Transcript
0:00 Mr. President? Oh, yes, sir. I gotta tell you something. It I I if it wasn't because of you calling, I would I'm on stage with the besties. I'm on stage with the besties. Jensen Huang is midway through an all-in summit interview when President Trump calls. Huang tells him he is on stage with the hosts in front of an audience of thousands. They have been debating AI safety, open models, and the pace of development. Within moments, the phone is on speaker, and the president joins the discussion. His first observation is about the gap between designing a computer chip and operating a phone.
0:45 You're now talking to the planet. You see, the great thing about life is that Jensen can develop the most complex computer chip in the world that nobody can copy for 10 years. But he can't figure out how to put me on SPEAKER Trump uses the call to argue for continued AI development and data center construction. He treats the build-out as a source of economic power and frames the competition internationally. He also says development should proceed prudently. The exchange puts Huang's argument in a much larger arena. Decisions about AI now reach from research labs to the infrastructure that governments and communities are being asked to accommodate.
1:28 Well, I know about AI. I know I also have common sense about AI. the robots will not be taking over. the AI will not be taking over the rest of the world. The whole thing is a hoax. Now, with that, we have to be a little bit careful. We have to very be, you know, we have to do things and we have to do them prudently. But that doesn't mean we're going to stop industry because, you know, as we work on the next 10 years about how to destroy it.
1:54 So, I'm with you all the way. I didn't even know how you felt about it. And I assumed you felt the same way as me. Yes, sir. And we if we're going to lead and I have an expression, it's whoever wins AI wins. That's how big it is. It's bigger than the internet. And whoever wins AI wins. Huang's answer to the safety debate starts with how systems are built and controlled. Earlier in the interview, he calls safety paramount.
2:17 He argues that labs moving from research into production need stronger engineering practices. That puts the focus on concrete incidents, what failed, how the failure happened, and what change would prevent a repeat. His confidence is a judgment about what engineering can achieve. demonstrating that confidence requires evidence from the systems themselves, especially when those systems are given more authority to act. the four incidents from one lab, the one giant incident from the other lab, the first thing that you have to do is just root cause the problem from an engineering perspective. what happened, what could we have done differently and what are we going to in to implement and institutionalize whether it's technology or methods or processes and make sure that we don't let it happen again. Now, I would bet you money that in every single one of those cases is within their control in the future to prevent it The practical question is whether an organization can show that its controls work. A useful incident review would identify the permissions an AI had, the signals that were missed, and the test that now catches the same failure. Huang later supports independent evaluators with several providers to reduce the influence of any one organization. The same logic appears in his answer about recursive self-improvement. When AI helps produce the next version of AI, the release still needs an evaluation process.
3:47 No. No, of course not. And the reason for that is because you could RSI all day long inside your company, but when you release a product, you've got to evaluate it, don't you? You have to test it again, don't you? You have to make sure that there's no regression, right? And so the basic process of control. These labs are going to as they move from labs to engineering, they will have much much better control, right? And when they have much better control that and control comes from methods and knowledge and practice and tools and technology all of those things that leads to better control verification and evals it's going to enable RSI to be done inside the company and for good products to be released outside.
4:29 Huang describes a feedback loop in which AI improves the work of building AI. He lists techniques such as reflection, reinforcement learning, and synthetic data. For a company using those tools, the important distinction is between a promising internal result and a version ready for other people to rely on. The release gate must measure capability and failure modes together. As development accelerates, the quality and independence of those tests become increasingly valuable. Huang then turns to who gets access to the technology.
5:02 The world needs both closed models and open models. you want you want to use I use as much closed models as I can. This weekend I I used four of them and and they work terrifically. They're frontier. They're great experience. They right they work incredibly well. They're getting better all the time. and and the way I think about closed closed closed models is kind of like bottled water. You know, water is free, you guys. I don't know if I've told you guys, but water is free. I I don't want to, you know, burst everybody's bubble, but water's free.
5:33 And this morning, I used a lot of free water taking a shower. And so, you use the right water in the right places. And this is no different than electricity. This is, you know, this is no different than all kinds of commodities that we use in the world. You need both. The bottled water analogy explains why Huang expects paid services and open models to coexist. A hosted service can offer convenience and a polished experience. Open models can give a team more control over adaptation and deployment. Huang specifically points to privacy, sovereignty, and proprietary technology. A sensible choice starts with the job. What information will the model handle? How much customization is needed? And who can operate it reliably? Running a model also requires infrastructure and engineering.
