Kari Briski at GTC 2026: The Future of NVIDIA AI & Nemotron 3

0:00This isn't science fiction. This is happening. You have this multi-agent fan-out of all working to get to a common objective and have the objective be more accurate, productive, and useful. Jensen described it as a new operating system. You're managing memory. You have file systems and knowledge. It's kind of like the new personal computing. We definitely know the importance of not just autonomous agents, but self-evolving, truly understanding the environment and adapting to what are you asking? Who are you? What do you need? What are the tools that you use? Prompting you doing things on your behalf.

0:33Now you're going to prompt it, like, go do this, go get this done. All of this isn't even about just the model anymore. It's about the total system working together. Open models are kind of having their moment because enterprises are starting to deploy in production AI. It's a part of the way you're going to develop software in the future. So the long-term vision is we're committed and we're going to keep doing it.

0:52Corey:Well, welcome humans to the Neuron AI podcast. I'm your host, Corey Knowles. Today we're recording from NVIDIA GTC 2026 in San Jose, California, and we're joined by a familiar face. This is Keri Briskey. She is the VP of Generative AI Software for Enterprise at NVIDIA. She's joined us before and is a driving force behind NVIDIA's push to bring open models to the masses, make them secure, make them cost effective, and all of the things. Keri, welcome back to the Neuron. Thank you for having me. Excellent. It's so great to have you again because you guys have been up to a lot of things. A lot of things.

1:30Corey:All of the things. So first, I'd like to talk a little bit, if it's okay, about Nematron 3 Super. That's a big step up for you in model size. What does that mean for NVIDIA's Agentex stack? We are really excited about Super. We had announced it and released it the week before GTC. And what it means to us. Well, we released Nano, Nematron 3 Nano in December at the end of the year. And when we did that, we decided, you know, we had a conversation. Should we? Should we not? And so we decided to kind of publish our roadmap. Another super is coming. Hey, ultra is coming after that. And so really having to deliver on that promise and deliver on the roadmap was something really important.

2:15When we were talking about the different architectures, you know, back in the middle of 2025 and the things that we were going to do with the hybrid model and then moving to MOE. We knew it was important, and we're really excited about the progress of Super.

2:29Corey:Nice, nice. Was this model designed primarily to make multi-agent systems economical? Was it meant to be more of a boost in intelligence? Yeah, and to make multi-agent systems economical. I'm actually, the answer is yes, and I'll go into more token efficiency. Maybe we'll get into that later. But I think that if you actually kind of step back and look at the progression of what's been happening in adoption of AI, you know, Jensen kind of talked about there was the beginning of end of 2022, beginning of 2023. Like, oh, OpenAI had ChatGPT and like the world woke up to what is generative AI and what does that mean?

3:07And then DeepSeek had its moment by opening up the reasoning models in January of last year, 2025. And some people are like, oh, there's these things called tokens and tokens produce intelligence. And what does that mean? And then and that's just about reasoning to get into a better answer. And then now we're talking about, OK, that we would define the definition of an agent last year. And then people would come to NVIDIA or come to our executive briefing center. And I'm like, OK, here's the definition of an agent. Here's how many tokens is producing. And like, but now when agents are talking to agents and everyone that came would be like, what?

3:41Is that really? Is that really happening? Like, yes, yes, it is. And so, you know, being the AI company that works with every AI company, like this isn't science fiction. This is happening. You have this like multi-agent fan out of the all working to get to a common objective and have the objective be more accurate and productive and useful for our work. And so, yes. So coming back to your question. Yes, we kind of saw what was coming. Yes, we knew that we needed to be more token efficient. We needed to have better latency. and there's no purpose in having an efficient and fast model if it's not smart.

4:17That's right. That's right.

4:19Corey:Well, of the architecture choices that you all made around 3Super, where do you think users are going to feel that different first? Is this in latent MOE, the Mamba Transformer? Yeah. You know, Nemotron 2 was actually, it was still a dense model, but it was the hybrid Mamba state space model plus transformer. And so you could see the latency improvements right away with that architecture. And then you add in the mixture of experts and latent MOE and then your compounds. And so you now see like a three to five X. I think you said three the other day in the elevator speed up just. And so we are always trying new architectures.

