NVIDIA’s Jensen Huang on Reasoning Models, Robotics, and Refuting the “AI Bubble” Narrative (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:06) Hson, thanks so much for joining us (00:00:07) today. (00:00:08) >> So great to have you guys. What an (00:00:10) amazing year. (00:00:11) >> What a year. (00:00:12) >> Happy Hanukkah, merry Christmas, (00:00:14) >> happy new year coming up. Yep. Happy (00:00:15) holidays. (00:00:16) >> So, uh, with everything that's happened (00:00:19) in 2025, (00:00:21) um, and you know, being in the middle of (00:00:23) the vortex with it, what do you reflect (00:00:24) on and say like this surprised you most (00:00:26) or this is the biggest change? (00:00:28) >> Let's see. There there's some things (00:00:29) that didn't surprise me like for example (00:00:31) the scaling laws didn't surprise me (00:00:32) because we already knew about that. The (00:00:34) technology advancement didn't surprise (00:00:36) me. I was pleased with the improvements (00:00:39) of grounding. I was pleased with the (00:00:41) improvements of reasoning. I was pleased (00:00:43) with uh uh the connection of all of the (00:00:46) models to to to search. I'm pleased that (00:00:50) it that uh there are now routers that (00:00:53) are in front of these models so that it (00:00:55) could depending on the confidence of the (00:00:57) answers go off and do necessary research (00:01:00) and and just generally improve the (00:01:03) quality and the accuracy of answers. (00:01:05) >> I'm hugely proud of that. I think the (00:01:08) whole industry addressed one of the (00:01:10) biggest skeptical responses of AI which (00:01:13) is hallucination and um generating (00:01:16) gibberish and all of that stuff. I I (00:01:18) thought that this year the whole (00:01:20) industry everything from every every (00:01:22) field from language to vision to (00:01:25) robotics to self-driving cars the the (00:01:28) application of reasoning (00:01:31) and the grounding of the of of of the (00:01:35) answers. Um big big leaps would you guys (00:01:38) say this year? (00:01:39) >> Huge. I mean things like open evidence (00:01:41) too for medical information where (00:01:42) doctors are now really using that as a (00:01:43) trusted resource like you Harvey for (00:01:46) legal you're really starting to see AI (00:01:47) emerge as one of these things that's (00:01:49) become a trusted tool or counterparty (00:01:51) for you know experts to actually be able (00:01:53) to do what they do much better. (00:01:55) >> That's that's right. And so so in a lot (00:01:57) of ways I was expecting it but I'm still (00:01:59) pleased by it. I'm proud of it. I'm (00:02:01) proud of all of the industry's work in (00:02:03) this area. I'm really pleased and and uh (00:02:06) uh and probably a little bit surprised (00:02:09) in fact that token generation rate for (00:02:13) inference especially reasoning tokens (00:02:15) are growing so fast several exponentials (00:02:19) at the same times it seems and uh and (00:02:22) I'm so pleased that that these tokens (00:02:24) are now profitable that people are (00:02:26) generating I heard somebody hurt today (00:02:29) that that open evidence speaking of them (00:02:31) 90% gross margins I mean those are very (00:02:34) profitable tokens. (00:02:35) >> Yeah. (00:02:35) >> And so they're obviously doing very (00:02:37) profitable, very valuable work. Cursor, (00:02:39) their margins are great. Uh Claude's (00:02:42) margins are great for the enterprise use (00:02:44) of OpenAI. Their margins are great. Um (00:02:46) so anyways, it's really terrific to see (00:02:48) that that um we're now generating tokens (00:02:51) that are sufficiently good, so good in (00:02:54) value that that people are willing to (00:02:55) pay good money for. And so I I think (00:02:57) these are are really great grounding for (00:02:59) the year. I mean some of the things that (00:03:02) the narrative that that um uh of course (00:03:05) the conversation with China really (00:03:06) really you know occupied a lot of my my (00:03:09) time this year. Geopolitics (00:03:12) uh the importance of technology in each (00:03:13) one of the countries. Uh I spent more (00:03:15) time traveling around the world this (00:03:17) year than just about any time in the h (00:03:18) all of my life combined. You know my (00:03:21) average elevation this year is probably (00:03:23) about 17,000 ft. You know so so it's (00:03:25) nice to be here on the ground with you (00:03:27) guys. Um and so so I think uh (00:03:30) geopolitics the importance of AI to all (00:03:32) the nations uh all worth talking about (00:03:34) later. You know of course I spent a lot (00:03:36) of time on expert control and and making (00:03:38) sure that our strategy is nuanced and uh (00:03:41) really grounded and um uh promotes (00:03:44) national security but recognizing the (00:03:46) importance of various uh various facets (00:03:50) of national security. Um a lot of (00:03:52) conversations around that. Um, you know, (00:03:54) of course, of course, uh, lots of (00:03:56) conversation about jobs, the impact of (00:03:58) AI, uh, energy, (00:04:01) >> um, uh, labor shortage. I mean, boy, we (00:04:03) covered everything, did we? Yeah. (00:04:05) Everything was AI. (00:04:06) >> Everything was AI. Yeah, it was (00:04:08) incredible. (00:04:09) >> Yeah, AI was definitely the center of (00:04:10) the storm for like every one of those (00:04:12) themes. Maybe one we can start with (00:04:13) actually um is jobs because or there (00:04:16) jobs and employment because when I look (00:04:17) at the traditional AI community even (00:04:20) before things were scaling and even (00:04:22) before AI was really working there was a (00:04:24) strong sort of doomsday component in the (00:04:26) people working on AI oddly enough right (00:04:28) the people who were most trying to push (00:04:29) the field forward were often the people (00:04:30) who are most pessimistic which is very (00:04:32) odd why would you do both at once (00:04:33) >> and I feel like that narrative has taken (00:04:35) over some subset of media or some set of (00:04:37) other things despite all the things that (00:04:39) we think are very positive about what AI (00:04:41) has done That's going to help with (00:04:42) healthcare, with education, with (00:04:43) productivity, with all these other (00:04:45) areas. (00:04:46) >> And in in general, whenever we have a (00:04:48) technology shift, you have a shift in (00:04:50) terms of the jobs that are important, (00:04:51) but you still have more jobs. (00:04:52) >> That's right. (00:04:52) >> Could you talk about how you think about (00:04:53) employment and jobs and sort of what (00:04:55) people are saying and what you think the (00:04:56) real narrative is there? (00:04:57) >> Maybe what I'll do is I'll I'll ground (00:04:59) it on uh three points in space, three (00:05:02) points in time. now. (00:05:04) >> Mhm. (00:05:05) >> Uh maybe uh uh very near future and then (00:05:09) some some point out out in the distance (00:05:11) and and maybe maybe some (00:05:14) counternarratives. (00:05:15) >> Um something else to think about with (00:05:17) respect to jobs in the near term. (00:05:20) >> Uh one of the most important things is (00:05:21) that that AI is not just AI is software (00:05:26) >> but it's not pre-recorded software as (00:05:28) you know. For example, Excel was written (00:05:30) by several hundred engineers. They (00:05:32) compiled it. It's pre-recorded and then (00:05:34) they distribute it as is for several (00:05:36) years. In the case of AI, because it (00:05:39) takes into the context, what you asked (00:05:42) of it, what's happening in the world, (00:05:44) right? Contextual information, it (00:05:46) generates every single token for the (00:05:49) first time, every time. (00:05:50) >> Which means every time you use the (00:05:53) software and and everything that we do, (00:05:55) AI is being generated for the first time (00:05:57) ever. Just like intelligence, our (00:05:59) conversation today relies on some, you (00:06:02) know, ground truth and some knowledge (00:06:04) and but it's every single word is being (00:06:06) generated for the first time here. The (00:06:08) thing that's really really quite unique (00:06:10) about AI is that it needs these (00:06:13) computers to generate these tokens every (00:06:15) single time. I call them AI factories (00:06:17) because it's producing tokens that will (00:06:20) be, you know, used all over the world. (00:06:22) Now, some people would say it's also (00:06:23) part of infrastructure. The reason why (00:06:25) it's infrastructure is because obviously (00:06:27) it affects every single application. (00:06:29) It's used in every single company. It's (00:06:32) used in every single industry every (00:06:34) single