI, physical AI, which by the way is a topic I am very excited about right now. I mean, over the past year, the idea of a humanoid in my house being able to do laundry for me, I need that really badly. We're going to be talking about everything from energy, autonomous vehicles, and so much more. Now, these conversations are going to cover not only what these thought leaders, these tech leaders are focused on right now, but what they are thinking about when it comes to the future of technology. We see it moving so quickly. Last year at GTC, I remember thinking to myself, there's no way Nvidia can outdo themselves. But this year, they have. The energy here. One thing I really noticed is last year the energy was more so okay what's next, what's coming up. Now this year it's more so where we see the results, we see how fast it's moving, where are we going to take this, what are we going to build next and that is so exciting and I think a lot of the conversations today are really going to cover that so you know where to put your focus, what's coming up next and as I mentioned earlier the robotics that is something that I personally am very very keen on. Now I think with Jensen's keynote today, it's going to really focus and shape as to what's coming over the next decade really where these technologies are headed, how these enterprises are building with them, which is something extremely exciting. Now I can hear behind me the crowd, the line. I mean, I remember last year there's some bridges that are going like 10 minutes, 20 minutes past the venue and people are still lined up. It's really incredible. So, stay tuned for some of those conversations I'll be having. But in the meantime, I'm going to throw it back to the stage.

Thanks, Tiff. Looking forward to it. Accelerated computing is driving breakthroughs in science, transforming manufacturing, and advancing healthcare. It's turbocharging how engineers imagine, simulate, and refine the technologies that shape our world before actual fabrication even starts. What began as a way to render pixels has become the foundation of a new era of computing. And its moment has truly arrived.

Here to unpack this pivotal moment for accelerated computing are Mark Edelstone, managing director at Morgan Stanley, Dinesh Nirmal, senior vice president software at IBM, Aki Jane, president and CTO at Palantir US government, and Anirudh Devgan, president and CEO at Cadence.

Mark, I thought we'd start with you. You've been involved with Nvidia as long as anybody outside the company as the analyst who took them public. It was quite a risk for you at the time given 3D graphics was not maybe the best industry in the world. I'm just curious what did you see in Jensen? What did you see in the company to make you take that risk? And I'd love some anecdotes from the early years.

Yeah. Certainly it's been an amazing ride from the very beginning. I first met the company back in 1996 and it was an interesting start. So they had just basically done their NV1 and NV2 chips. Didn't go so well. And so I met some other people and began to think about them much like the other 60 or so 3D graphics and 2D graphics wannabe companies effectively that were coming to market at that period of time with their products. I met Jensen a year later in '97, completely turned things around. It was pretty amazing what they did. They brought out their RIVA product line that just completely transformed the industry. And so you know what I saw there was this company that had this incredible view that you got to build platforms. Basically it's a combination of an architecture, algorithms, software compatibility and do that over and over and over again. And so they really just changed the whole environment what we had seen historically where in this market for graphics everybody would leapfrog one another. And so companies could maybe hold together one or two generations of success and then the next player would come in. And Nvidia really just changed that. And they changed that in part with a whole different approach to how they'd bring products to market. Jensen would basically call that three teams for two seasons. And back then effectively you had fall and spring refreshes for PCs in the graphics marketplace. And so ultimately people would just kind of miss a product window and then therefore they would kind of lose their position in the market. That three teams for two seasons really changed it. So you constantly could leapfrog yourself and do it faster and better than other people. The vision of the beginning was basically that of computer vision that ultimately then transitioned into accelerated compute and the story has just continued to get better and better ever since. So just tremendous combination of vision, technology, roadmap, execution that really has allowed this company to stand the test of time. So at the end of the day, I didn't take that much of a bet. When I met the people and understood what they were trying to do, it was pretty clear to me they'd be very successful. Now it wasn't that prescient to know that it would be the most valuable and most profitable company in the world, but you could see that they'd have long-term success. And now it's more like maybe nine teams for four seasons with all the products they have.

It's such an amazing expression of a platform company and it was amazing foresight of yours, Mark. Dinesh, you have been part of a platform that's reinvented itself over multiple generations of technology. When did you start as a company paying attention to this generation of AI and say we need to do it again?

I think if you look at today for the generative AI revolution as a platform to continue, I think one of the core elements is data and the reason I say that is because if you look at the data in the past it was always data as an endpoint meaning you have data stored in a database you write a structured query language to extract the data. Today data as tools I call it meaning you have MCP server that is enabling access to data. There's multiple tools that's available for you to go access the data. Data is chunked and stored into a vector database. All those things and for that platform vision that you talked about to come alive tomorrow I call it data as code or data as skills. The reason I say that is because so let's say you have a document that you want to extract something today. You just say how much was my bill or how much did I pay for my phone? It extracts from the document and you get an answer. Tomorrow you want to do the same thing when I talk about data as code or data as skills what's going to happen is you can go to cloud code and say create me a program or an application that enables to do that. That is the trick because that enables you to one, flexibility because now the next time you have the same set of questions or different questions against the document it can modify the code rather than regenerate the code. So the flexibility is there. Two, optimization and scale, right? Any platform, a data platform for generative AI needs that. How do I optimize it? How do I scale it? And the third is the velocity and the velocity of the data that's coming. I mean, if you look at enterprises, I would say 90% of the data is unstructured. How do you correlate with your transactional data is where the value is. So that platform vision to come alive, I would say three things, right? We have gone from endpoint to today which is tools and tomorrow how do I create data as skills or data as code. So I think that transformation is what we are excited about. That's what I think the future is and how do we unlock that value for enterprises where 90% of the data is sitting in unstructured and how do we correlate to the transactional data to bring the real value out.

Makes total sense that you would have the early insight building, you know, as IBM was the data infrastructure for all of these enterprises and saying, 'Aha, we finally see that be it tools or skills is going to get a lot cheaper, a lot more accessible, a lot more powerful from here.'

Exactly. Exactly. I think that's where the future is because there's no shortage of data. I mean, you know, the chats that you're doing, the documents that you're creating, everything is data. But how do you really create the value for that data is where the enterprise are struggling because how do we make sure that you know there's so much transactional data happening but the velocity of data or the volume of data is coming from unstructured side of things. So every enterprise is looking to say how do I really create value on the unstructured side but also bring the transactional correlation to really make generative AI reality especially on real time.

Awesome. Aki, first thank you for keeping America and our allies safe. But second would love to hear, you know, Palantir is I think autonomous with AI at the application layer and would just love to hear when Palantir started what AI could do, where we are today, and what do you think we'll be talking about next year at GTC from a capabilities perspective.

Yeah. No, absolutely. Thanks for having me. And I think kind of the points Dinesh made are really critical if you think go all the way back to the origin story of Palantir. We kind of as a bunch of software engineers were doing in order to enable insights into that transactional and unstructured data, recognized that we need to be able to treat data the same way we did software. We needed things like revisioning. We needed the ability to actually put structure on it similar to original object-oriented programming. And to make that really accessible to end users in order to enable humans, right? And at the time, you know, AI was not mainstream, right? And there was a whole lot of AI exoticism back in the early 2000s. And unless you were a really really big internet company who had lots of data, right? You were Google, something like that. It was really hard to actually leverage machine learning in ways that drove ROI for your business or for your mission. And so we really focused on the data problem and making data relevant to humans, curating it, and enabling large teams of humans to operate on that data together, similar to how software engineering teams worked. And so if you fast forward to kind of like 2016, 2017 where computer vision now on the back of Nvidia's investments and deep investments in CUDA and GPU all of a sudden became a lot more useful, right? This is when Elon said hey we're going to do everything with computer vision and autopilot like we don't need these types of technologies we can do it in the low SWaP form factor. On the government side that became really interesting, right? And on top of Nvidia's GPU infrastructure, working with a number of companies the ability to bring those computer vision models to keep soldiers and sailors safe when they're doing their missions kind of became a really core part of what we were doing. But the way you do that is with data, right? And both the unstructured, so unstructured isn't just text, it's video, it's audio, it's pictures as well as text. Putting all that together with your transactional systems in order to enable the AI to actually make sense of your ecosystem. And I think where we are today with generative AI and a lot of where we've invested with AIP and closely with NVIDIA and the team has been a function of saying just you know we have all these systems we have the ontology we have the ability to understand the organization through data right now we know how to onboard humans to our organizations how do we onboard our AI teammates to work closely with humans hand in hand to actually accelerate a set of outcomes but then to do it in a way that's governed that's secure and that ultimately scales, right? And we kind of think about it as we're moving into a generation of technology where the limiting factor on scale for an organization is not just neurons, it's now electrons actually, right? And the more electrons we can pump through to actually accelerate and scale out our outcomes, the better. And that's really kind of where we're focused right now with kind of the next couple years.

