NVIDIA GTC 2026 Open Models Panel Highlights with Jensen Huang

0:12 Everybody, welcome!

0:14 Great to see all of you.

0:15 I have a special treat for you.

0:18 We have two sessions.

0:19 We have so many great speakers for you.

0:21 We broke it up into two sessions.

0:26 Let's get right to it.

0:28 I think we love a world where there's proprietary products,

0:32 but we also we also need a world where,

0:36 a whole bunch of companies and different industries

0:40 in different domains need models as a technology,

0:45 that we could then transform into products.

0:47 So today, really what I wanted to celebrate is not the proprietary versus open,

0:56 because I don't think that that's a thing.

0:58 Proprietary versus open is not a thing.

1:00 It's proprietary and open.

1:02 And we we love the fact that there

1:04 are these terrific models that are really products.

1:09 Folks think that there are only two types

1:12 of companies up at the software level of AI.

1:16 I think that they think that there are

1:18 foundation model companies that build very large general models,

1:22 sell access to those models through APIs,

1:25 do lots of products in different verticals and then there

1:27 are application companies that don't really do much AI work themselves,

1:31 but they build great products on top of the models.

1:33 And I think that we are really seeing the birth and flourishing

1:37 of a third type of company that sits in between the two.

1:42 A company that uses the best the market has to offer at an API level,

1:46 but then also does great work on the modeling front too,

1:50 and takes both the best the market has to offer on an API level,

1:55 their own models, and wraps them all into one,

1:58 into the best product for a certain vertical.

2:01 And in particular, I think we're seeing the rise of a new

2:05 type of agent happening over the course of the next year or two,

2:08 where the way you used AI models, it started as you're just calling a model.

2:15 And then it got a little bit more complicated.

2:16 It became calling a model, and that model has a bunch of tools that it can use.

2:21 And I think we're soon going to see agents really be coworkers that can

2:25 take on tasks that take many hours

2:28 or many days and do incredibly complex workloads.

2:31 We've experimented with things in this domain,

2:34 in software, where we've experimented with building,

2:36 for instance, prototype browsers from scratch over the course of many weeks,

2:40 entirely intent with agents.

2:42 And when you start to get

2:43 to these much more complicated workloads, under the hood,

2:46 you want to be farming out that workload to different models,

2:49 because different models have different strengths.

2:51 And they're going to be times when you want to use the computer,

2:53 use abilities of a foundation model from one of the APIs.

2:57 And there's going to be other times

2:58 where you want to use the industry-specific intelligence

3:01 you have as a company that's focusing on one domain to format your own models.

3:06 And so I think we're going to see the rise of these compound agents

3:08 that can be smarter than any one model on their own and mix them all together.

3:12 As Jensen said, AI is not the model—it's the system.

3:17 It's the computer.

3:20 Perplexity computer is the idea that you should

3:24 build the orchestration system of everything AI can do.

3:27 Every single capability—coding, writing, generating multimodal content.

3:34 So what you want is a multimodal,

3:39 multimodel and obviously multi-cloud orchestra of every

3:43 single two-model file system connectors put together.

3:49 So that all you got to do is delegate your task.

3:52 You don't have to worry about which model is good at what.

3:54 It's for the orchestration system to figure it out.

3:57 These sub-agents are like musicians, and the models are just instruments,

4:02 and the work that AI gets done for you

4:05 is the symphony or the music that they play.

4:07 And that makes it all pretty simple.

4:12 And, you know, you get to basically think of any task that AI can do today,

4:17 any different model.

4:18 You don't have to feel any vendor lock-ins.

4:20 And as Jensen was saying at the beginning— It doesn't

4:23 have to be a dichotomy between open models and closed.

4:27 We have open models in computer.

4:31 We have closed models in computer too.

4:33 And there are different needs for each of them.

4:36 Open models tend to excel at being very token efficient, cost efficient.

4:43 Closed models are very good at orchestration, reasoning, two calls,

4:47 and basically what's happening is models are essentially

4:51 becoming just tools like file systems and connectors.