country. Therefore, it's part (00:06:35) infrastructure like energy and and (00:06:37) internet. Now, because of that and the (00:06:40) amount of computers that's necessary to (00:06:41) generate these tokens and it's never (00:06:43) happened before and because we need (00:06:44) these factories, three new industries (00:06:47) have emerged. Number one, well, three (00:06:49) new type of plants have to be created. (00:06:50) Number one, we have to build a lot more (00:06:52) chip plants. (00:06:53) >> Mhm. TSMC is building, right? SKH Highix (00:06:56) building a lot more plants and so we (00:06:59) need more chip plants. We need more (00:07:01) computer plants. These computers are (00:07:03) very different. These are supercomputers (00:07:05) that the world's never seen before. (00:07:06) Right. Grace Blackwell looks like a very (00:07:09) different type of computer than anything (00:07:11) that's ever been made. And entire rack (00:07:13) is one GPU. (00:07:15) >> And so we need new supercomputer plants. (00:07:17) And then we need new AI factories. These (00:07:20) three plants are currently being met (00:07:22) being built in the United States at very (00:07:25) large scale quite broadly all over the (00:07:27) United States for the very first time. (00:07:29) >> The number of construction workers, (00:07:31) plumbers, electricians, technicians, (00:07:33) network engineers, you know, right? The (00:07:36) the number of the skilled labor that's (00:07:39) necessary to support this new industry (00:07:40) in the near term, (00:07:42) >> it'll be enormous. Let's just face it. (00:07:45) Uh I'm [clears throat] so excited to (00:07:46) hear that electricians are seeing their (00:07:48) paychecks double. They're being they're (00:07:50) being paid to travels like like us. We (00:07:52) go on business trips. They're going on (00:07:53) business trips. And so it's really (00:07:55) terrific to see, you know, that this (00:07:58) these three industries are now three (00:08:01) types of plants, factories are just (00:08:03) creating so much so much jobs. The next (00:08:06) part is the the near-term impact of AI (00:08:10) on jobs. And one of my favorites is um I (00:08:14) love Jeff Hinton. uh he said uh you know (00:08:18) some five six seven years ago that in (00:08:21) five years time uh AI will completely (00:08:25) revolutionize radiology that every (00:08:28) single radiology application will be (00:08:30) powered by AI and that radiologists (00:08:35) uh will no longer be needed and that he (00:08:38) would advise this the first profession (00:08:40) not to go into is radiology and he's (00:08:42) absolutely right 100% of radiology (00:08:46) applications are now AI powered. That's (00:08:49) completely true and in some eight years (00:08:51) time it is now completely pervaded uh uh (00:08:56) radiology. However, what's interesting (00:08:58) is that the number of radiologists (00:09:00) increased (00:09:01) >> and so now the question is why and this (00:09:04) is where the difference between task (00:09:07) versus purpose of a job. A job has tasks (00:09:10) and has purpose. And in the case of a (00:09:13) radiologist, the task is to study scans, (00:09:18) but the purpose is to diagnose disease (00:09:21) >> and to research (00:09:22) >> and and that exactly and they're doing (00:09:24) research. And so in the case in their (00:09:26) case, the fact that they're able to (00:09:28) study more scans more deeply, (00:09:31) um they're able to uh request more (00:09:33) scans, do a better job diagnosing (00:09:36) disease, the hospital's more productive, (00:09:39) they can have more patients, which (00:09:41) allows them to make more money, which (00:09:43) allows them to want to hire more (00:09:44) radiologists. And so the question is (00:09:47) what is the purpose of the job versus (00:09:50) what is the task that you do in your (00:09:52) job? And and as you know, I spend most (00:09:54) of my (00:09:55) >> day typing. [snorts] That's my task, but (00:09:58) my purpose is obviously not typing. And (00:10:01) so the fact that somebody could use AI (00:10:03) to automate a lot of my typing, and I (00:10:05) really appreciate that, and it helps a (00:10:06) lot. (00:10:07) >> Um, it hasn't really made me, if you (00:10:09) will, less busy. In a lot of ways, I (00:10:12) become more busy because I'm able to do (00:10:13) more work. So, I think that that's the (00:10:15) second part to consider is the task (00:10:17) versus the purpose of the job. This (00:10:19) example really strikes home because my (00:10:21) my sister-in-law Erin actually leads um (00:10:23) in nuclear medicine at Stanford, right? (00:10:25) So, she's in radiology and with all the (00:10:28) technology advancements that are coming, (00:10:30) >> these doctors really welcome it and they (00:10:32) are working 20 hours a day trying to do (00:10:34) more research and serve more patients. (00:10:36) Exactly. And and I think one thing that (00:10:38) is often missed beyond the sort of um uh (00:10:42) diversity of jobs being created by this (00:10:44) investment in infrastructure is actually (00:10:46) how much latent demand there is for (00:10:49) different goods that we we need in (00:10:52) society like better healthcare. I don't (00:10:53) think anybody feels like you know what (00:10:55) we have reached the the tiptop uh (00:10:59) mountaintop of like what American (00:11:00) healthcare or global healthcare could be (00:11:03) and um the more we can make these people (00:11:06) productive the more demand there will be (00:11:07) >> that's exactly right if I if Nvidia was (00:11:09) more productive it doesn't result in (00:11:12) layoffs it results in us doing more more (00:11:15) things (00:11:16) >> I met your new hire class today you seem (00:11:18) to be hiring every week anyway yeah (00:11:20) >> that's exactly right right the the more (00:11:22) productive we are the more uh ideas we (00:11:25) can explore uh the more growth as as a (00:11:28) result the more profitable we become (00:11:30) which allows us to pursue more ideas and (00:11:33) so I think you're you're absolutely (00:11:34) right that that if if the job if if your (00:11:37) if your life if the world the problems (00:11:40) is literally already specified and (00:11:43) there's no other problem to solve then (00:11:45) productivity would actually reduce the (00:11:48) economy but it's clearly going to (00:11:50) increase the e economy I think that the (00:11:52) Next part that I would consider is, you (00:11:55) know, people say, gosh, all of these (00:11:57) robots that we're talking about, it's (00:11:58) going to take away jobs. As as we know (00:12:00) very clearly, we don't have enough (00:12:03) factory workers. Our economy is actually (00:12:05) limited by the number of factory workers (00:12:07) we have. Most people are are having a (00:12:10) very hard time retaining their workers. (00:12:13) Um, we also know that the number of (00:12:15) truck drivers in the world is severely (00:12:18) short. And the reason for that is people (00:12:20) don't want those jobs where you have to (00:12:22) travel across the country and live in (00:12:24) different parts of the world, different (00:12:25) parts of the country, you know, every (00:12:26) single night. So people want to stay in (00:12:28) their town, stay with their families. (00:12:30) And so I think that I think the first (00:12:32) part is that having robotic systems is (00:12:36) going to allow us to cover the labor (00:12:39) shortage gap which is really really (00:12:41) severe and getting worse because of (00:12:43) aging population. This is this is not (00:12:45) only United States, it's all over the (00:12:47) world as you guys know. (00:12:48) >> And so we're going to cover the labor (00:12:51) shortage. But the second part that (00:12:53) people forget and and as a result we'll (00:12:55) go there are shortages as well in other (00:12:57) places that people talk about AI being (00:12:58) relevant. Accounting would be an example (00:13:00) where there's shortages there. Nursing (00:13:02) is another example. So you know you can (00:13:04) you can go through multiple other (00:13:05) industries and say okay there's gaps (00:13:07) right (00:13:07) >> and AI is trying to help fill those (00:13:08) gaps. (00:13:09) >> That's exactly right. And so so um (00:13:11) automation is going to help us increase (00:13:13) and solve the the the the labor gap. Now (00:13:17) people also don't don't remember that (00:13:19) when we have cars, we need mechanics to (00:13:23) take care of our cars. (00:13:25) >> And if you look at the robo taxis that (00:13:27) are that are even on the streets today, (00:13:29) it's taken 10 years for that to happen. (00:13:32) Look at all the maintenance crews and (00:13:34) all of the the the various, you