I am going to ruthlessly adopt LIFAC. It took me a second to figure it out, but I love it. Yes, limiting factor. But it would be great to hear a very specific example maybe with your forward deployed engineer model for the last three months where Palantir was able to really help a mainstream Fortune 500 American company.

Yeah absolutely. So one of the things we've really started to invest in is what we call an AIFDE, right? So these are AI forward deployed engineers and the whole concept of forward deployed engineers early on was to say a lot of organizations and companies don't understand how to model data how to actually build an ontology in a way that's going to be useful to answer their long-term analytic questions. We've all done this. We've built five-year schemas and road maps of organizing all our data and like on day three you abandon it. It's very hard, right? Yeah. It's got to be iterative. And so really a good example of this would be some of our work with GE Aerospace as an example in 2024. They're really focusing on actually modeling. So building the simulation modeling around a particular engine. They kind of presented this I think last week at AIPCON. They were basically explaining how they started with modeling and sim to actually have like a digital twin of the capability. From there they actually took it and they built ontology around it to be able to understand all the parts components and different elements that went into it. Then they kind of extended that all the way to their supply chain to have kind of full end-to-end understanding of how should these engines perform when they don't and they're out of band how do we actually start to get ahead of that from a supply chain perspective to build capability to get ahead of it if you can now do that not just with humans but with AIFDEs which we're really investing that are AI oriented so it's a human plus AI effectively an AI FDE enabling them to now scale and do that for the next 10 components so what used to take three months will now take three days because you have that ability to scale out. That's a tangible example of something that we've seen in the last three months leveraging generative AI.

Aki, I have so many questions about reasoning in such a mission-critical field and thinking about these different areas of engineering and operations that Palantir's working in. But maybe the other pinnacle of humanity's engineering prowess is chips and chip design. Anirudh, you run the platform that enables the platform, right? Can you talk a little bit about what Cadence has been doing to enable more innovation in chips and how AI's changed your workflow to date? I think everybody at this conference hears continually about coding agents and the productivity and power that comes from them, but for this area of engineering, are you seeing the impact?

Yes, absolutely. I mean we have worked with AI and Nvidia for some time, you know we have partnership with Nvidia for last couple of decades and it has really accelerated in the last five to seven years and there are two parts to that partnership one is what I call design for AI, you know our software is used to design Nvidia's GPUs and most of the AI systems in the world and the second part is AI for design, you know we can apply AI to our own products like you mentioned and the effect can be quite transformative. You know I think one thing of course we've seen application of AI to coding, you know software, you know C, C++ coding but in chip design also there is coding there is SystemVerilog and RTL, you know it's less known is register transfer language. So once you have that, you know we have all kinds of tools over last 20 years which optimize and generate the chip but in the past we have not been able to generate RTL code or RTL test benches. So with agentic AI flows now we can generate that top layer and then put the automation that has always been there below it. So I always said for years that you have to look at it as a three-layer cake, you know and the top layer is AI now agentic AI, the middle layer is ground truth, you know things like physics, chemistry, you know how molecules work, people forgot that's important too, you know and then the bottom layer is computer and data like it was mentioned. So any successful application a vertical slice of this cake, by the way the reason I call it a cake is unless you're two years old you have to consume the slices together. So it has to be AI plus ground truth, you know like the fundamental how physics, chemistry and then compute and data and that definitely applies to chip design just like it does to other slices of that cake.

So Anirudh, what will this mean in kind of the most specific way for semiconductor design? Are we going to have better chips? Is the pace of progress going to accelerate help us offset the slowdown or end of Moore's law depending on who you talk to? Are we going to be able to design more chips more cheaply like what is this going to mean what does the future hold?

Yeah it has profound implications. So first of all on Moore's law, Moore's law some people say is completely dead some people say it's alive. I think what you have to remember is what parts of Moore's law is completely alive is area scaling, you know maybe performance is not improving but the chips are getting smaller and smaller, you know you go from five to three to two to 1.4 to one. So every generation there's double the transistors on the chip. Okay. So that's likely to continue for next 10, 15 years and further accelerated by 3DIC, you know just like Blackwell or you can pack more chips in a package. So as a result the size is going to explode, you know it's going to be 10, 20, 30 times bigger than now and the history of chip design is always the workload is exponential. See that sometimes people get worried okay if you are too efficient does it change how the things are done, you know like do you need less tools because you know tools are 10 times faster but in case of chip design the workload itself is exponential so we need 10x just to keep pace with that exponential and this is the history of chip design, you know if you look Mark has witnessed this just like I have over the last 20 years, you know our software has made chips design 100 times better but still there is more design and more activity. So I think to keep up with this exponential we have to get 10 times better at least 10 times better with AI just to keep on that exponential complexity.

Aki I want to come back to you just because Palantir has been going for a while serving some of these very risk-sensitive or at least very serious customers. I think we talk about this period as the era of accelerated computing becoming mainstream. How does more capacity or the exponential that Anirudh and Nvidia as a platform enables how does it enable you to do more? Are you compute constrained in the use cases you have?

Generally we're not necessarily compute constrained actually. I mean really on the government side of things, we're generally capability constrained, right? Getting the same capabilities that, you know, we can pull out our phone and use into the hands of war fighters and analysts and operators that need those tools to have some kind of compelling edge in their mission has traditionally been the area that we've seen be the limiting factor. But um

And sorry, when you say capability constrained, you mean in the product sense for the military sense versus the model sense?

Yeah. Well, no, actually both. Traditionally, I think that's changed actually, which is why this is so exciting. I think if you 15 years ago, you know, when we were kind of getting off the ground, there was massive disparity in what was available to those users versus your traditional commercial or kind of public or private sector users. Now, what we're seeing is there's a very very short lag. In fact, it's incredible how quickly this technology is being leveraged. And so, to that degree, then you do start to become compute constrained, right? And I think kind of on Anirudh's point from a workload perspective there's only more work to do. There's only more data to analyze. Like these are not the things that we don't have. And so what's really cool we're seeing right now is that traditionally the problem has been there's been a lag in the software side and there was tons of hardware, tons of compute. Now the software is actually the baseline is actually quite similar. And so now actually compute and scaled compute for the purposes of these missions are probably the limiting factor. But the data, the mission, the space, the humans, those are all available in abundance.

Dinesh, I'd love to hear from you. I've heard Jensen say that all enterprises in the future are either going to be token producers or token consumers. And I think kind of the logical conclusion of this is if you're not a token producer, your fate as a business is going to be entirely determined by how efficiently and effectively you consume tokens as an organization. So I'd love to hear from you because I know how many companies IBM works with. What are some common features of your customers that are successfully consuming tokens?