4:54 And we're able to operate at an abstraction about models.

4:59 Finally.

5:00 The first one is that model companies are not actually model companies.

5:04 Like, they don't just build a model.

5:06 They build a whole stack, end-to-end.

5:08 So when you're buying a model from a proprietary or otherwise,

5:13 you're actually buying the chips, this orchestration software,

5:17 the inference and product—and all of that has been optimized end-to-end.

5:22 Now that yields great products.

5:25 And what openness allows is for other

5:28 people to basically optimize the whole thing end-to-end.

5:30 So you're not just buying a model— you're buying a whole system.

5:34 I think the other big misconception is on open models,

5:38 that somehow open models are fundamentally going

5:42 to be behind the frontier, that, you know,

5:44 the hit that you take when you adopt an open model is you get control,

5:49 but you get something that’s a few months behind.

5:51 I think that's just an artifact of the time where we are today.

5:55 There's nothing fundamentally different between an open and a closed model.

6:00 And, you know, models in general in this system,

6:05 these are it's fundamental knowledge infrastructure.

6:09 And fundamental knowledge Infrastructure yearns for openness.

6:13 You know, like a like an animal, you know,

6:15 it yearns for the hills or for the forest, like it wants to be open.

6:19 Like books used to be closed, and the printing press made them open.

6:22 Science used to be you know, the real alchemists,

6:26 and then the scientific journal made it open.

6:28 And strong encryption, there was a debate between closed and openness.

6:32 And actually, it ended up being both that started out closed,

6:36 and then there's a whole flourishing

6:38 ecosystem of strong encryption that became open.

6:40 And I think the same thing is about to happen in AI,

6:42 where there's a flourishing ecosystem of powerful,

6:45 closed models but equally capable open models that are

6:48 going to be coming over the next couple of years.

6:50 I think that's really exciting.

6:52 One last thing I did is that, progress is extremely fast.

6:56 We are on an exponential and everything is very compressed,

7:00 and there is a lot to learn.

7:01 There's a lot of study to be done,

7:04 and it cannot be done completely in the large labs.

7:09 Because there are tons of smart people out there.

7:11 But they're lacking access to knowledge, access to tools.

7:15 And this is where openness can be very helpful.

7:19 And it doesn't have to be just models— also infrastructure data,

7:23 general research insights.

7:25 And this can enable a ton of people

7:27 out there that can study various aspects of research.

7:32 And it advances the science of AI, science of intelligence.

7:36 I see this as a very positive sum.

7:39 Pretraining is memorization and generalization and some basic knowledge.

7:44 That basic knowledge gives you the foundation to go learn skills.

7:48 If you didn't have that basic knowledge, you wouldn't even be able to teach.

7:51 I mean, you know, you can't teach someone how to be an engineer

7:56 if they don't have any basic skills in math and science and, you know,

8:00 some basic understanding of technology.

8:01 And so the pretraining part was just the beginning.

8:04 Most people misunderstand that, in fact,

8:07 all of your labs do enormous amounts of model development,

8:12 particularly in post-training and the post-training area.

8:16 If you were to think about the amount of computing scope in the future,

8:19 the amount of computing use in pretraining was like 90% of training two,

8:23 three, or five years ago.

8:25 But in the future, the amount of training

8:27 percentage in pretraining is probably going to be tiny.

8:30 It's going to be mostly post-training.

8:33 It is also probably the case that these proprietary

8:38 models are going to be the best generalists.

8:43 But it's very unlikely they're the best specialists.

8:46 And most value is derived from specialists.

8:50 We need generalist capability all the time.

8:53 And they're going to get better and better.

8:55 And as you say, when we integrate them into a system,

8:58 you get the benefit of an insane generalist as well as an incredible specialist.

9:03 The first inflection point that I believe

9:06 at least I saw that basically made me switch

9:08 from theoretical physics to AI was a system

9:11 that my co-founder Ioannis has helped build called AlphaGo,

9:15 which was the first super- intelligent agent at scale,

9:20 and it was a 60 million parameter network.