know, (00:13:37) hubs that they're in where you have to (00:13:39) take care of these robo taxis and just (00:13:41) imagine we have a billion robots. (00:13:44) >> Mhm. (00:13:44) >> It's going to be the largest repair (00:13:46) industry on the planet. So I I think a (00:13:48) lot of people don't they they just have (00:13:49) to think through (00:13:50) >> and this is the part where you said um (00:13:52) when we create this type of automation, (00:13:54) we create this other job. Right now look (00:13:56) at AI is creating so many jobs. Mhm. (00:13:59) >> The AI industry is creating a boom of (00:14:01) jobs. (00:14:02) >> I think one of the core challenges here (00:14:03) is it's very easy to draw a straight (00:14:07) line of extrapolation from like oh you (00:14:10) know uh there are tools that help (00:14:12) lawyers be more productive. It's going (00:14:14) to replace the lawyers but it's actually (00:14:15) it takes like a step of incremental (00:14:18) reasoning to say there's a sucking sound (00:14:21) in the economy for everything in AI (00:14:22) infrastructure. there's actually a (00:14:24) sucking sound toward all of this demand (00:14:26) that is latent in the places where we (00:14:28) have gaps where um I think a lot of (00:14:31) policy makers have focused on you know (00:14:33) we can't replace or reduce what we have (00:14:35) when it's really there's there's far (00:14:37) more demand in what we actually are not (00:14:39) >> and in the case of lawyer what's the (00:14:41) what's the purpose of the lawyer versus (00:14:43) the task of the lawyer (00:14:45) >> reading a contract writing a contract is (00:14:48) not the purpose of the lawyer the (00:14:50) purpose of the lawyer is to help you (00:14:52) resolve conflict (00:14:54) And that's more than reading a contract. (00:14:56) It's more than writing a contract. The (00:14:59) purpose is to protect you. That's more (00:15:01) than reading a contract. It's more than (00:15:02) writing a contract. And so I think just (00:15:05) it's really really important to go back (00:15:06) to what is the purpose of the job versus (00:15:09) the task that we use, (00:15:12) >> you know, to perform that job that (00:15:13) changes over time. (00:15:14) >> Yeah. The other big theme of the year (00:15:15) that you mentioned that I think is (00:15:16) really important to touch upon is both (00:15:18) uh China is sort of in the rise of (00:15:20) Chinese open source in particular where (00:15:22) you know some of the highest scoring (00:15:24) models against benchmarks now are (00:15:26) Chinese models on the open source side (00:15:27) on the closer side it's still a lot of (00:15:28) the US models but things like Quinn (00:15:30) Deepseek etc (00:15:32) >> are doing very well you've long been a (00:15:34) proponent for open source in general (00:15:35) could you could you share views about (00:15:37) both China emerging for AI for open (00:15:40) source and what the US should be doing (00:15:41) in terms of both open source as well as (00:15:42) its own industries (00:15:44) >> when you Think about these complicated (00:15:46) interconnected dependent (00:15:49) um networks of problems. These this you (00:15:51) know big goop of a mesh of problems it's (00:15:55) always good to to go back and find a (00:15:57) framework for what it is that we're (00:15:59) talking about. In the case of AI um what (00:16:04) is AI? (00:16:06) Well, of course, the technology of AI (00:16:09) and the capability, the capabilities of (00:16:11) AI is about automation. It's about (00:16:14) automation of intelligence for the very (00:16:16) first time. And you could combine it (00:16:19) with megatronics technology to embody (00:16:23) that megatronics and and make it perform (00:16:25) tasks. (00:16:27) >> So that's what's AI automation. But what (00:16:31) what is the stack that makes AI (00:16:33) possible? What's the technology stack? (00:16:34) functional stack. And of course the e (00:16:37) the easiest way to think about that is (00:16:39) it's kind of like a fivey year five year (00:16:41) five layer cake which is at the lowest (00:16:43) level is energy. (00:16:45) >> Um it transforms energy to the output (00:16:47) that I just described. The next layer is (00:16:49) chips. The next layer is infrastructure (00:16:51) and that infrastructure is both hardware (00:16:53) software right this is where land power (00:16:56) and shell this is where construction is (00:16:58) data centers are the software stack (00:17:01) >> you know for orchestrating the so it's (00:17:03) software and hardware the layer above (00:17:05) that is where everybody thinks about (00:17:07) which is AI which is the models (00:17:10) >> we know this but it's really helpful to (00:17:12) understand that AI is a system of models (00:17:15) >> and AI is a um a techn technology that (00:17:20) understands information and there's (00:17:22) human information and so we often times (00:17:26) think about AI as a chatbot (00:17:28) >> but remember there's biological (00:17:30) information there's chemical information (00:17:32) there's physical information of all (00:17:34) kinds there's financial information (00:17:37) there's healthcare information there's f (00:17:38) there's information of all modalities (00:17:41) all kinds AI is really really broad and (00:17:44) of course human language is at the (00:17:46) foundation of of many things but it's (00:17:48) not the essence of everything because as (00:17:51) you know you know biology molecules (00:17:53) don't understand English (00:17:55) >> they understand something else right (00:17:57) proteins don't understand English they (00:17:58) understand something else I think the (00:18:00) next layer the important thing is is uh (00:18:02) that's where the AI models are but (00:18:04) there's a whole the AI is very very (00:18:06) diverse and then the the layer above (00:18:08) that is is applications and it depends (00:18:10) on the industry and you already (00:18:11) mentioned open evidence there you (00:18:13) mentioned Harvey there's cursor there's (00:18:15) all kinds of right there's all kinds of (00:18:16) applications full self-driving is really (00:18:18) an application, an AI application that (00:18:21) is embodied into a mechanical car (00:18:23) >> and figure is a AI application that has (00:18:26) been embodied into a mechanical human. (00:18:28) And so, so you got all these different (00:18:31) applications. Well, this five layer (00:18:33) stack is one way of thinking about it. (00:18:36) And then the next way of thinking about (00:18:37) I just mentioned is AI is really (00:18:39) diverse. When you now have this (00:18:40) framework of what the the technology (00:18:43) capabilities are, how to how to build (00:18:46) the technology and how diverse it is, (00:18:49) then you can come back and think about (00:18:51) okay, let's ask the question, how (00:18:53) important is open source? Well, without (00:18:56) open source, you know, today, of course, (00:19:00) the frontier models, the the the leading (00:19:02) labs have chosen to to use a closed (00:19:05) source um application approach, which is (00:19:08) just fine. you know what people decide (00:19:10) to do with their business models is is (00:19:11) really in the final analysis. It's their (00:19:13) business and they have to they have to (00:19:15) calculate what is the best way for them (00:19:17) to get the return on investment so that (00:19:18) they could scale up and and make better (00:19:20) advances. Um however they they made that (00:19:23) calculus is fantastic. On the other (00:19:26) hand, uh without open source, as you (00:19:28) know, startups would be challenged, uh (00:19:31) companies that are in in uh uh different (00:19:33) industries, whether it's manufacturing (00:19:36) or transportation or um it could be in (00:19:39) healthcare. Without open source today, (00:19:42) all of that AI work would be suffocated. (00:19:46) >> And so, they just need to have something (00:19:47) that's pre-trained. They need to have (00:19:50) some fundamental technology about (00:19:51) reasoning. from that they could all (00:19:54) adapt, fine-tune, you know, train their (00:19:57) AI models into exactly the domain and (00:20:00) application they want. And so what (00:20:02) people really really miss is just the (00:20:05) incredible pervasiveness and the (00:20:08) importance of open source to all of (00:20:10) these industries. large companies uh (00:20:13) without without open source some of some (00:20:15) of 100-year-old companies that I work (00:20:17) with (00:20:17) >> in in industrial spaces in healthcare (00:20:20) spaces they would be suffocated they (00:20:22) wouldn't be able to do that (00:20:23) >> open source at this point is driving all (00:20:24) of our data centers is driving a big (00:20:27) chunk