Yeah. So I go back to the data side of things. So if you look at enterprises, I would say there's three or four core things they look at, right? One is security. Obviously, it's very critical. Scalability. How do I scale? And that's where the tokens piece comes in because it's not just producing and consuming. How effectively and efficiently are you doing that? Because there's a cost associated with it. The third one is reliability because once again coming back to the token if the answers are not correct it causes even more damage for an enterprise. So if you go back to the billing question which is what was my bill last month and why did it go up by $5. Now when you ask that question obviously everything is a token but when you ask that question you're expecting the correct answer not somewhat correct answer but accurate answer. If you don't get that you as a consumer is not happy immediately what do you do you cause some damage by putting it on Facebook or whatever else. So for every enterprise it's not just about the token but the reliability of that answer becomes so critical in this world where we live in where the data information flows so fast. It's so important that reliability. The last piece that I would say in that token producing piece is responsibility. Meaning can I answer why that answer was given. In the generative AI world it becomes a bit more hard and that's where the companies like IBM, Palantir can come help because the explainability of that answer becomes so important. Why did I produce that answer and how do I if a regulator comes especially in the regulatory industries you got to be able to explain why you made that decision or you gave that answer. So 100% right that token producing and token consumption is important but the work that needs to get done and this is where Palantir or IBM others are excelling because you got to take that data that's coming in the ingestion phase how do I clean that data make it trustable and make it available for these models to eventually for a consumer to use it as token consumer or producer. So one more example and then I will. So if you go back to the past right it was all about events everything at an enterprise is an event and events are causing a lot of data but events also was a consumer and a producer you have a producer of events and you have a consumer of events and that's where Kafka was born. Kafka took off. Same thing is happening here, right? I mean, you have a token producer and a token consumer. But in that process, the reliability of answers matter and the responsibility or explainability matters.

So in the same way that effectively you know Claude Code to me the reason coding really exploded as this use case is people figured out how to have a really good harness for it that gave the model the context and explain maybe not as much explainability but there's a big opportunity to build harnesses for every use case imaginable.

Correct and even on that to touch base as a really good point that you bring up which is you know it's not about producing code I mean you can use Cursor, you can use Claude Code, you can use many tools out there as an IDE to produce code but what is the quality of that code? Because if I'm spending more time fixing that code then I don't get the value. So that's why many vendors don't talk about the quantity of the code but they're more focused on the quality also how do I get task done meaning I want to go from Java 7 to Java 17 can I what used to take me one month can I do it in 2 days so I think it's moving from code production because everybody can claim I created 30,000 lines of code but if that 30,000 lines of code is not good quality then you're spending even more time fixing it so that's where the focus is going to How do I make sure? That's why I said the reliability matters. How reliable is the data? How reliable is the answers? How reliable is the code? I think that is where the focus is and that's where we are really helping enterprises.

As enterprises work through this hinge moment of struggling through the usefulness, productivity, interaction mode with these models. Can we look ahead a little bit? Mark, you have been watching this company and this industry for 20 years. There's never been more attention paid to semis and its impact on all of us. If you were going to make a prediction on the next era of Nvidia and what it means for all of the businesses that rely on it, what would it be?

Well, I think we're still quite early in this revolution that's happening here on AI. Obviously, it's been initially on training, moving to inference, but the whole world of physical AI is yet to come. And when I look at what Nvidia is so good at, which is driving parallel computing, taking advantage of their platform approach to things where you really have to enable the entire ecosystem, that's what I think we'll see tremendous growth for them in the future. And it's been amazing. You look at this company and how they've evolved over time. It's been pretty spectacular, and I think we're still early on in terms of its evolution.

I can't wait to have my own multiple robots in my house. And hopefully that happens later this year or sometime in '27.

Yeah, it is amazing. You know, you were talking about the early days of having 60 plus competitors in the graphics space. It's kind of unthinkable to people looking at semis now that it would be so competitive, but the idea that you're just chasing generation after generation transitioning to a platform and an ecosystem. What I'm hearing from you Mark is actually Jensen would tell us we're at the beginning of another S-curve around inference and you're saying you definitely see one beyond that in physical AI.

Absolutely. Yeah. I think inference is still very early days. We've obviously taken a tremendous amount of infrastructure build that's been happening to drive training but the reality is we all want to basically take advantage of those models. So that will be in first thing but I think there's a whole another wave that comes past that and we see it now with early days of automated cars and things of that nature but I think we're very very early in terms of how we'll use all this intelligence at the edge.

I'm a huge believer. Thank you so much Mark and guests. It's been a great conversation. Our friend Tiffany will be with us all day with some amazing folks. Tiff, who are you with now?

Thanks Sarah. I'm here with Dan Ping who is the VP of R&D at Siemens. Dan Ping, you know, AI isn't just transforming software, it's also transforming the physical world and Siemens plays such an important part in that. Can you expand on some examples of this at scale?

Oh, yeah, sure. Tiffany, it's great to be on your show. We live in a digital age. The cell phones we use, the cars we drive, the tools we use generate tons of data which contains vital information to the success of a company. But to extract those information, we need a piece of smart software algorithms so that we can make sense of the data. And that algorithms is called AI. You know AI models like large language model has been very powerful and versatile but at the same time with its breadth it comes with its weakness it's like jack of every trade master of none, right? It needs to be combined with domain knowledge and fortunately we have this generative AI so today with agentic AI we can actually build in the domain expertise of people working in the industry day in and day out and then we can actually using the MCP using the workflows and we suddenly empower the people who do not necessarily to be tech savvy. This is a game changer, right? This is very very important. You know at Siemens we are in the manufacturing business our software is run by thousands of manufacturers from the design simulation to the manufacturing flow to the supply chain management and we know how hard it is to achieve that. With AI with the GPU we can do much more and do it faster, right? At Siemens at CES the Siemens CEO Roland Bush and Jensen was on the stage they announced that there is going to be a joint partnership between Siemens and Nvidia in four areas the AI in simulation AI in supply chain management AI in manufacturing and AI in EDA, right? So I came from the EDA industry I can tell a little bit more about the EDA. So for years the designers has been using the EDA software to design the chips that is faster and better with GPU it's like a monster right compare with the chips you find in iPhone and today we are using AI and GPU to design GPU we do this by leveraging you know we run the physical simulation on the GPU with much improved speed. We use AI to build a process model with unprecedented accuracy and speed. So this is one area. The other area is we do use generative AI and agentic AI to control our workflow so that the software will think and take actions. That's very critical, right?

It is. It is. And Dan Ping, one thing you mentioned is just how quickly these workflows are speeding up now due to this. I really want to cover quickly here about digital twins. Do you have an example of a company that is using digital twins and you've seen the workflows accelerate?

Yes. You know as Siemens working with Nvidia's Omniverse we built this tool called Accelerator. Now the teams across the globe can use this software in a virtual environment with photorealistic, you know, in the world they can change the design they can tweaking make decisions that is great.

Dan Ping thank you so much for that example and for your time. I am so excited for what Siemens is doing and what comes next but for now I'm going to pass it back to the stage.

Thank you, Tiffany.

Just as electricity powered the industrial age and the internet powered the digital age, AI infrastructure is the transformative driving force driving the next wave. AI requires a new kind of computing. Purpose-built systems designed to produce intelligence at massive scale. Here's a look at how AI is becoming the foundational infrastructure of the modern world.

We're joined now by leaders building that global infrastructure. Michael Dell, founder, chairman, and CEO Dell Technologies. Our second year in a row. Michael. Michael and Mike Intrator, co-founder, chairman and CEO of CoreWeave. Lynn Chow, co-founder CEO of Fireworks AI and Joe Creed, the CEO of Caterpillar.

Michael, you are a legend at understanding the hardware business that comes with each technical revolution. You were there for the PCs, for the cloud. Now with AI, you are building these gigantic data centers. And yet there's also renewed interest in having a ton of compute at the consumer edge with you know OpenClaw and these types of projects. How do you think about where the distribution of compute is going forward?