6:20 Access to the weights is only one part of that decision. So what exactly is the race? the race. Yeah, I think that's that's a really good point. My point is the race is really about who exploits the technology best. You know, the last industrial revolution, all of the inventors were Maxwell, Volta, Ampère. None of them were American. They were right. The last industrial revolution came from Europe. But we exploited it. We took advantage of it socially better than anybody else in the world. Look how it turned out for us. I want to make sure that this next generation happens just like this.
7:00 Yeah. Yeah. Huang measures the AI race through adoption across an economy. That makes the number of useful deployments matter alongside the quality of a leading model. A business gets value when a tool improves an actual workflow. Identifying a problem, connecting the right information, checking the result, and putting it into use. The same idea helps explain NVIDIA's strategy. More builders and more applications can create more demand for computing. Huang's argument for broad access, therefore, connects directly with the business that supplies that computing.
7:34 Now, in order for that to happen, it's got to produce the intelligence. And so that's a production process which is the reason why this infrastructure has to get built. But once you get the infrastructure built, the question is what about all of the other layers across the United States? this industry isn't just about the model. It's not just about the chips. It's mostly about the applications on top. It's mostly about the infrastructure layer, the data centers and all the infrastructure, the the the the construction, the electricity, the power generation that all of that is involved.
8:07 And so I look across the entire ecosystem and look for bottlenecks and if there are places where extraordinary companies are being built constraints extraordinary companies being built maybe it's supply chain that has to get scaled up so that when we're ready to deploy compute that they'll be ready for us land, power, shell. A powerful chip creates value when the surrounding system can use it. Huang describes looking for constraints across land, power, buildings, manufacturing, and the companies that supply them. A delay in one layer can hold back capacity in another. Supporting the wider ecosystem can help more of NVIDIA's technology reach working deployments. It also links the company's progress to partners executing on very different schedules. The questions become concrete. When will power be available? When will a site be ready?
9:03 And when can customers put the installed capacity to work? And so all of these labs are building on NVIDIA. And the reason for that is because as a company, I rather for us to help everybody succeed instead of taking a slice out. And so we would go up as far as we need to but as low as possible. Our strategy is go up as far as we need to and as low as possible. And the reason for that is because if I do that, if I solved the if if not for Nvidia creating cuDNN, all of the frameworks wouldn't exist. If not for us creating Megatron megatron core, then all of the large scale training wouldn't have happened.
9:46 Wouldn't exist. so we we go and we invent all the technology necessary as far as we need to and then we let a thousand flowers bloom. Huang says NVIDIA goes as far up the technology stack as necessary while leaving room for customers to build their own products. He gives examples of enabling software and later models for specific industries. The logic is to solve a shared technical problem so many companies can develop applications on top. That is a strategy with a built-in tension. Customers need a capable platform and they also want space to differentiate. NVIDIA's partnerships will depend on how it manages that boundary as its software and model ambitions expand.
10:26 Well, I mean it it feels apparent, I think, to most of us in the industry that we're kind of in the AGI moment. And it's a definition obviously just as smart as any other human. I think we're already there. We're there, right? And so then super intelligence is the next way point based on what you see, based on your customer base, based on your history here. But Jason, I think we're there, too. You think we're at super intelligence?
10:47 Yeah. Yeah. When you when you when you take a narrow segment I mean my my self-driving car I don't want you to make me an omelette I just want you to drive the car right that is super intelligent super it's better it's better than a human yeah yeah onetenth the the accident rate exactly synthesizing proteins you know doing virtual screening of proteins we're already there Huang ends with examples of superhuman performance in specific tasks, including driving and work with proteins. His definition makes the task boundary essential. What can the system do, under which conditions, and how reliably? That question ties the interview together. Stronger capabilities need better controls. Wider adoption needs useful products. More computing needs functioning infrastructure. The surprise phone call makes the stakes unusually visible. But the work Huang describes runs through every layer, research, engineering, deployment, and the people who ultimately use the result.
11:50 Follow us for AI News That Matters.