of telefan in the world in terms (00:20:28) of Android or other devices it's driving (00:20:30) exactly (00:20:31) >> you know to your point a lot of the (00:20:32) industrial applications so it's already (00:20:33) pervasive and I think the big question (00:20:34) is (00:20:35) >> open source without open source higher (00:20:37) ed (00:20:37) >> higher ed wouldn't happen (00:20:38) >> education research (00:20:40) >> startups I mean the list goes on, you (00:20:42) know, and so so (00:20:44) >> we talk we talk all day long about the (00:20:46) tip but the most visible part of that (00:20:49) the most the part that's most newsworthy (00:20:51) maybe but underneath that is such an (00:20:56) important space of open source AI and (00:20:58) whatever we decide to do with policies (00:21:01) do not damage that innovation flywheel. (00:21:05) So I spent a lot of time uh educating (00:21:08) educating uh uh policy makers to help (00:21:11) them understand whatever you decide (00:21:12) whatever you do don't forget open (00:21:15) source. Whatever you decide whatever you (00:21:17) do don't forget biology. (00:21:20) I think the counternarrative here that (00:21:22) is worth addressing is that essentially (00:21:25) like you know there should be a (00:21:28) monolithic vertical player and (00:21:30) monolithic asset in the like one model (00:21:32) that does it all and that we can't give (00:21:34) away that crown jewel to other countries (00:21:37) or non-American companies and and your (00:21:39) your argument is like we actually need (00:21:41) this huge diversity of AI applications (00:21:44) and and the American advantage is (00:21:46) actually or any any sovereign advantage (00:21:48) is in the whole stack right? The (00:21:50) capability to deliver any piece of it. (00:21:52) >> I guess someday we will have God AI. (00:21:56) >> But when is that day? (00:21:57) >> But but that someday that someday is (00:21:59) probably on biblical scales, you know, I (00:22:02) think galactic scales. Um I I think it's (00:22:05) it's not helpful to go from where we are (00:22:08) today to God AI. (00:22:11) >> And um I don't think any company (00:22:14) practically believes they're anywhere (00:22:16) near God AI. And nor nor do I do I see (00:22:20) any researchers having any reasonable (00:22:22) ability to create god AI. The ability to (00:22:25) h understand human language and genome (00:22:29) language and molecular language and (00:22:31) protein language and amino acid language (00:22:33) and physics language all supremely well. (00:22:36) That god AI just doesn't exist. (00:22:39) >> And and yet we have a lot of industries (00:22:41) that need AI. Mhm. (00:22:43) >> AI is if if you will at the simplistic (00:22:47) level, it's just the next computer (00:22:50) industry. (00:22:51) >> And give me an example of a company, an (00:22:54) industry, a nation who doesn't need (00:22:56) computers. (00:22:57) >> Mhm. (00:22:58) >> And we all don't have to wait around for (00:23:00) God AI for us to advance, right? So God (00:23:03) AI is not showing up next week. I'm (00:23:04) fairly certain of that. Okay. And God (00:23:06) [clears throat] AI god AI is not not (00:23:08) going to show up next year, but the (00:23:10) whole world needs to move forward next (00:23:11) week, next year, next decade. I think (00:23:13) that that the idea of a monolithic (00:23:16) gigantic (00:23:18) company, (00:23:20) >> country, nation, state that has got AI (00:23:24) is just (00:23:25) >> it's unhelpful. (00:23:26) >> It's unhelpful. It's too extreme. (00:23:28) >> Then in fact, if you want to take it to (00:23:31) that level, then we ought to just all (00:23:32) stop everything. (00:23:34) What's the point of having even (00:23:36) governments? I mean, why why why are (00:23:38) they doing policies? God AI is going to (00:23:39) be smart enough to avert, you know, work (00:23:41) around any policy. And so, what's the (00:23:44) point? And so, I I think that that we (00:23:46) ought to bring things back to the ground (00:23:48) ground level and start thinking about (00:23:50) things practically and and use common (00:23:53) sense. (00:23:55) >> This seems to be like a big theme in (00:23:56) general in terms of this conversation (00:23:58) where there's been a lot that's been (00:23:59) kind of put out there that seems very (00:24:02) extreme if you actually think about it. (00:24:03) It's the jobs and employment. Nobody's (00:24:04) going to be able to work again. It's God (00:24:06) AI is going to solve every problem. It's (00:24:08) we shouldn't have open source for XYZ (00:24:10) reason despite open source powering much (00:24:11) of our industries already. (00:24:12) >> That's right. (00:24:13) >> And so it seems like in general maybe (00:24:14) one of the themes of 2025 was there's a (00:24:16) lot of extremes that were sort of (00:24:18) painted in the public with AI that if (00:24:21) you look at them very closely don't (00:24:22) really follow a logical change in terms (00:24:23) of happening anytime soon. (00:24:25) >> Yeah. And so it's it's it sounds like (00:24:26) it's really important to have this (00:24:27) conversation. (00:24:28) >> Extremely hurtful frankly. And I I think (00:24:30) we've done a lot of damage uh with very (00:24:33) wellrespected people um who have who (00:24:36) have painted a doom doomer narrative um (00:24:40) end of the world narrative science (00:24:42) fiction narrative and um you know and I (00:24:45) and I appreciate that that many of us (00:24:47) grew up in and enjoyed science fiction. (00:24:50) >> Um but I but it's not helpful. It's not (00:24:52) helpful to people. It's not helpful to (00:24:54) the industry. It's not helpful to (00:24:55) society. It's not helpful to the (00:24:57) governments. Mhm. (00:24:58) >> There are a lot of many people in the (00:25:00) government who obviously aren't as (00:25:03) familiar with as as comfortable with the (00:25:06) technology (00:25:07) >> and when PhDs of this and CEOs of that (00:25:12) >> goes to governments and explain and (00:25:14) describe these end of the world (00:25:16) scenarios and extremely extremely (00:25:19) dystopian future the future. Um, you (00:25:22) have to ask yourself, you know, what is (00:25:24) the purpose of that narrative and what (00:25:26) is their what are their intentions and (00:25:28) what do they hope? Why are they why are (00:25:31) they talking to governments about these (00:25:32) things to create regulations to (00:25:35) suffocate startups? [clears throat] (00:25:36) >> For what reason would they be doing (00:25:38) that, you know, and so (00:25:39) >> and do you think that's just regulatory (00:25:41) capture where they're trying to prevent (00:25:43) uh new startups from showing up and (00:25:44) being able to compete effectively or (00:25:46) what do you think is the goal of some of (00:25:47) these conversations? you know, I I can't (00:25:50) I can't um uh guess what they what they (00:25:53) have in mind. I know that the concern is (00:25:55) regulatory capture. As a policy, as a (00:25:58) practice, I don't think companies had to (00:26:02) go to (00:26:04) um governments to advocate (00:26:08) for the regulation on other companies (00:26:12) and other industries. just in practice (00:26:15) their their intentions are clearly (00:26:18) deeply conflicted and and uh their (00:26:22) intentions are clearly you know not (00:26:24) completely in the best interest of (00:26:26) society. I mean they're obviously CEOs (00:26:28) are obviously companies and obviously (00:26:30) they're advocating for themselves (00:26:32) >> and so so I think if we can all (00:26:36) >> come back to where are we today (00:26:39) >> and think about where the technology is (00:26:41) going to be. I mean look lit literally (00:26:43) in one year's time as we were talking (00:26:45) about in the beginning uh some of the (00:26:48) most proud moments is when the industry (00:26:52) was able to invest very aggressively in (00:26:55) advancing AI technology instead of being (00:26:57) slowed down. (00:26:58) >> Remember just two years ago people were (00:27:00) talking about slowing the industry down (00:27:02) >> but as we advanced quickly what did we (00:27:05) solve? We solved grounding, we solved (00:27:08) reasoning. We solved research. All of (00:27:11) that technology was applied for good (00:27:14) improving the functionality of the AI (00:27:17) not you know (00:27:19) >> yet the end has not come (00:27:20) >> yet the end has not come it's become (00:27:22) more useful it's become more functional (00:27:25) it's become able to do what we ask it to (00:27:27) do you know and so the first the first (00:27:30) part of the safety of a product is that (00:27:32) it perform as advertised (00:27:36) >> the first part of safety is performance (00:27:40) that