You know I think it's going to be everywhere right? You know there's been the debates over the years about you know on what side of the network does computing power grow and the answer is both right it's kind of always been both but now with autonomous agents certainly we see tons of distributed inference growing and you know two years ago we introduced the Dell AI factory right and we had customers that sort of knew what they were doing But a lot of customers needed help. Now we have this full AI data platform with an orchestration engine to organize all the data. That's incredibly important. Of course, the infrastructure continues to evolve and expand at an incredible pace. And you know we have the GB300 in the workstation level and increasingly powerful machines that can run these autonomous agents all the way to the you know Vera projects that we're doing with Mike and many others and it's incredible to see the pace of the buildout and the demand for compute and inference is highly distributed tons at the edge tons in smaller data centers and tons in the larger data centers too.

I think one of the things that's really striking for everyone on this panel is the new level of scale you are figuring out how to operate at and able to enable this ecosystem, right? It's not a single unit, it's a factory. The complexity and engineering in that is something that you're enabling tons of other people to take advantage of in partnership.

Yeah. And it's a combination of of course all the advances in compute but also the network fabrics the storage you know we build a whole parallel file system our Lightning file system the fastest file system in the world you got to feed these GPUs the data super fast and as the networks grow and of course KV caching is incredibly important as you get more complex queries and multi-agent systems and so Having this sort of layered memory system with KV caching is in some ways the growth of the memory and storage and data is more than the compute itself because the way the algorithms work. And you know a big plot inside the enterprise is how do you unlock all this cold data that has never really been used. And this is what we see happening with these 4,000 plus customers now. They're figuring out how to activate that data and really make this real inside their businesses. But it's a lot more than just hitting a button and all of a sudden you've got AI. You know you got to organize the data and curate it and also work on the processes in a company to make it effective. But we're now seeing thousands of companies do this and you know I believe we're still at the beginning of it actually.

You know you make this really good point. I was thinking about just the overall scale of the clusters that we are being asked to build today. But the fact that it is a memory hierarchy problem, a system software problem, a model caching problem, a power and you know dirt and electricity like management problem. I think it's incredibly amazing what the ecosystem has put together and I think we sometimes overly focus on this one really powerful thing which is how much matrix multiplication can we do.

And importantly bringing all the data together because you know in the previous IT architectures you had silos of data hundreds or thousands of them inside a large corporation and there was never any reason to bring all the data together well now there's a reason to do it.

Absolutely. And an insightful observer who we're going to hear a lot from maybe I guess in 90 minutes described it as a five layer cake. And Mike you are CoreWeave Mike you are an important part of that cake and would love to hear I think a big focus of today's keynote is probably going to be the disaggregation of inference into prefill and decode. What could this mean for your business?

Yeah, it's a great question.

It's going to be one of the important themes that we're going to be focused on here and kind of as we move forward. It's such a fascinating panel because it's giving you, and we were talking about this in the back, a real insight into how broadly this is impacting every aspect of the economy. So, it really does stretch beyond compute, stretches beyond data centers down into the energy structure all the way up into inference. And so, you know, it's a fun panel because it really kind of touches on so many of those thematic requirements to deliver the AI factories.

I get really excited about the deconstruction of the workloads because it allows a company like CoreWeave to really kind of laser focus on each aspect of that to be able to optimize in areas that perhaps in earlier waves of computing it didn't pay because it wasn't at the order of magnitude that required that level of specialization. As far as the depreciation question you're asking, it's really interesting because what will happen is as you go about deconstructing these workloads, the ability to route the portion of the workloads into those part of the computing architecture progression will allow for extracting more value from existing infrastructure, right? And so what you're probably going to see is folks begin to pair older generations of GPUs to pick up certain portions of the workload while the most bleeding edge tech forward components of the infrastructure stack are able to really deliver on those computationally enormous workloads. And so I think, you know, talking my book a little bit here of course but I think what you're going to get is once again a realization that there's a lot of different workloads that are coming to the market and a lot of those different workloads will enable us to extract value from the installed base as well as the base that we're installing as we move forward through time, which leads to the ability to extract value in excess of the depreciation cycle that we're working on which we're really excited about and you know we've built our company around being able to really target how we go about using this infrastructure.

Absolutely. I believe firmly the disaggregation of inference is going to lead to an extension of useful lives of GPUs. Love that because you can put a decode optimized solution in front of a three-year-old Hopper, a five-year-old Ampere and get a lot more tokens out of those old GPUs. That's going to be great for GPU residual values, useful lives. And who would have thought that the disaggregation of pre-fill and inference was going to save the private credit industry because those four to five year useful lives people underwrote to might be eight or 10 years and those financing rates are going to drop which will be great for everyone.

Yeah. Listen, the more value that can be extracted from each dollar invested into the infrastructure, you know, certainly the better it is for the lenders, the better it is for the credit stack. But it's also really great for new use cases that are coming into existence because what it does is it allows ideas and new companies that may have been struggling to get their hands on or access to infrastructure to build whatever their new vision was or that whatever their new company was. And it's incredibly important for all of us to be able to seed that next generation of companies, right? And if they can't get access to the compute that they need to build their companies to drive their companies to monetize their ideas, it really it's bad for the ecosystem at large. And so, you know, I think it's really great as you move through the stack to extract more value from what we have already invested. I mean, those dollars are out the door. They're in the data centers, you know, they're in the servers, they're in the supercomputers, they're in the energy infrastructure, we need to milk all the value from that we can to be able to invest in the next wave.

Absolutely. Getting more compute out of what we have and lower financing rates. It'll help us get the thousandx more compute we need, the 100,000x more compute or even more. We need to finance this.

Right. We start to see increasing significant explosive increasing adoption of customized inference and Fireworks is on the critical path of delivering that value and we provide customized inference with customized model for purpose built for applications. And before we talk about like 2026 is the year of efficiency, I would like to talk about 2026 is the year of explosive innovation, AI innovation that's happening. And we can see that from demo to production scale it drops from multiple quarters to multiple days and Open Clawbot is just a poster child of beginning of this journey. And Fireworks, we massively agentize our day-to-day work workflow. This can generalize the whole entire industry where we build our recruiting agent to source top tier candidates. And our finance team build agents use Fireworks to automate finance forecasting and the budget planning. Our legal team are building using Fireworks to build their legal tag. We obviously we develop inference software and we use coding copilot massively. And what we see is across the whole entire industry the tools are getting better and therefore agents can build more efficiently on top of new tools and new tools get further better and the new agent get emerged to replace old agents. So there will be a huge amount of energy in the application agentic space to continuously innovate. And the whole ecosystem will take some time to settle down and we're really happy to power all this innovation at light speed. Now talking about 2026 as the year of efficiency. We have seen the explosive usage consumption of token from our own company and from all of our customers. And it's kind of interesting phenomenon that now today having product market fit doesn't mean having a durable business. It's very strange because in the typical SaaS era having a product market fit you just scale as fast as you can and today there are so many cases whether they are startups or incumbents they will literally scale to bankruptcy. So this is because fundamentally inference is not as cost efficient yet as before because we the whole entire supply chain we are going through the evolution of being operational efficiency. All the way from the software stack to hardware to data center to energy the whole entire system is going through this revolution. I think we are all working together everyone is all working together to bring down the cost of ownership. We are approaching this in a very unique way. What we think about this is we don't think application or agent builder should be API wrappers. They should activate their private data produced by their application or locked inside enterprise to bring their model to the next level of intelligence and use that to also make the model faster and a lot more cost efficient. So this is something we call three-dimensional optimization that Fireworks produced as a unique way of approaching problem and we can bring down your TCO by five to 10 times. At the same time we're also working on cutting edge technology. We're thinking why we separate out model training from inference. We should all mix it together. While your model work with your application to do inference, your model should continuously learn the specific pattern from your application. Therefore to make it faster, better and more cost efficient. So we are working a lot of those cutting edge technologies to bring customized inference into fruition and obviously our innovation would not happen without significant innovation from all of you the technology at basic infrastructure level and particularly we are very excited about the new Vera Rubin and Grace integration because this is a massive disaggregated deployment at hardware level but we also think hey when we blend training and inference together we can also do a lot more aggressive disaggregation as in every single bottleneck they shouldn't be blended together they can scale complete independently to drive the maximum amount of efficiency. So this is extremely exciting time for us and to see all the impact building on top of us and work with all of you to drive the innovation here together.