9:23 It was tiny, relative to what it is now.

9:26 And it beat the best kind of player in the world at Go.

9:30 And the thing is, that system never stop learning.

9:33 It was just an economics problem.

9:35 How much compute are you willing to put in to get it,

9:37 you know, to be 10 times better?

9:39 And at some point, they cut it off because, you know,

9:41 would you put 10 billion more dollars

9:43 to make AlphaGo beat Lee Sedol even harder?

9:46 Probably not.

9:48 But you know, now that RL has started working on language models,

9:52 those are the kinds of questions we'll be asking now,

9:55 and a year from now, of am I willing to put in $10 billion,

10:00 $100 billion to solve, to cure a particular disease, right?

10:04 When you have RL working at scale, the things that you can solve,

10:08 because these are mechanical brains with endless capacity to learn,

10:12 become just a matter of economics.

10:14 And we're seeing the first generation of that in coding,

10:17 in agentic kind of enterprise applications.

10:19 But we will be coming to a point

10:22 where there's kind of fundamental scientific problems,

10:24 and we're just making economic decisions of whether we

10:27 want to allocate those resources to have a breakthrough.

10:30 And now we're at a point where models are extremely capable.

10:33 And in order to unlock usefulness,

10:35 you need to work on this orthogonal capabilities—connecting the context,

10:40 being able to operate within your data, within,

10:44 you know, what you're trying to do in your domain.

10:47 And this is an example of that, having an agent

10:50 that sort of can operate in your domain within your data,

10:54 connecting everything together, building this system that's more of an operator.

10:59 And we have a long way to go to actually make this very

11:02 reliable and take actions that you want to do with intent that you have.

11:09 But it sort of like shows that capability and where we need to go,

11:15 in terms of unlocking more usefulness.

11:17 The models and the systems orchestrating the models

11:21 are going to get much more capable.

11:22 And so you'll be able to have personal productivity agents

11:25 that can take on more complex tasks that run for longer.

11:28 And I think that also these compound agents

11:31 they will get much better at using tools.

11:34 Open models is how we got known, with Mistral 7B in 2023.

11:39 And effectively, we started Mistral to create an enterprise

11:43 business on top of an open model foundation.

11:46 And we've been lucky to work with you on Mistral Nemo, which was a 2024 release.

11:52 Which is quite good.

11:53 And the reason why we believe that open wide models should actually be

11:57 the basis for building all the AI software in the world is really two things.

12:02 The first is control, because the AI agent sits at the execution layer.

12:06 You want to have control over where this gets deployed.

12:10 You want to make sure that you have the turn-on

12:12 button to all agents that you're deploying in your company.

12:14 And so if you actually own the systems entirely,

12:17 so from the models to the orchestration layer,

12:19 and you know what code is actually running and you can actually modify it,

12:23 suddenly, you're much more confident than if you're depending on only APIs

12:27 that can be turned off or it can fall into certain problems.

12:30 So that resilience is actually very important.

12:33 The second reason is really customization.

12:36 Agents are great when they're operating in the virtual world,

12:38 where they're operating on knowledge.

12:40 They are great on health care because they see a lot of documents.

12:43 But whenever you have something where you

12:45 have a physical footprint, you have a machine,

12:48 you want to model a physical system,

12:50 you have an engineering team that is actually to build things,

12:53 in the physical world.

12:54 Well, suddenly you have a lot of IP

12:57 that you want to put into the models themselves.

12:59 And it's never going to go back to the general

13:01 purpose models that are trained fully on the virtual world.

13:04 And so you suddenly you have access to models

13:06 that you can modify to whom you can plug, time series, stream, etc.

13:11 You can actually build agents that understand

13:13 the physical world and actually make them useful

13:16 for teams like engineering teams that currently

13:19 do not benefit a lot from generative AI.

13:21 So I think that customization aspect, we've released the product called Forged

13:25 that is precisely meant to connect models

13:27 to various sources of data that may come from the physical world is,

13:31 the second reason why open models are very important.