5:03This is kind of the idea of extreme co-design. We always have an agenda to exploit the latest features and functionalities out of our, I don't want to say hardware, but our total infrastructure. And so, again, we knew where it was going. And we know that we need to do more and a smaller amount of compute footprint so that everyone can scale to even more tokens and intelligence and productivity.

5:26Corey:That makes sense. That makes sense. Which do you think hurts more in production right now? Context failure or maybe? Unreliable tool use? Yeah, a lot of people, sometimes the model can follow instructions to a T. And, you know, it's kind of like a genie when you say, like, you know, you want world peace and then, like, everyone's gone in the world. Like, it's very peaceful. So it's kind of like, what's the, you know, how far do you need to follow instruction? I mean, I think it depends. Like, you asked what the failure modes are. It really depends upon what you're trying to do and the evaluation of it.

6:04So, you know, everyone wants more long context. And so I think it's more about memory management and efficient systems. So what's your harness? Because all of this isn't even about just the model anymore. Yeah. Right. It's about the total system working together. And so how do you evaluate the total system altogether and then be able to tune it to bring it in?

6:24Corey:Yeah, it makes sense. Yeah. There's a lot of the harnesses all the talk lately. It's very much we keep shifting from like, okay, we're going to talk about prompting. We're going to talk about context engineering. And then we're going to shift into this phase here where suddenly everyone cares about the harness. Yeah. And I guess that's agents that have done that to us. That's right. Or that have brought us to that point, I should say. That's right, because the harness is kind of the opinionated way of how to, you know, orchestrate, prompt, and manage not just the model, but, you know, systems of models.

6:54Yeah.

6:54Corey:Yeah. Yeah. I want to take a step back, too, on your comment about the speed. When you mentioned I was getting 3X, that's spot on. I'm using an RTX 4000 and running Nemetron 3 Super, which is a 120 billion parameter model. At triple the speed, I've been able to run a 70 billion parameter. I mean, it's still slower because it's a normal house machine with a good GPU, but it's significantly faster and very bearable when you want that extra lift. I've been really impressed. You're seeing that because of the combination. I'm assuming you said 70 billion. Is that just like a Lama-dense model? Yeah.

7:34Yeah. And I think that, and that's kind of the innovation architecture that's happened over the last two years, right? Okay. Because a 70B model is still just a transformer, and it's still a dense model. And so now you have a model that's almost double the size, 120 billion parameters. And so it takes up more memory space, but it performs as if it's a 12 billion parameter model. Because that's the point of the, it's a factor of 10. So 120 billion parameters to the 12 active. And so it's as fast as a 12 billion parameter model, but it has the knowledge of a 120 billion parameter model. Wow. Wow.

8:10Corey:That's such a big step.

8:15Corey:Is that done through finding efficiencies? I mean, I know there's a new architecture as well, but I feel like that's a big shift. How do you come to a point where that's doable? Well, I think actually a mixture of experts is kind of proven in the research world and almost all the large, really awesome open models that are out there right now are a mixture of experts. And so being able to do routing efficiencies, like which expert to route to and how fast and why is all part of the training paradigm. Makes sense. where, where do you feel like you wish you could still get this to grow and get better with, with say with super than even trying families?

8:56Like, where's your, here's where I still want

8:58Corey:to see some improvement. Um, yeah, I think just still being more robust. I think like long horizon and you kind of ask like, what would people see in the failure modes? It's not just a complex multi-turn, but also complex multi-turn questions over a long horizon or period of time. So we're working on that with, you know, post-training more, you know, different rubrics and the verifiers in order to kind of achieve that. So that's one for me. I think even kind of expanding out of language and better, being better at modalities. Like I'm really excited about our speech models. Our ASR models have always been really popular.