Awesome. So Joe, thanks for being with us here. The world is structurally short wafers and watts. And we just heard a lot from your fellow panelists about how we're going to get a lot more tokens out of the wafers we're making because of the new technologies that we're going to hear about in a little bit. More tokens per watt means Jevons paradox. We're going to consume way more tokens and way more watts. So if you thought the world was lacking in power before, you know, look out 12 to 18 months but I know you guys have been increasing capacity. So Joe can you just talk about how Caterpillar, you know, in that five layer cake nothing can happen without that bottom layer of Watts.

Sure. I mean I think it's amazing again this panel to be up here. You know, you think about founders of tech companies and everybody that's here and that as a hundred-year-old industrial company, it shows really how widespread AI and what we're seeing with technology in the world. I mean, the infrastructure buildout's going to be like nothing I've seen in my almost 30 years of career at Caterpillar. And, you know, I love how Jensen lays out the five layers. And you talk about energy and chips and infrastructure, those first three layers. I mean, that's what we enable at Caterpillar, right? We are construction, we are mining and we're power and energy. So when you think about what our customers do, you know, we need critical minerals for chips. That's through mining. We need iron ore and copper to build data centers. Our construction equipment builds the data centers. Our power and energy equipment, you know, we're massively short on enough energy and watts to keep up with this buildout. So, you know, we kind of refer to ourselves as the invisible layer that kind of supports all that infrastructure buildout. And I think it's just an incredible time for a 100-year-old company to be looking at, you know, what we have in front of us and how we play into the bottom three layers of that. What's also kind of fascinating is for Caterpillar, we're also using the top two layers. We're big users of the technology as well. So when you think about we just launched an AI assistant for our customers. We're going to launch it in the cab of the machines to be an operator coach later this year. And so our mission is to solve our customers toughest challenges. And the things that I hear from our customers are, hey, first off, job site safety. Same in our factories. We want our employees and our customers to be safe. Shortage of operators. You know, we need to be able to quickly train operators and have coaches in the cab for them. And then also job site efficiency. There's a lot of inefficiency in our factories and on job sites where we can use digital twins and AI to really unlock a lot of value here. And so, you know, as users of the technology, I mean, we are just at the cusp of really getting out there taking putting people at the forefront and moving them from the dangerous jobs to where they're overseeing equipment working instead of necessarily in the cab. And just having coaches. I mean one of the things Mike and I talked about this last night. There's a lot of talk about the power and having enough power infrastructure to keep up with the growth that we have, but we need more skilled labor. You know, our customers are struggling to get operators. We're trying to help them through autonomy solve that challenge, but we need more electricians. We need more technicians to work on our equipment. And so we're making big investments in the workforce at the same time because, you know, right behind power, I believe, is having enough skilled labor to really keep up with this huge demand that we have for infrastructure across the globe.

That's awesome. So AI is not only creating jobs, it's creating better, safer jobs. We need that AI to meet the demand we have for watts and wafers. Makes sense that you would be investing in, you know, these supply gaps on labor that we can't solve. Beyond that, how else are you making Caterpillar go faster? Right? Because you can't train new people to match a thousandx or 100x demand. What else are you doing to compress the bottleneck in the physical construction?

Yeah. So, we have a number of things going on. One, we're just trying to add enough capacity to keep up with the product and demand that we have. But internally, you know, changing how we work and how do we use technology and now we're moving into inference. You know, how can we embed it inside our company to help our people develop product move faster? How do we take the inefficiencies out and get people off of sort of mundane tasks? You'd be surprised as a 100-year-old company, the processes we have and how many opportunities we I think that we have in front of us that we've barely even scratched the surface on of really making our people more efficient and kind of elevating them out of what I would say is kind of mundane work and putting them into where they use their mind and creativity to really move the needle. So I think for us, you know, it's going to really be super impactful for our culture and just how our people work every single day, both in the office and in the factory and in the field. You think about our technicians that have to support and our dealers technicians supporting customers in the field, having quick access to an AI assistant to diagnose problems, you know, remote monitoring and diagnostics for us so they show up ready to get customers back to work and that's what we're focused on.

I think this is a really important point because you know organizational structures and systems are a function of the tools that were available at the time they were created, right? And now we have this incredible breakthrough technology that is evolving rapidly, continuing to advance. And so how you get to the outcomes in 2029 is very different than you did in 2019, right? And so you really have to reimagine the entire processes from a tops down standpoint to get to the outcomes and then you get incredible speed inside your company. And you better do that or else you know you're going to get destroyed by somebody who does. 100%. The third quarter was actually the first time that we saw kind of real S&P 500 companies in a variety of industries talk about AI driven productivity have it show up in their financial statements and have the stocks go up significantly. You know, CH Robinson, a trucking company was a big AI winner in the third quarter. But it'd be really helpful kind of maybe a little bit of a lightning round. And maybe Michael since you posed it, you could go first. But what is the most specific mind-blowing productivity gain you've seen at your own company in the last six months through AI where you were just like wow I didn't think we'd get that until 2030 and it's here now.

So many examples I could take it the rest of the panel. Won't do that. But the one that I'm most excited about is the speed of innovation, the from idea to product delivery in our product groups, right? So 20,000 plus engineers enabled with the latest tools and they're able to do things in two weeks that would take nine months, right? And so that is making us a far better company and we're just innovating at scale much faster and that is super valuable and when I think about all companies sort of on that path I mean that's just an enormous speed up for drug discovery, solving all sorts of unsolved challenges out there and ultimately as we get these better tools we're going to do more the economy is going to expand.

Absolutely. Michael, are you seeing that kind of speed up in such a real world operationally intensive business?

Absolutely. So one of the great things about this panel here is that for me is that I'm getting like affirmation that every single part of the stack is seeing similar types of challenges, similar types of hurdles that they have to address in order to maintain the pace of the buildout of infrastructure. And you know whether or not we ultimately hit Jensen's lofty numbers of how much more compute is required or not directionally that is where we are going. And so the challenges that all of us are going to face are going to be sustained. One of the things that I do believe is that when you introduce new technologies like this, they tend to come from the fringe inward towards the core of what a business does, right? And that there's a numerous different...

Who's the fringe right now?

So you start out with, you know, hey, I'm going to make my email a little bit better. I'm going to do this. I'm going to do that. And then all of a sudden, you start to see AI presenting itself through engineering, right? In the coding. And you know when you ask me about what's astounding or what's magical around artificial intelligence when I sit with my CTO and he says that basically unless I'm on a phone call I'm 10x more effective, these are pretty talented engineers to be able to magnify the impact of those engineers who are building this by 10x unless they're working with a client going through the kind of normal cadence of how you build a business. That's incredible, right? And so my expectation of how our company will feel the impact of this was that you start on the margin and kind of work your way down towards the core of what a business does. And so you're going to see enormous strides in terms of the velocity with which we're able to drive our technology stack, the software layers that sit on top of it. How we go about configuring our data centers to optimize and futureproof them so that we're able to support not just the current iteration of technology but into the future. You see it working its way through HR so that we are not hiring for tasks but really hiring for people that are going to be able to carry the company forward. All of these things are driving more productivity and more scaling within our organization. And I think that that's probably a progression that you see across industries and you know like we're still years away from where it's going to really get to the core of some of our heavy industries. Just going to take some time and it's just really exciting. And absolutely we're I think we're tracking well ahead of Jensen's lofty predictions from previous GTCs. So let's hope we stay on that path.