it's is supposed like the first (00:27:42) part of safety of a car isn't that some (00:27:44) person is going to jump into the car and (00:27:47) use it as a missile. The first part of (00:27:49) the car is it works as advertised. (00:27:52) >> Mhm. (00:27:53) >> 99.999% (00:27:55) of the time working as advertised. And (00:27:58) so it takes a lot of technology to make (00:28:00) that car or make that AI work as (00:28:03) advertised. And I'm really glad that in (00:28:05) the last couple two three years the (00:28:07) industry has invested so much in (00:28:09) enhancing the functionality of the AI as (00:28:12) advertised. And I think if if we were to (00:28:14) to look at the next 10 years, we have so (00:28:17) much work to do to make it work as (00:28:21) advertised. Meanwhile, as as you know, (00:28:23) you both of you invest so much in in the (00:28:25) in the ecosystem, you see so many (00:28:28) companies being built for um synthetic (00:28:31) data generation so that the AIs could be (00:28:34) more grounded uh more diverse uh less (00:28:38) biased more safe uh you're investing in (00:28:40) a whole bunch of companies in cyber (00:28:42) security using AI for cyber security you (00:28:44) right people think that there's this AI (00:28:47) um the marginal cost of AI is going to (00:28:49) go go down significantly and it is (00:28:52) >> and therefore the AI is going to be (00:28:54) dangerous. It's exactly the opposite. If (00:28:57) the marginal cost of AI is going to go (00:28:59) down significantly, that one AI is going (00:29:01) to be monitored by millions of AIS. (00:29:04) >> Mhm. (00:29:04) >> And more and more AI is going to be (00:29:06) monitoring monitoring each other. People (00:29:08) don't can't forget that an AI is not (00:29:10) going to be an agent by itself. It's (00:29:12) likely the AI is going to be surrounded (00:29:14) by agents monitoring it. And so it's no (00:29:17) different than if the if the marginal (00:29:19) cost of of keeping society safe was (00:29:22) lower. We have police in every corner. (00:29:25) >> So one thing that that we were talking (00:29:27) about a little bit earlier was just the (00:29:28) cost of AI and how it's been coming (00:29:30) down. And so (00:29:31) >> I I think um in 2024 the the cost of (00:29:35) GPT4 equivalent models if you look at a (00:29:37) million tokens it came down over 100x. (00:29:40) Um you know somebody in my team did this (00:29:42) analysis to show that. Uh so the costs (00:29:44) are dropping pretty dramatically and (00:29:45) very rapidly and part of it is all the (00:29:47) advancements you all have been driving (00:29:48) on and the Nvidia level but also just (00:29:49) across the stack getting big efficiency (00:29:52) gains. (00:29:52) >> Yeah. (00:29:53) >> Um at the same time model companies are (00:29:54) talking about how the costs are rising (00:29:56) how there's enormous sort of capital (00:29:58) modes to building these things out. How (00:30:00) do you think about cost of training and (00:30:01) cost of inference over time and what (00:30:03) that means for the average end user or (00:30:05) the average startup company trying to (00:30:07) compete or people trying to do more in (00:30:08) this industry? I forget the statistic (00:30:10) that but but you know Andre Andre (00:30:12) Cararpathy um estimated the cost of (00:30:15) building the first chatbt I think (00:30:18) >> versus now I think you could do that on (00:30:20) the PC now. (00:30:20) >> Yeah. Yeah. It's probably tens of (00:30:22) thousands of dollars at this point or (00:30:23) maybe even less. (00:30:24) >> Right. And so it costs nothing. (00:30:26) >> Mhm. (00:30:26) >> And and (00:30:27) >> he has an open source project that you (00:30:28) can do in a weekend. (00:30:29) >> Oh, is that right? Okay. That's (00:30:31) incredible. Right. We're talking about (00:30:32) three years. Mhm. (00:30:33) >> Mhm. (00:30:34) >> What people people said cost billions of (00:30:37) dollars (00:30:39) um supercomputers built raising billions (00:30:42) of dollars in order to do all that now (00:30:45) >> cost you know something that you can do (00:30:47) on a weekend on a PC. And so that tells (00:30:50) you something about how quickly we're (00:30:52) making making AI more cost effective (00:30:54) >> or Spark sorry probably not quite a PC. (00:30:56) >> Okay. Not quite a PC. Yeah. We're (00:30:59) improving our architecture and (00:31:03) performance (00:31:04) um every single year. The first GBTU I (00:31:07) think was trained on Voltus. (00:31:08) >> Mhm. (00:31:09) >> And then uh Ampear (00:31:11) um you know and and it wasn't I think (00:31:14) the first breakthroughs none of it (00:31:16) included Hopper. (00:31:17) >> Mhm. (00:31:18) >> And um of course Hopper the last couple (00:31:20) two three years and um uh we're off in (00:31:22) Blackwell for the last year and a half (00:31:24) or so. And um every single one of these (00:31:27) generations the architecture improves (00:31:30) and of course the number of transistors (00:31:32) go up and uh the capacity goes up every (00:31:35) single generation very easily every (00:31:37) every single year from a computing (00:31:39) perspective. The combination of all that (00:31:41) getting 5 to 10x every single year (00:31:44) >> is not unusual. And here comes Reuben (00:31:46) just around the corner. And so we're (00:31:48) seeing 5 to 10x every single year. Well (00:31:51) compounded it's incredible. Moore's law (00:31:54) was two times every year and a half (00:31:56) >> and over the course of five years is 10x (00:31:58) over the course of 10 years is 100x (00:32:01) >> in the in the in the case of AI over the (00:32:03) course of 10 years is probably 100,000 (00:32:05) to a millionx okay and that's just the (00:32:08) hardware (00:32:09) >> then the next layer is the algorithm (00:32:11) layer and the model layer the (00:32:13) combination of all that the fact that if (00:32:16) you were to tell me that in the cost in (00:32:17) the in the in in the span of you know 10 (00:32:20) years we're going to reduce the cost of (00:32:21) token generation about a billion times. (00:32:23) I would not be surprised. (00:32:24) >> Mhm. (00:32:25) >> Okay. And so that's the tokconomics (00:32:28) of of of AI. On the training side, it's (00:32:31) not quite as aggressive in in cost (00:32:33) reduction, but it's close. If you were (00:32:35) to say that that every single year we're (00:32:38) increasing by two or 3x over the course (00:32:40) of 10 years, incredible. But the (00:32:43) important idea is when somebody says it (00:32:45) cost $und00 million to train something (00:32:48) or half a billion dollars to train (00:32:50) something. (00:32:51) Well, next year it's 10 times less. Next (00:32:54) year it's 10 times. (00:32:54) >> For people to scale these things up, (00:32:56) though, right? So the counter argument (00:32:57) is, well, we'll just get bigger every (00:32:58) year by 10x or 100x or, you know, we'll (00:33:01) try to offset that decrease in cost by (00:33:03) scale (00:33:05) >> and others can't keep up. (00:33:06) >> Yeah. But really what's happening is is (00:33:09) you're and and this is where come in as (00:33:12) you know the scale went up by a factor (00:33:14) of 10 but the computational burden did (00:33:16) not go up by a factor of 10 because (00:33:18) you're getting the compounded benefits (00:33:20) of all three things. The hardware is (00:33:22) going up the the algorithms of the (00:33:25) training models are going up and of (00:33:26) course the model architecture is going (00:33:28) up and we're getting the benefit of (00:33:29) learning from each other. This is, you (00:33:31) know, let's face it, Deep Seek was (00:33:33) probably the single most important paper (00:33:37) that most Silicon Valley researchers (00:33:39) read from in the last couple years. (00:33:41) >> It was the only thing that felt frontier (00:33:43) that was open. (00:33:44) >> That's right. (00:33:45) >> In years, (00:33:47) the value of open source again putting (00:33:49) out these papers. (00:33:49) >> Literally, Deep Seek (00:33:51) >> benefited American startups and American (00:33:54) AI labs all over (00:33:55) >> and infrastructure companies (00:33:56) >> and infrastructure company all over. (00:33:59) probably the single greatest (00:34:00) contribution to American AI last year. (00:34:03) >> And so if you said this out loud, of (00:34:05) course, you know, people (00:34:08) >> kind of shudder um that we're uh (00:34:11) American AI is actually getting learning (00:34:13) from and benefiting from uh AI from (00:34:16) other nation. But why would that be (00:34:18) surprising? You know, AI researchers in (00:34:20) all over America, all over America are (00:34:22) uh Chinese natives and come from (00:34:24) different