9:35And now having a true end-to-end voice chat model is really exciting. But then bringing that in, so you're going to see things from, you know, Most of our language models will release a vision version of that model. But now we're going to start omni models. So omni understanding of bringing in speech, language, video, images, and then being able to describe, you know, what are you what are you hearing understanding? You know, like if you heard me knock off camera, like, was that me? Was that a door? I can just be being able to understand what's happening. Like, what are you seeing and hearing and be able to describe it?

10:08Corey:Yeah. Yeah. That's amazing because at this moment, that's a front that still struggles in open source, in my opinion, on computers like an individual can run. And I really feel like that's this horizon that will be really valuable and really open that market. Yeah, and computers an individual can run. And I think it was Andre Carpathy, you know, he posts a lot. Yeah. And he had mentioned, you know, we're going to need a lot more compute to get where this is going because of all the things that he was excited and doing. And I think that that's true. I mean, the Build-A-Claw tent outside is like, you know, really there's so many people that have so much energy.

10:48And I think it's kind of interesting because this is like, you know, Jensen described it as a new operating system. Because, you know, you're managing memory. You have file systems and knowledge. And so it's kind of like the new personal computing.

11:05Corey:It is. It is. Yeah, yeah. And I would say OpenClock has brought a lot of things to the forefront. Like I care more about token usage in my personal life than I used to because it gets out of hand quickly. And the OpenClock thing has been really interesting to me. And I think it's really telling that in a matter of 60 days, we went from Pete Steinberger having dropped this thing on GitHub to Jensen in his keynote on Monday saying that if your company doesn't have a claw strategy, you're in trouble. You need to be on that. And I'd love to hear your thoughts about maybe what does that look like in a real workspace?

11:44Yeah, I think that that's why we're putting it out in the open to have the community contribute, because we're not going to go solve security and policy privacy all on our own. This is something that needs to be decided in the community and all together. But Jensen saying that you need to have a class strategy is because just like we saw the importance of reasoning models, I think we definitely know the importance of not just autonomous agents, but self-evolving, truly understanding the environment and adapting to what are you asking? Who are you? What do you need? What are the tools that you use?

12:23Prompting you actually is like doing things on your behalf. And so I think that that's why everyone kind of realizes the full potential of this is not just a, you know, how or what. You know, you always ask kind of you ask questions of how or what or knowledge. Now you're going to prompt it like go do this, go get this done. And that's what's most exciting.

12:43Corey:It's so exciting. It's so exciting. And I, you know, even in the last 60 days myself, I feel like every time I'm using these or I'm talking to my, you know, my desktop open claw rig at home through Telegram, I think, man, this is a this is different fast. And I look forward to the proactive end. Like you mentioned, I'm a busy guy. I forget things. There's so much stuff that I could get done if I just had someone to poke me occasionally and give me a reminder and not a reminder app. I need to go set up like I want it to just know me and have my calendar. Exactly. We had a developer actually at NVIDIA and he had set up his open claw machine and he has a lot of sensors around his house.

13:23And so he had gone away for the weekend and it was monitoring all the sensors like his sprinkler system. And it saw that the water gauge was going up even though he was out. And it WhatsAppped him and texted him and said, hey, the water is, there might be a leak. Should I email a plumber? And so he wrote back, yeah, here's my plumber's name. It crafted the email, sent him the draft and said, can I send it? He said, yes, the plumber came out. Sure enough, there was a leak in the sprinkler system. Did the plumber call and ask if it was a scam? But I mean, it's like that's the thing, the proactive, you know, doing this rather than finding it, you know, a month later and your water bill is super high.

13:59Corey:Yeah. Well, I'd like to talk about NemoClaw a little bit, too, because I think the idea that that NVIDIA has this this vision that this needs to be a thing that's secure, that businesses can use. and in a quick turnaround has put together a system that's, my understanding is it's more of a layer for OpenClaw than a, is that correct? Yeah, it's more of a runtime. So it does, you know, first of all, you lock it down so the agent has no ability or authority at first, right? And so you slowly give it access to things. But what's interesting is that not just the runtime, but it does, not just setting up guardrails, but also this policy routing.