It'll keep us busy. Thank you so much to this group. You guys are the builders behind the builders. So, we are counting on you. And thank you to my awesome co-host, Gavin.

Thank you, Sarah. Thank you, team.

Thank you. Thanks, everybody. Let's check in with Tiff and see who she's talking with backstage.

Thanks, Sarah. I'm here with Claudia Blanco who is the chief innovation and AI officer at GE Vernova. Claudia, you know, when we think AI, we're really starting to realize just how much AI and energy are converging. What are some opportunities there for GE Vernova?

Thank you for the question, Tiffany and first of all, I'm really glad to be here today with all of you. So, we are in a unique moment in time, right? So, we see like large AI data centers are demanding more power, more energy and is where we came into the intersection of things, right? So, we need AI for AI actually is what I like to say. So how those large data center will require energy when the power system requires the combination of things to happen right. So from how to generate more energy to how to transfer this energy with a more capacity but also how the distribution can be done more efficiently right and also combining with sustainable solutions. So it's a large mix of things that has to happen at the same time to make it work and AI at the same time play a key role. So at the time that we have to provide this power system and all that is required to make it work. It's also very important to know that AI can help a lot to manage the power systems. Right.

Absolutely. I love that you said AI for AI. I think I'm going to steal that from you. I'm going to use that one. You know, I'm curious to get your perspective on this. What are some parallels when it comes to the energy grid and AI infrastructure that you're seeing?

A very interesting question because actually we see that parallel lines. If we think about how the power system infrastructure was built 100 years ago, it was a unique moment as well. Many things had to happen, right? from the power architecture to be built to also the transmission systems and also the subsystems to the distribution as well as the automation and control at the edge. So all of these had to happen 100 years ago, right? And we see something very similar today with the AI infrastructures, right? How the large data centers will require a combination of multiple things to happen as well. That will require a big transformation. So how to merge the power supply, the power systems with computing with also all that is going to be the cooling systems and the infrastructures management of an AI data center. it will require also a big transformation. So I see very important parallelism there.

Absolutely. I think it's so important to be able to look back. You mentioned a hundred years of energy grids and taking learnings from that into our future and what we are building next. I'm curious to get your take. How is software really reshaping energy management?

Yeah. Well as well this is something that is going to be critical for us you know that with AI actually we can give for the first time the capabilities to our operators to manage the grid to manage the power system in more intelligent way right so which means that if we provide this intelligence power system is a complex system right so it's a large amount of data to be managed at the same time that system hardware physical system at the edge and then how can you operate that in a larger scale right so AI will provide ways to handle this massive data in a more comprehensive way more intelligent and is where it will play a key role and connecting to the previous question I think the AI factory concept play a key role here because we will start designing from the beginning the infrastructure require end to end merging all those system power systems with what the AI data centers will require.

Absolutely. Claudia, thank you so much for shedding some insight on that. I am very excited to what the future holds. But for now, I'm going to throw it back to the stage.

Thanks, Tiff. We'll check back with you in a bit. I'm now joined by Sequoia Capital Partner Alfred Lin, who will help lead us for the remainder of the show. Welcome, Alfred.

Thank you, Sarah. Thanks for having me here. Thanks, Jensen. Thanks, the Nvidia team. You guys did a great job leading up to this. This is awesome. It's a bright day. It's sunny. It's awesome to spend time with you, too, because usually I spend more time with your husband.

Well, you know, one of us has to work today. It's great to be here with a friend, a OG Nvidia investor, and a co-tra as we try to figure out what's happening in the next wave of the ecosystem, which we're going to talk about now.

Yeah. Well, yeah, Sequoia has been very fortunate to be an early investor, a seed investor in Nvidia and Mark Stevens made that investment and he's still even though he's retired from Sequoia is still on the board and so we're very very fortunate to be associated with Nvidia. I wasn't there when the Nvidia investment was made that was just genius. But when I joined Sequoia I did check Nvidia stock price and it was a quarter. I think Jensen's added $180 to that which is a 721 times return in the last 10 years, 15 years, 16 years. So, it's amazing. It's an amazing company. It's pretty cool to see also, you know, hearing from Mark at the beginning of the day what the journey is from a competitive market of, you know, 60 vendors fighting about graphics to platform company with 3 S-curves in front of it and we're maybe in the beginning of the second one. I cannot wait to talk to some of the companies we are speaking to in the rest of the pregame about the platform journeys they're going on.

Yeah. And you guys set it up so well. We had the accelerated compute, the infrastructure. Now we're going to talk about open weight models and systems about AI agents and then we're going to end with physical AI. Great to have you here. Before we get to the next chapter, Tiffany has got another great guest on deck. Tiff,

Thanks Sarah. I'm here with Alex Atala who is the co-founder and CEO of Open Router. Alex, why is open so important when it comes to AI innovation?

A vast majority of the economy is going to move towards inference and compute. And a major principle we want to make sure is true in that future is that we have competition in the economy around inference. Open models allow multiple companies to compete to build incredible inference for the models that come out of model labs. And what that results for the end consumer and for businesses is faster inference at higher quality and more interesting integrations. And you just need an open economy to enable a really competitive future for that. And open models are the way to do that.

Absolutely. I couldn't agree more. Open models. On that note, when it comes to developers, how are you seeing developers navigate so many different models today?

Developers tend to follow the vibes which most people are familiar with. They do look at benchmarks but we find that happens when a new model launches we've observed something that we call the Cinderella effect which we wrote a paper about last year where the first cohort the first week of developers there's a section of it that figures out what the model is good for and they stick to it. They have the highest retention among all cohorts that follow them in the future. So they discover these novel use cases or what these open models are good at and then they stick to them and what we do is try to expose those insights to the rest of the world so people can see what the power users and what the LM enthusiasts are doing with these new models. We also see that closed models tend to boost the total addressable market for a particular use case and then open models come in and help people reduce costs which becomes a huge thing when you realize like wow our costs are ballooning. All of our developers are doing crazy things we didn't expect.

Absolutely. And was the Cinderella effect? Is that what you call it?

That's what we called it.

I like it. I need to read that paper. That sounds very fascinating. I'm curious, what startups are you starting to see emerge?

We're seeing a lot of startups try to reinvision how to organize a company now with AI agents. Like what will it look like when you actually try to give them responsibility and when you try to give them full context because these two things are the two big gaping holes in AI today. Like agents don't have responsibility for anything and they don't have the full context of the company. So we're seeing more startups try to work on those two problems. We're seeing a lot of startups just integrate, try to get more integrations to AI agents so that they can, you know, act on behalf of humans in different places. We're seeing a lot of interesting agents emerge to just help developers improve their productivity. Help developers manage multiple workflows that are happening simultaneously. A lot of startups are realizing, wow, I have the problems that 100 person companies usually have and much more.

When it comes to AI agents. What is an example? Just a quick specific example with AI agents that you're seeing a startup use them for?

Very specific example. Well, I could just think of all the cases our own company is using them for. I mean, we're building agents to help our sales team prep for calls. Very bespoke agents that run automatically before the calls happen. So that they know everything they need to know about who they're talking to. We have agents around auditing for security. We have security audits running constantly all the time on all PRs for code that was merged in a week ago. And now we have interesting rollup agents that are just kind of running ambiently in the background and looking for interesting insights in our data like what apps are emerging. We have data about public apps like apps that opt into public visibility and we expose all this data to the world in our new apps rankings and in our models rankings.

Thank you for those examples. I love hearing specific examples because it really gets me so excited as to what other companies are working on, what people are thinking about and we're going to continue a lot of these great conversations but for now I'm going to give it back to the stage.