13:35 And I would say, like, finally, open models are great because it allows

13:38 us to make cheaper versions of everything,

13:41 by working together, by sharing R&D cost.

13:44 That's the new neutron and coalition that we're

13:46 happy to be part of, by having an open

13:49 ecosystem of people that have aligned incentives to create

13:52 assets that are going to be great for humanity.

13:55 We can actually accelerate progress and make sure

13:58 that everybody gets access in a fair way,

14:00 across the world, to artificial intelligence.

14:02 And so that's why I think it matters a lot.

14:05 Open models are strictly better than closed models.

14:08 Because I actually think in some contexts, closed models are great.

14:11 But there is one context in which open models are extraordinary.

14:15 If you think about all the applications that you guys just mentioned,

14:19 they're increasingly mission critical.

14:21 You know, 3 or 4 years ago, I think the questions we often used to get is,

14:25 will AI be useful for anything more than chat bots?

14:28 And now because of RL and agentic infrastructure,

14:33 these are getting useful at mission-critical applications.

14:36 And the thing about mission critical applications is you’ve got to trust

14:40 that they will perform the way you do in a high-stakes situation,

14:43 where the margin of error is very low.

14:46 And the reality is, if it's not an open model,

14:49 and you don't control where, you don't know where,

14:51 you can't introspect it, you can't host it,

14:54 and you're dependent on third parties for that.

14:56 Hey, they're good at some things,

14:58 but at the end of the day, you're delegating trust.

15:01 And I think that will become more and more a topic of conversation.

15:04 As those of you who are deploying agents in critical industries,

15:08 I think we will come back to trust.

15:10 And it's much easier to trust an open

15:13 system that you can introspect and you can say,

15:15 look, this parts of this system are closed and they're good.

15:17 It's good at that.

15:18 I'm going to deploy it.

15:19 and I'm gonna have some guardrails.

15:21 I'm going to manage the risk.

15:22 But at the end of the day, especially in healthcare, defense, you know,

15:28 places that if we want to welcome these agents

15:30 into the most mission-critical parts of our lives,

15:32 I think we're going to have to find a way to trust them.

15:35 And as far as I know, open models are one of the fastest ways to trust a system.

15:42 Now, the flip side is, I think the infrastructure also has to be open.

15:46 And that's not happening as quickly as it should.

15:49 We are heading into a phase

15:50 of the industry where infrastructure is consolidating very fast.

15:53 If you study the history of the industrial revolution—you asked us to consider

15:56 the industrial applications— and I think we are in a new era.

15:59 But there's some clues that the the 1800s gave us,

16:02 where if you looked at factories that had a steam

16:05 engine and knew that you could make interesting products,

16:08 they were starting to realize that, hey,

16:10 the inputs to production are pretty critical.

16:13 And they started hoarding them.

16:14 And if you just do a flyby of 1800 Victorian England,

16:18 you'll see factories with generators running at half capacity,

16:21 because everyone's hoarding generators and trying to run their own generator,

16:24 and you have piles of coal just stockpiling, not being used.

16:28 And I think what we need to go to is open models and open infrastructure.

16:33 We need to develop a mechanism that allows us to share,

16:37 not to overprovision per peak,

16:40 but have enough secure infrastructure that we get our baseload,

16:44 but then they can spike up and down.

16:46 And that's called a grid.

16:47 And that's what we're working at AMP.

16:48 As you know, we're building an AI grid.

16:50 But I really do think for open models to be developed at the frontier,

16:54 we're going to need open infrastructure too.

16:56 When we have a language model or an AI system these days,

17:00 we just look at the final snapshot of this whole very long,

17:05 curated process, of how these models are developed.

17:08 However, we argue that or I argue that there

17:12 is a long process and releasing different parts of it,

17:16 enabling researchers and developers to use different parts of it,

17:20 it enables infinite customization,

17:22 not just the final way, but also on top of different checkpoints of the model.

17:27 So my team built Olmo, and through that, we developed model flow,

17:31 which we call it the full development

17:33 cycle of the whole process, including the data,

17:36 all the model weights, all the infrastructure, and everything,

17:39 so that researchers on hard-core developers could actually play with it,

17:44 integrate their finding, and even build on top of that.