countries. We benefit from (00:34:26) every country. become benefit from every (00:34:28) researcher and no all of the world's (00:34:31) ideas don't have to come from the United (00:34:33) States and so I I think um back to your (00:34:36) your original question it is the case (00:34:39) that (00:34:40) you know some of the narratives around (00:34:42) around the cost of AI is about scaring (00:34:45) everybody out of the market you know (00:34:47) nobody ought to do pre-training but us (00:34:49) nobody should do you know training these (00:34:51) frontier models but us because the (00:34:53) because of innovation of models (00:34:56) algorithms (00:34:57) and the computing stack, the cost of AI (00:35:00) is actually decreasing well more than (00:35:03) 10x every single year. And so if you're (00:35:04) just one year behind or even six months (00:35:06) behind, you could you could really stay (00:35:08) close. (00:35:09) >> And I think one thing that felt very (00:35:11) different to me about 2025 is um Ilia uh (00:35:15) said recently that uh you know we're in (00:35:17) the age of research again versus an age (00:35:20) of scaling. I think both things are (00:35:21) happening by the way. Everybody is also (00:35:23) trying to scale on multiple dimensions. (00:35:24) >> Yeah, exactly. both are happening. (00:35:26) >> You know, being 6 months behind or being (00:35:28) at 100 versus a 200k cluster, I think (00:35:31) matters if you are competing (00:35:33) symmetrically, but now you have people (00:35:35) from Frontier Labs or um at the very top (00:35:38) of the game who have very different (00:35:40) ideas about how to progress from here or (00:35:42) who are working on diversity of (00:35:43) problems, right? Uh and and I I think (00:35:45) that felt different from 24 maybe where (00:35:48) there was a lot of energy focused on (00:35:50) just pre-training scale and LLM. (00:35:52) >> Yeah. And several several other (00:35:54) dynamics. Um, as the market grows, each (00:35:57) one of these models could choose to have (00:36:00) verticals (00:36:02) >> or segments where they want to (00:36:03) differentiate. (00:36:04) >> Somebody could decide to be a better (00:36:06) coder. Somebody could decide to be just (00:36:09) better at being easier to be accessible (00:36:11) so that it could be a greater consumer (00:36:13) product. (00:36:14) >> You know, the diversity of these models. (00:36:15) As a result, you could you could (00:36:18) probably make a niche leap without (00:36:22) having to be great at everything else (00:36:23) and still be super valuable to the (00:36:25) market. (00:36:26) >> It's no longer necessary to boil the (00:36:28) entire ocean. The f two years ago, (00:36:31) because it was called pre-training pre, (00:36:34) you know, people people said, well, you (00:36:36) know, pre-training is over. First of (00:36:38) all, pre-training is not over. But the (00:36:41) point of pre-training is to train (00:36:43) yourself for training. That's why it's (00:36:45) called pre-training to prepare yourself (00:36:47) to do the real training. And now we call (00:36:49) it post-training. It's kind of weird. I (00:36:52) I think it's just training, but (00:36:54) pre-training is pre-training and (00:36:55) therefore it's training. Training as you (00:36:58) as as we all know is is where uh compute (00:37:02) scaling directly translates to (00:37:04) intelligence. You you've you've largely (00:37:07) now now this the the data the the data (00:37:10) necessary to train a model is actually (00:37:12) pretty small. Maybe it's just the (00:37:14) verifiable results. Now it's really (00:37:16) algorithmic, very compute intensive and (00:37:19) so and you don't have to be good at (00:37:20) everything in life as you know just like (00:37:23) all of us we don't we could decide (00:37:24) because we don't have time to learn (00:37:26) everything equally well. We decide to (00:37:28) choose a specialty and focus all of our (00:37:31) energy on it and we become superhuman or (00:37:33) incredibly good at something that other (00:37:35) people are not. And so I think AI labs (00:37:37) are going to start doing the same. (00:37:39) They're going to start bifurcating into (00:37:41) various segments and over time you're (00:37:44) gonna and startups will do the same. (00:37:46) >> They'll find a micro niche and they'll (00:37:47) take something open and then be (00:37:49) incredibly good at it. (00:37:50) >> Well, I think one of the most optimistic (00:37:51) views here is uh actually that these (00:37:53) microniches are quite valuable, right? I (00:37:56) was talking to Andre um because I've (00:37:57) been talking to a lot of people about (00:37:58) their predictions for next year. We'll (00:38:00) ask you yours as well of course. Um um (00:38:03) but he asked you know what is a what's (00:38:05) an example of a prediction that would (00:38:06) have been preient last year uh and my (00:38:10) answer everything's easy in retrospect (00:38:12) is that coding would be the first (00:38:14) application level business that gets to (00:38:16) a billion of AR as an AI native app (00:38:18) right and I I think if you taken an old (00:38:21) world view of this (00:38:23) >> um you would have believed like one of (00:38:24) two narratives right one is uh single (00:38:28) model does everything and it'll all just (00:38:29) be subsumed into something monolithic (00:38:31) Mhm. (00:38:31) >> And two is that developer tools never (00:38:33) get very big, right? Well, kind of (00:38:35) depends on how valuable the developer (00:38:36) tool is. Now, I think many more people (00:38:38) understand software engineering is in a (00:38:40) niche and there's more demand than ever (00:38:41) for it, (00:38:42) >> but I think we'll see more like that. (00:38:44) >> Also interesting, uh we are using we we (00:38:47) use cursor here and we use cursor (00:38:50) pervasively here. Every engineer uses it (00:38:52) and the number of engineers, you just (00:38:53) mentioned it, the number of people we're (00:38:55) hiring today is just incredible. (00:38:56) >> Yep. (00:38:57) >> Right. Monday is come to work at Nvidia (00:38:59) day and and um uh why is that? Uh this (00:39:03) is now the purpose and the task. (00:39:05) >> The purpose of a software engineer is to (00:39:08) solve known problems (00:39:10) and to find new problems to solve. (00:39:15) Coding is one of the tasks. (00:39:18) >> And so if the purpose is not coding, if (00:39:21) your purpose literally is coding, (00:39:23) somebody tells you what to do, you code (00:39:24) it. All right? Maybe you're going to get (00:39:26) replaced by the AI. But most of our (00:39:27) software engineers, all of our software, (00:39:29) their goal is to solve problems. And it (00:39:32) turns out we have so many problems in (00:39:33) the company and we have so many (00:39:34) undiscovered problems. And so the more (00:39:37) time they have to go explore (00:39:38) undiscovered problems, the better off we (00:39:40) are as a company. Nothing would give me (00:39:42) more joy than if none of them are coding (00:39:44) at all. They're just solving problems. (00:39:46) >> You see what I'm saying? And so I I (00:39:48) think that this framework of purpose (00:39:50) versus task is really good for everybody (00:39:52) to apply. For example, somebody who's a (00:39:54) waiter, their job is to not to take the (00:39:59) order. That's not their job. As it turns (00:40:00) out, their job is so that we have a (00:40:02) great experience. And if somebody if (00:40:04) some AI is taking the order, their job (00:40:07) or even delivering the food, their job (00:40:09) is still helping us have a great (00:40:11) experience. They they would reshape (00:40:13) their jobs accordingly. And so so I (00:40:16) think the um the question about about (00:40:19) cost of compute um uh is really (00:40:23) important. Let's let let me come back to (00:40:25) one the the reason why we are so (00:40:28) dedicated to a programmable architecture (00:40:32) versus a fixed architect. Remember a (00:40:34) long time ago (00:40:35) >> uh a CNN chip came along and they said (00:40:37) Nvidia is done. (00:40:39) >> And then and then a transformer chip (00:40:41) came and Nvidia was done. (00:40:42) >> People are still trying that. Yes. (00:40:43) >> Yeah. NP and and the benefit of these (00:40:46) dedicated AS6 of course it could perform (00:40:49) a job really really well and (00:40:50) transformers is a much more universal AI (00:40:55) network but the transformer as you know (00:40:58) the species of it is growing incredibly (00:41:00) >> the attention mechanism (00:41:01) >> the attention mechanism how it thinks (00:41:03) about context (00:41:05) diffusion versus auto reggressive (00:41:08) >> these hybrid SSM transformation (00:41:10) >> hybrid SSM