14:43Or if it knows it needs to go do something, instead of leaking my information about who I am, be able to on the fly use synthetic data generation in order to say, hey, this is describing the situation or the person, but you don't actually need to know who Kerry is and kind of go get this done. So it's kind of bringing together the know-how of the developers were security experts at one point at NVIDIA. And that's why, again, no one person or companies can figure out a loan. They're putting out in the open. Again, it's got guardrails. It's got policy. It's got synthetic data generation. And it's got this runtime.

15:16And Peter's been really great to work with the team. They were whiteboarding over the weekend at headquarters. Yeah, he came.

15:21Corey:I fanboyed out when I got to meet him at the Build-A-Clock tent. I was like, oh, it's Peter. I'm so excited. Yeah, really, really nice guy. So I think that just like we've done with PyTorch or Kubernetes or OpenGL, we kind of had to explain that to Peter like we're not we want your authenticity this is your project it's in the open you've given it to the open let's work together and contribute to it and you know where do you need the most help and he's like oh I've got like you know 2 ,000 pull requests definitely pull requests yeah that's really cool and I love that as other companies have gotten involved in NVIDIA with touching it that there has been this attention to this is a very personal thing that we don't want to come in and, and, and like muck up.

16:08Corey:We want to, we want to keep it what it is, but help you make it better. Yeah, exactly. Yeah. It's so awesome. Um, so as has always been the tradition with NVIDIA, uh, super three, the new voice, the new speech models, uh, Nima claw, everything's very open. And I'm wondering how does that benefit enterprises? Are enterprises is seeing a real value proposition out of open source yet? Yeah, I think of two ways that enterprise has seen value out of open source. One, when we put everything out in the open, one, it was, you know, I like to say trust through transparency because they said, okay, you're showing your data sets.

16:45They showed how you trained it. I can go interrogate the data sets and see, or I can add them to my data sets. And so they like that. The other way that we're seeing enterprises embrace open is that I think open models are kind of having their moment because enterprises are starting to deploy in production AI. And so just with any software development platform, you put out a piece of software, you watch how it runs, you observe it, and you start to optimize for performance, right? And so you get features, you get bugs. It's the same thing. When you put an LLM or an agent or a system of agents out there, you start to observe it and say, okay, I can probably, I know that data flywheel, get this data, retrain on a smaller model.

17:25But then you're able to use that system of models of the open model, train it, and then grow with even more workloads, new types of use cases, right? Because now you're optimizing off. And so we've seen that the open model token generation has gone up about 35x across the board over the last year. And, you know, again, we work with all these AI companies that were able to kind of say, hey, how many tokens are we generated by open models? But Open Router, I don't know if you're familiar with that. It's like a switchboard for developers to be able to, right? Proprietary and Open Models, they've seen it exactly kind of grow almost to the T, that 35X.

18:04Oh, wow. Yeah. And so there are a lot of tokens being generated by Open Models right now. And we will continue to see that.

18:10Corey:I was in a conversation yesterday where we were trying to just guesstimate what token usage will look like in 2030. And I was like, if we're in that. I said, I saw another company say they're around five trillion in a day or so. I said, we're going to be in the quadrillions, the quintillions, the sextions, these numbers that sound funny. It sounds funny. It was so funny because we were putting together keynote facts and just like trying to put our head around like what is happening. Like how do we describe what that is without sounding like it's a funny number? Like how can you like really feel it?

18:45You know, it was things like, well, we can compare it to the Library of Congress. It's like writing 30 million books every single day. It was just like, how do you put it into context for people to feel like a quadrillion, a billion? Exponential doesn't do it for most people. Yeah, exactly.

19:01Corey:That's so neat. something I just briefly want to touch on is that I saw a lot of conversation the day that 3Super dropped about how this was kind of the U.S. finding its way back into being a legitimate push in the open source community around AI because it's been a very honestly China-led direction for the last couple of years. Yeah. Um, was that a, an intentional thing that like, you know, we need to get back in this race or was that just the nature of development? I think it's the nature of development. Um, you know, again, we like to work and support, we, we love open models. I think there's another great model, open model out there for open class called a Cree.