17:47 So why is this important?

17:49 So this is particularly really important to make

17:52 advances in the next generation of AI.

17:55 Because these days, you might argue that the progress

17:58 in AI is getting limited into a few close labs,

18:02 but it's actually very important to the vast

18:05 majority of academia and researchers or kind

18:08 of nonprofit and other places who want to also be part of this progress.

18:13 And we've seen that these all this progress

18:16 already has happened by everything being open.

18:19 So for example, you've probably all seen that hybrid

18:23 models are being used in many of these open models,

18:26 and even closed model like Nemotron, a very good example of that.

18:30 So most recently, we—actually our team— has started releasing and studying

18:34 why hybrid models are probably only better than only transformers.

18:39 And based on analysis and more theoretical findings,

18:42 we actually found that okay,

18:44 there are theoretical reasons why hybrid models are better.

18:48 And then empirically we show that yes— it actually is much more efficient,

18:52 much more token efficient in training.

18:53 So having access to this infrastructure, training,

18:56 and being able to kind of publicly talk about it, and also,

19:01 kind of study these models,

19:03 enable the next generation of even model architecture,

19:06 better AI systems, and so on.

19:08 So enabling research is also a very important topic.

19:11 Open models are strictly better than closed models.

19:14 I think they enable diffusion of innovation, research, and competition.

19:19 I guess like competition in the end is good for like for you.

19:22 Right?

19:22 And, in addition to that, it's really also,

19:26 I don't know, l’m here in Silicon Valley,

19:29 every two months, then I'm based in Germany for my usual day-to-day work.

19:34 And I come here now and just like this weird

19:36 AI psychosis going on, where people even like AI researchers,

19:40 think that they cannot really contribute anymore because

19:43 of the advances in coding agents, stuff like that.

19:47 But I think even if you're, like,

19:50 really convinced of that, then like the way that you can have impact should be

19:55 thinking about how can I make an open

19:57 model that actually can replicate these capabilities.

19:59 So I think, it's actually one of the most exciting times to work on, you know,

20:04 like the frontier models,

20:05 the big models, or more specialized open models that then get deployed,

20:09 like on device and all that.

20:10 And there's like so many different frontiers,

20:13 and all of them should have some open component.

20:16 Even in a company that is a closed model company,

20:20 I really believe that open models will be used as part of the agentic system,

20:26 where the closed model is your crown jewels.

20:30 And I'm fairly certain that even in visual

20:33 intelligence that used to be text-to-image in the future.

20:37 well, you're already there now— increasingly,

20:41 it's in robotic visual intelligence system.

20:45 and so it's reasoning, it’s trying, it’s taking inputs,

20:49 and the prompts could be text and many other conditionings.

20:52 It's solving problems and ultimately generating an image or a stream

20:59 of images that you can even interact with in a very robotic way.

21:02 And so I think the future of visual

21:05 intelligence that you're working on is just so exciting.

21:08 And even though it's a proprietary crown jewels today,

21:11 you're going to surround it with a whole bunch of open models.

21:13 It's my prediction.

21:15 And so all of these different industries are starting

21:18 to get solutions that are very impactful and very capable.

21:21 And so I think that this year we're going to see,

21:24 really instead of the question, “What is the ROI of, you know,

21:29 the last three years?” All of the questions are,

21:31 “What are the ROIs’ of AI?” I think now,

21:34 this year, starting with with coding, of course.

21:37 And coding is not just software engineering.

21:39 Coding is the description, the codifying of business processes,

21:43 codifying basically the rules of all businesses

21:47 and all engineering and almost all work.

21:50 And so coding is a very big deal.

21:52 But nonetheless, we're going to see an inflection

21:54 this year of real business economics taking off.

21:58 And you guys are all experiencing that.

22:00 Ladies and gentlemen, hey, let's get the last crew back up here.

22:07 That’s it.

22:08 The future!

22:11 Thank you, guys.

22:12 Good job!