Thank you, Tiffany. We'll check back with you soon. AI's impact won't come from chatbots alone. The real breakthroughs happen when developers take open models and build on top of them, adapting them to new industries, new problems, and entirely new kinds of applications. That's why open models are so important. They give developers, startups, researchers, and enterprises the freedom to experiment, specialize, and push AI into new frontiers. So, what happens when open models meet a global community of builders? Let's dive into our next chapter.

Open source models have become incredibly useful for developers. They are now the lifeblood of startups.

Each one of the industries have its own use case, its own data, its own flywheels. All of that capability, that domain expertise needs to have the ability to embed into a model. Open source makes that possible.

Here to discuss how open models drive innovation, expand the AI ecosystem are Aravind Srinivas, co-founder and CEO of Perplexity AI, Arthur Mensch, co-founder and CEO of Mistral AI, Robin Rombach, co-founder and CEO of Black Forest Labs, and Aidan Gomez, co-founder and CEO of Cohere. Welcome.

Welcome everyone. First, let me start with Aravind. You had a very successful launch with Perplexity Computer and you've described it as a computer that should be the personal computer that should exist in 2026. I wish I had that where it runs 20 agents simultaneously helps you do a whole bunch of things. Tell us about your launch. Tell us how it's going and tie that back to open models. How are you using them and how's that helping the ecosystem?

Yeah. So, Perplexity Computer is an orchestra conductor of everything AI can do today. It runs in the cloud and all the musicians in that orchestra is basically all these different frontier models, the connectors, you know, they're playing a music that's equivalent to the work that you want AI to do for you on the computer. One other way to think about it as a person doing work on your computer, what do you really need? You need a browser. You need access to the internet. So giving AI's access to the internet was the core idea of Perplexity. And then giving AI's access to the browser was the core idea of our Comet browser. Now we complete the puzzle where we gave them access to a file system and a code execution sandbox and 20 different models, hundreds of connectors, lot of CLIs. Turns out when you do that AI can become a director directing all these sub-agents to do all your workloads for you every day asynchronously always on. So that is the core idea of Perplexity Computer. Open source has always accelerated new applications and open models naturally fit into computer. We're already using models from Alibaba the Qwen series and we're going to be using a lot more. So models in our mind and as far as computer is concerned are just tools just like the file system or the code sandbox or these connectors or the CLI. Models are just tools what actually matters is the orchestra.

And how are people orchestrating computer today?

So many different applications people are building their own like many versions of Bloomberg terminal that's connected to their own portfolio that they have access through the Plaid connector. People are building web applications that famously monitor the situation around the world whatever is going on because you have access to the internet you can create your own live streams of what's going on. You can build applications that pull internal context. Our entire company Perplexity is just running on Slack where this computer is yet another person on Slack that is just orchestrating these workloads. We estimate around $90 million worth of labor for our max users within the first couple of weeks of launch of enterprise product.

That is really terrific. There's such massive increase in continuous agent usage. That obviously requires compute behind it. Arthur, Mistral has always had models that punch above their weight from an efficiency perspective. We just had a panel talking all about 2026 as the year of efficiency. Where are you seeing demand for this in your own customer base where people care not just about max capability but efficiency?

I mean the demand comes from all of our enterprise customers. It depends a little bit on the vertical but efficiency matters when you're deploying agents at scale and then suddenly you care about token cost and you care about the unit economics and that's where we build open weight models so that our customers can have cheaper access to that technology. So it matters for cost also matters for the UX and the experience generally because when you have a small model it can actually go faster and if it goes faster you can solve your task faster as well. So that aspect around latency is actually critical. If you lower the latency you can actually change the kind of applications you're building. And so when our customers are building for their own customers it typically is a problem because they want to have low good UX and good UX means to have specialized models that are small enough. There's a big use case as well that we target quite significantly and that we work with a lot of companies which is edge deployment. And edge deployment is about mobile deployment but oftentimes it's also about putting models into robotic systems into industrialization systems into factories. And so you need to have these small models that have audio as input that have video as input and that can process that and that can do it in a way which is fully disconnected. So we're actually announcing today Mistral Small 4 which is really punching above its weight in that respect and it's a reasoning model that can actually be deployed on the edge in that respect.

Congratulations. That's super exciting. I think one of the things that is somewhat forgotten when we talk about use of models is you want this iterative loop either in the models generation and verification capability itself or between the human and then I haven't been waiting for computers for I don't know a decade and a half before the last few years and I'm waiting all the time right and so I do think efficiency is going to be a huge front of competition.

Yeah if you put it that way it's super important and the other thing that is important is also to be able to modify those systems and to learn from interactions in between user and the models themselves and that's where having open weight models and a way to access the parameters and to change them to steer the behavior is actually critical when you go into the depth and the productivization of the AI systems you actually need that deep access.

Thank you. Arthur, on the sort of importance of latency, latency is very very important for the creative world for the generative AI world and Robin, for Black Forest Labs, how are you reducing latency? How are you sort of supporting the creative community in their quest to make better and better content out there?

Yeah. I think for the creative community in particular, open models are super important because they are more formable. We see that actually and I kind of want to disagree with this take of models are just tools. I think models are a super important substance for the creative community. It comes, you can fine-tune them, you can interact with them in all different kinds of ways. It's not just input output. You can take them apart. You can plug them into workflows, node based systems, stuff like that. And then they go way beyond the creative community. It's like a basis for research, for innovation. I think it's fundamental technology. And really not just a tool. For latency, yes, 100%. It's one of the most important aspects of all kinds of models today. We work on a lot of distillation research that we do. And yeah, I think we are pretty good at getting the number of inference steps that you need with a diffusion model or flow matching model down to let's say four or something like that which then really helps in combination with stuff like quantization to make these models really efficient.

You mentioned diffusion models. They started the whole wave for creatives and creative community to build things that they didn't think they could build as efficiently and as powerfully as before. What's next? What's on the horizon?

Yeah, I think it's really interesting that we see how especially visual models are expanding in scope from content creation to simulation to physical AI applications and it's like the shared technology underneath is the same or very related modeling approaches but we're approaching new scales of data, of models and multimodality that then enables these models to expand into what we find really really exciting physical AI simulation use cases real-time video generation.

Aidan, whether we look at these models as just tools or a moldable substance as Robin was saying, they've been relatively hard to work with for most of the industry even the engineering audience for the past few years. What are you seeing in your enterprise customers and the people you interact with about sophistication in terms of post-training and customization?

Now, it's increasing dramatically and I think open source models are contributing to that because it gives the customer the freedom to explore and experiment themselves without having to depend on any other organization. One of the things I'm most excited about with open models is the notion of control. So, Cohere does a lot of work with high security settings. Privacy is not just important but essential for the customers that we work with inside telco and financial services etc. And what open models do is give the customer the ability to choose where and how they deploy. So when we're delivering that technology when we're delivering North our agentic platform we have to make sure that our models can be deployed on prem airgapped on device anywhere. And open models enable that. In terms of customization I think that these models they're fantastic general reasoners. They can be applied to pretty much any domain anything. But there's a production gap that needs to be closed. There's a level of performance that you need when you're deploying into these extremely important high stakes scenarios that you're not going to get out of a generic model. So the ability to customize and close that gap is essential to actual deployment and scale of the technology into the economy and the ecosystem.

I have a hot take question. There's this meme in VC that closed is easier to monetize, open is easier to mobilize. Do you agree with that? Is it easier to get distribution? Is it easier to work with so that you're not locked in because you're having to pay one or two closed models. What do you think?

Well, I think distribution at the end comes down to value you provide. For example, there were a lot of open models, but you could argue that the pivotal moment in open source among model providers was when DeepSeek surpassed OpenAI o1. That's when everyone started to...

So, and I would say like the same thing keeps repeating, like Kim Moonshot, like I think even Elon called them out for their work yesterday night. So in general, closed or open, the capabilities are basically what fundamentally define distribution, and that's kind of why we've taken the position that we would use whatever is the best capability for each model and put it all into like one orchestrated system. One advantage obviously in open source is cost because you can take the model and do your own inference. You are obviously going to have a massive cost advantage serving that over paying a closed model provider. But what usually has happened over the last two years is every time there's a great open source model, the closed model providers always lower their costs, which is good for application layer companies like us, and we very much welcome that competition.

I think what we find is that the value is in the integration in between the models we do and the distributed systems that we use to deploy the agents to make sure that execution is durable, observable, to make sure that the models can improve when the users are giving feedback. So at the end of the day, if you only bring an open model to an enterprise, it's not going to know what to do with it. But if you bring it the right platform with the right sandboxes, execution with the right inference at scale that is efficient, and the right customization mechanism that they don't need to care about, they don't want to care about fine-tuning, they care about seeing the system improve over time. And so what we're really bringing to enterprises is the right level of abstraction so they can work on our open source models, on other open source models if need be, and make sure that they can actually create the business applications that are driving the ROI. And so the motion there, which has been successful for us, is to really bring that platform but also bring the experts so that they work with the subject matter experts and actually customize the systems at a level where the ROI can be found. So that's the gap bridging the gap between open weight models that effectively bring control and low cost and strategic autonomy, and the AI that you need to have if you're running an enterprise or running a line of business is the gap that we solve, and it's the gap where we build our business model.

Yeah, 100%. And I think making images or making videos is just like one application for these models as I said, but they are becoming more general. So I think like a lot of the demand is, it's actually not latent, it's like very obvious. Like if you look at the frontier models today, having that capability available open source, it's just like a no-brainer, you have to work towards that if you want to maximize distribution. So I think that's to answer your question of like distribution versus monetization, I think if you want to maximize distribution, just make the most capable model and get that out there. And again, it's like a basis for so much more than just image generation or video generation. And I think that when you have an open model available and these use cases are getting explored on top of the open model, that really helps to unlock new use cases that then other companies implement on top of these models. And I think like one pretty important aspect, especially when you look at agentic applications, is that there's a lot of model-in-the-loop kind of applications that will be enabled in the future or that people are already working on, where image generation is a great example, right? Like that's one that is being productionized already where low latency really helps you to make repeated calls to the model and then embed that into your product. For stuff like coding agents currently, I think there's a lot of work being done to implement the code and then you make the product out of it. But actually going towards more efficiency and having repeated model calls enabled all the time is something that needs to happen and I think where open source can also really, really help.

That's a really hard... I feel like it's a surprise every three, four months that comes out. The extent to which we've been able to create a general platform for intelligence that can be applied in so many different domains is a huge shock. When we were building the transformer, it was built for Google Translate. That was the objective. That's what we were climbing on. I don't think any of us who were involved in creating the transformer had any foresight into what might come and how people might run with it. And I don't think we can take any credit for it either. I think it's really the community that picked up the transformer as a platform and carried it forward into all of these different domains.

Market penetration on the consumer side of house in terms of like the average person using these models seems to be well on its way to everyone will be able to use this model in the same way that they used Google for search, right? That seems to be quite safe and steady and on a track towards success. On the enterprise side of house in terms of deploying this technology into global economies, there are huge barriers. There's the trust barrier. There's the privacy barrier. Trying to bring these systems into an organization, give them the access and authority to take actions that previously we would only be able to consider a human for. Those sorts of trust barriers, I think, are the ones that need to be overcome. Privacy is a huge component of that. Sovereignty is a huge component of that. And it's something that, you know, at Cohere we're super excited to make progress on.

I'll tell you an interesting stat. All right. So first of all, Perplexity started as a consumer company, but it's an interesting emerging outcome is our enterprise adoption is growing fast even if we don't want to work on it. So one thing that our enterprise users are doing, more than 50% of every user, not just organization, is picking different models every day they use the product.

So they can't... so at the beginning of 2025 my worldview was models are going to get commoditized, and towards the end of the year, and I think different frontier lab CEOs are saying this too, that models are specializing. Even within coding, people would say Claude Code and Codex are good at different things, and our engineers use both of them. So what's happening right now is models are specializing in different things, and so that makes the need for selecting models for different use cases very important. Of course, we're going to do that orchestration for you. But if you have a certain opinion yourself that this model suits my use case the best, you can go and pick it, and it's happening. The user behavior indicates it's happening, and it's one of the biggest value propositions of our app and our platform is that you don't need to feel locked into one vendor. And by the way, we are supporting both open and closed models here. Like on Friday, we brought up Nemotron, the latest open source model from Nvidia, and they want to keep improving it. And we're going to keep putting all their newer versions on our app every time that happens. And that also unlocks a different kind of application like people are talking about sovereignty and AI. You obviously need a model, but you also need the orchestration, the actual product that connects your internal data, web data, proprietary data, all these different tools, runs securely on sandboxes. Like that's the work we've already done except we can do it all with open source models too now and that'll unlock enterprise on-prem or sovereign applications for us.

Sorry, the open space?

So, I'd say the reality is that enterprises, they don't really want to care about what's underlying. What they care about is indeed controlling the systems, knowing that nobody can turn it off, and they care about the things improving over time. And so to Aravind's point, today some of them still care about what model to pick. But eventually that's kind of a bug. It's just a sign of immaturity of the space because what they want to have is really business applications that are solving their problems. And so when we use Mistral to actually do this kind of business applications and work with our customers to really create the software they need to operate their processes, they don't really care about it being open or it being closed. They care about having control. They care about if we are to disappear, they actually can maintain business continuity, etc. So that's the sovereignty part of it. So eventually I think in five years, the technology and abstractions that are needed to operate agents at scale reliably in a way that is improving, etc., this is going to happen. And we think there is kind of two worlds: either a world where you have just a couple of players that are locking the space from model to orchestration to applications, or a world where you have a choice in between the different bricks of the platform that you can use. So you can choose your own infrastructure provider, you can choose your own models, you can actually see the code of the orchestration side and have enough transparency and control over the entire stack. And so in five years I would expect that we'll come up with the right abstractions that's going to be the Linux of AI, where you have open-source inference, you have open source models, you have open source orchestration, open source coding tools, everything is actually going to be quite cheap because when it's open source you can deploy it yourself, so it's hard to make margins, but the integration part is where the value is going to be created, just like I guess Red Hat for instance has a pretty good business model. So the bet that we're taking is invest significantly on making open source win and then build the orchestration and the help that enterprises need to stitch all of these bricks together.

We are making our own models. So we are primarily using them. We've actually been working with Mistral for our latest Flux 2 model. So there's like a VLM component that is able to process the inputs. That's the Mistral 3 model and then Flux 2 is like our own from scratch trained model architecture.

I think our models are more efficient and better. So I think that's... I don't know, like our mission is really here to push the frontier of visual intelligence and these visual generative models. And I think to your point of like how do you monetize these open models, right? There's a lot of talk about like open source but then also about open weight, and I think there's a case to be made for new business license models around open source or open weight models in AI because it is so resource intensive obviously that you have to find something. And for us it's actually been interesting, we have a lot of open weight models that we put out there and they are contributing like the majority to our revenue, it's more than our API that we host. And yeah, we keep iterating on it, but I think there is a real opportunity to figure out something new.

Thank you guys so much. This is a great panel and core to the ecosystem here at NVIDIA. I did not realize that the commitment to open ecosystem came with a commitment to cool dark sunglasses. Alfred and I will sign up next time. You got to send a text, guys. Come on.

It's about the community working on things together.

Thank you guys. Really appreciate it. Thank you. Thank you.