for example Neotron we just (00:41:12) announced a new hybrid SM SM and and so (00:41:15) the architecture of transformer is in (00:41:18) fact changing very rapidly and over the (00:41:20) next several years it's likely to change (00:41:21) tremendously and so we we dedicate (00:41:24) ourselves to an architecture that's (00:41:25) flexible for this reason so that we can (00:41:27) on the one hand adapt with remember (00:41:30) because MOS law is largely over (00:41:32) transistor benefit is only tens 10% (00:41:37) maybe a couple of years (00:41:39) >> and yet we would like to have hundreds (00:41:40) of X every year and so the benefit is (00:41:44) actually all in algorithms and an (00:41:46) architecture that enables any algorithm (00:41:48) is likely going to be the best one right (00:41:50) because the transistor didn't it didn't (00:41:52) advance that much and so I I think the (00:41:54) the our dedication to programmability is (00:41:57) number one for that reason we have so (00:41:58) much optimism for innovation and (00:42:01) algorithms and iteration software that (00:42:03) we protect our programmability for that (00:42:06) reason the second thing is is by (00:42:09) protecting this architecture (00:42:11) our installed base is really large. When (00:42:13) a software engineer wants to optimize (00:42:16) their algorithm, they want to make sure (00:42:18) that it doesn't run on just one this one (00:42:20) little cloud or this one little stack. (00:42:21) They want it to run on as many mo on as (00:42:24) many computers as possible. So the the (00:42:25) fact that we protect our architecture (00:42:27) compatibility then flash attention runs (00:42:30) everywhere. So SSM run everywhere, (00:42:32) diffusion runs everywhere, auto (00:42:34) reggression runs everywhere. Just (00:42:36) depending it doesn't matter what you (00:42:37) want to do. CNN still run everywhere. (00:42:39) LSTM still runs everywhere. And so that (00:42:41) this this architecture that is (00:42:43) architecturally compatible so that we (00:42:45) have a large installed base programmable (00:42:47) for the future is really important in (00:42:50) the way that we help to advance and as a (00:42:52) result all of this drives the cost down (00:42:55) [clears throat] and and I'm super proud (00:42:56) that that um uh our latest innovation (00:42:59) MVLink72 (00:43:01) we're the lowest cost token generation (00:43:04) machine in the world by enormous amounts (00:43:07) and the reason for that is because are (00:43:10) really really hard (00:43:11) >> and so you know people didn't expect (00:43:13) that um that forees it's probably easier (00:43:17) to train but for inference is incredibly (00:43:19) hard to generate tokens on but as as (00:43:22) cost drop usually you open up new (00:43:23) applications or new verticals that (00:43:25) become more and more accessible (00:43:27) >> and we talked a little bit about coding (00:43:28) like cursor and cognition and other (00:43:30) companies that are really benefiting (00:43:31) from that in this last year do you have (00:43:32) any thoughts or predictions in terms of (00:43:34) what the next breakthrough industries (00:43:35) will be or new applications or areas (00:43:37) that you're most excited about coming in (00:43:38) 26 in particular like Are there one or (00:43:40) two things that you think will (00:43:41) >> because of three things I because of (00:43:44) because of a couple two three things I I (00:43:47) think I think several industries are (00:43:49) going to are going to experience their (00:43:50) chat moment. Um I believe that (00:43:54) multi-modality (00:43:57) and um very long context is going to (00:44:01) enable of course really really cool chat (00:44:04) bots. Um but the basic architecture that (00:44:08) in combination with breakthroughs in (00:44:10) synthetic data generation is going to (00:44:12) help create the chat GPT moment for (00:44:16) digital biology. (00:44:18) >> That moment is coming. (00:44:19) >> And by digital biology, do you (00:44:20) specifically mean other aspects of like (00:44:23) protein folding or protein binding or (00:44:24) protein diagnosis? I see proteins. (00:44:26) >> I think we're good at protein (00:44:28) understanding. Mhm. Now multi-proin (00:44:30) understanding is coming online and we (00:44:33) recently created a model called LA (00:44:35) prina. It's open. Um it's for (00:44:37) multi-proin (00:44:39) >> um understanding and and represent (00:44:41) representation learning and generation. (00:44:43) Uh so so I think that the protein (00:44:46) understanding is is advancing very (00:44:47) quickly. Now protein generation is going (00:44:49) to advance very quickly. Chat GPD moment (00:44:52) proteins. (00:44:52) >> Yeah. There are a lot of interesting (00:44:53) companies working on molecule design in (00:44:55) endtoend way like chai. (00:44:57) >> Exactly. And then and then of course (00:45:00) chemical understanding and chemical (00:45:01) generation and then protein chemical (00:45:05) >> confirmation understanding and (00:45:07) generation. Is that right? And so that (00:45:09) combination the chat GBT moment the (00:45:11) generative AI moment all of that stuff (00:45:13) is coming together for for um digital (00:45:14) biology (00:45:15) >> and to your to your point about like new (00:45:17) industries or you know the way I think (00:45:19) about it is like investing in the inputs (00:45:21) for this AI as well. All of these things (00:45:23) around biology and chemistry and (00:45:25) material science, they require real (00:45:27) world data generation and (00:45:28) experimentation, right? And that's new (00:45:30) infrastructure too. (00:45:30) >> New infrastructure, uh, synthetic data (00:45:33) is going to be really important because (00:45:34) they just have such sparse, right? Spar (00:45:36) sparity of data and they just don't have (00:45:38) as much as human language. And there the (00:45:40) the real breakthrough is going to be (00:45:42) when we can train a a world foundation (00:45:45) model, a foundation model for proteins, (00:45:48) a foundation model for cells. I'm I'm (00:45:50) very excited about both of those things. (00:45:52) Once we have a a foundation model, our (00:45:55) understanding capability, our generative (00:45:57) capability, that data flywheel is really (00:45:59) going to take off. (00:46:00) >> The this this the second area that I'm (00:46:02) excited about, um, of course, reasoning (00:46:04) made huge breakthroughs in language, but (00:46:06) because of reasoning, cars are going to (00:46:09) be able to perform better. So, instead (00:46:10) of just perception cars and planning (00:46:13) cars, they're going to be reasoning (00:46:14) cars. So, these cars are going to be (00:46:16) thinking all the time. And when they (00:46:17) come up they come up to a circumstance (00:46:19) they they've never en encountered before (00:46:21) they can break it down into (00:46:22) circumstances they have encountered it (00:46:24) before and construct a reason reasoning (00:46:28) system for how to navigate through it. (00:46:30) And so the out of domain out of you know (00:46:33) out of distribution (00:46:35) >> part of AI is going to very much be be (00:46:38) addressed by reasoning systems or and as (00:46:41) a result we could do more things than we (00:46:43) were taught to do between uh generative (00:46:44) AI uh and um multimodal uh you know (00:46:49) vision language action models and (00:46:51) reasoning systems. I think we're going (00:46:53) to see big breakthroughs in human robots (00:46:55) or multi-mbodiment robots. you know does (00:46:58) >> what do you think what do you think is a (00:46:59) time frame for that because if you look (00:47:00) at the self-driving analog and obviously (00:47:02) self-driving technologies were based on (00:47:05) very different types of neural networks (00:47:06) than what we're using today in terms of (00:47:08) you know there's been a big swap over (00:47:09) the last two three years (00:47:11) >> in terms of how we do a lot there (00:47:12) >> we started too soon (00:47:14) >> self-driving cars really had four eras (00:47:17) era was smart sensors (00:47:20) >> connected into a car (00:47:22) >> the mobile [clears throat] eye era (00:47:23) >> the mobile eye era and even even the (00:47:25) very earliest days of of (00:47:27) Yeah. (00:47:28) >> Yeah. Even the earliest days of Whimo, (00:47:30) >> the the um you're talk you're using (00:47:32) smart sensors um a lot of human (00:47:35) engineered algorithms (00:47:36) >> and education severe mapping as far (00:47:40) >> extreme mapping (00:47:41) >> mapping and then different systems for (00:47:42) planning and perception. (00:47:44) >> Exactly. And so so you're essentially (00:47:46) creating a car that is driving on (00:47:49) digital rails, right? It's no different (00:47:50) than than the rails at Disneyland. There (00:47:53) are digital rails. And so that's the (00:47:54) first generation. the second generation. (00:47:57) Um and during that generation you have (00:47:59) perception, world model and planning. (00:48:01) >> Mh. (00:48:02) >> And and the these modules um and each (00:48:05) one of these modules have the limits of (00:48:07) their technology and and perception was (00:48:09) first imple was was first affected by (00:48:11) deep learning uh first and then and then (00:48:14) uh and then it propagated through the (00:48:15) pipeline. (00:48:16) >> And so that but that system was too (00:48:18) brittle (00:48:19) >> and it only knows how to perform what (00:48:20) you taught it. And now where we are are (00:48:23) endtoend models and then and then where (00:48:27) we're going to go next are end to end (00:48:28) models. There you go. So that those are (00:48:31) kind of the four eras in a lot of ways. (00:48:34) If we would have started self-driving (00:48:35) cars probably three years ago, (00:48:38) >> we would probably be exactly the same (00:48:40) place. (00:48:41) >> All our poor friends who were working in (00:48:42) self-driving. Yeah. (00:48:43) >> And and I don't I don't mind it. I've (00:48:44) been working on on it for 10 years. (00:48:46) Nvidia's self-driving car stack, by the (00:48:48) way, number one rated safety in the (00:48:51) world today. (00:48:53) >> Number one, we just got we just got that (00:48:55) rating today uh last week. And number (00:48:57) two is Tesla. So, I'm very proud that (00:48:59) two American companies are up on the (00:49:00) >> Are you um So, from a robotics (00:49:02) perspective, you think because we've (00:49:03) already built all these sorts of (00:49:04) technologies in the modern era, robotics (00:49:06) won't have the same 10, 15 years. That's (00:49:10) right. (00:49:11) >> I'm much more optimistic with robotics (00:49:13) because we we've kind of (00:49:15) >> advanced foundational technology. (00:49:18) >> Now, you know, people are thinking about (00:49:20) human robotics. Human robotics has a lot (00:49:22) of challenges. I mean, there's all the (00:49:24) megatronics challenges there. You know, (00:49:26) like for example, (00:49:27) >> it's not helpful if the robot weighs 300 (00:49:29) lb (00:49:30) >> and what happens if it falls over and (00:49:32) interacting with kids and so on so (00:49:33) forth. And so, so you got all kinds of (00:49:35) challenges to deal with. I'm certain (00:49:37) that we're going to we're going to solve (00:49:38) those. But remember the fundamental (00:49:40) technology that goes into a human robot (00:49:42) robot can go into a pick and place (00:49:44) robot. (00:49:45) >> Um it could be it could be um how do you (00:49:48) think about one thing I've been curious (00:49:49) about for robotics in particular is if I (00:49:51) look at who won or who who who's (00:49:53) perceived as winning in self-driving. (00:49:55) >> It's largely incumbents, right? It's (00:49:57) Whimo, it's Tesla. You mentioned uh the (00:49:59) safety rating Nvidia's gotten. And so (00:50:01) it's people who've been working on this (00:50:02) for a long time. It took a lot of (00:50:04) capital. It was really intensive to get (00:50:05) there. You have supply chain, you have (00:50:06) hardware, you have all this extra (00:50:07) complexity. Do you think the same thing (00:50:09) will be true in robotics? Are the (00:50:10) winners basically going to be Tesla with (00:50:12) Optimus and other people who have both (00:50:14) been in the industry for a while but (00:50:16) also have all those sort of incumbent (00:50:17) effects? Do you think there's room for (00:50:18) startups? (00:50:19) >> They will be one of the leader one of (00:50:20) the one of them and and and (00:50:24) surely a major one. Um but everything (00:50:28) that moves will be robotic. (00:50:31) >> Everything that moves will be robotic. (00:50:33) And everything that moves is a very (00:50:35) large space. It's not all human or (00:50:37) robot. And yet every AI will be (00:50:40) multi-mbodiment meaning you know just (00:50:43) like just like a human with our m our (00:50:46) multi-mbodiment (00:50:48) AI ourselves (00:50:49) >> we could sit in a car (00:50:51) >> and embody that (00:50:53) >> we could pick up a tennis racket embody (00:50:54) that we could pick up a chopstick embody (00:50:56) that (00:50:57) >> and so we could embody the (00:50:58) >> people are general purpose right they (00:51:00) can do all these things (00:51:00) >> exactly and so AIS are going to become (00:51:02) general purpose so you have one arm pick (00:51:04) and place maybe it's two arms pick and (00:51:06) place could be six arms pick and place, (00:51:08) you know. So, so I think you're going to (00:51:09) have all kinds of different sizes and (00:51:11) shapes. It could be a caterpillar. It (00:51:12) could be, you know, it could be an (00:51:13) excavator. It could be all kinds of (00:51:15) stuff. And so AI will embody those just (00:51:17) as a just as a a construction worker (00:51:20) embodies an excavator embodies a (00:51:22) tractor. You know, they you know, (00:51:25) >> could there be a small number of (00:51:26) companies then that do the embodiment (00:51:28) for everything or are you saying more (00:51:29) there's going to be niche applications? (00:51:30) You should definitely see a lot of (00:51:31) software companies and then those that (00:51:33) software (00:51:34) company could serve a lot of a lot of (00:51:36) different (00:51:37) >> verticals but each one of the verticals (00:51:39) will still have solution providers that (00:51:41) then grounds it all turns it into (00:51:43) something that works perfectly. Does it (00:51:45) make sense? Because in the case of AI (00:51:47) for consumers if it works 90% of the (00:51:49) time you're delighted you you're you (00:51:51) know you're mind blown. If it works 80% (00:51:53) of the time you're satisfied. In the (00:51:55) case of most industrial and physical (00:51:57) AIs, if it works 90% of the time, nobody (00:52:00) cares about that. They only care about (00:52:01) the 10% that it fails. Basically, you (00:52:03) know, 100% dissatisfaction. And so, you (00:52:06) got to take it to 99.99999. (00:52:09) So, the core technology might be able to (00:52:11) get get you to 99%. (00:52:13) >> And then a vertical solution provider (00:52:15) like a Caterpillar or somebody, they (00:52:17) could take that core technology and make (00:52:19) it 99.999% (00:52:21) great. Do you think that's what happens (00:52:23) like earliest on because in in markets (00:52:25) that are this immature it seems one of (00:52:27) the fastest paths to market could be (00:52:28) full verticalization right because you (00:52:30) just have control of iteration speed (00:52:33) >> the different the the difficulty (00:52:34) difficulty of of verticalization for (00:52:37) technology that that is general purpose (00:52:39) is that you don't have the R&D scale to (00:52:42) build a general purpose technology. Now, (00:52:43) of course, open source helps that (00:52:45) tremendously, (00:52:47) >> which is the reason why you're going to (00:52:48) see a, you know, a a big surge of (00:52:51) vertical (00:52:52) opportunities in AI in the next several (00:52:54) years. (00:52:55) >> My my prediction would be over the (00:52:57) course of the next five years, the (00:52:59) excitement is going to be (00:53:00) verticalization. (00:53:02) >> Notice we we're excited about Open (00:53:05) Evidence, we're excited about Harvey, (00:53:06) we're excited about Cursor. cursor is is (00:53:09) a horizontal but it's kind of a (00:53:11) horizontal vertical (00:53:12) >> you know and so um I'm I'm super excited (00:53:14) about all the verticals (00:53:16) >> you know a lot of people said yeah AI is (00:53:18) gonna get so god AI is going to get so (00:53:19) good that all these rapper companies are (00:53:22) going to be obsolete it's just it misses (00:53:24) the big point (00:53:25) >> you know the reason why you could talk (00:53:27) about the reason why somebody can talk (00:53:29) talk about somebody is creating (00:53:32) technology could talk about the life of (00:53:33) a surgeon is because they've never been (00:53:34) a surgeon the reason why somebody who (00:53:36) builds at AI and talk talks about the (00:53:38) life of a accountant and a tax, you (00:53:40) know, a tax expert because they've never (00:53:42) been a tax expert, you know, and so so I (00:53:45) I think they just the reason why (00:53:47) somebody could talk about being a bus (00:53:48) boy without being a bus boy is they (00:53:50) never been a bus boy. And so so I I (00:53:51) think you you you've got to be a little (00:53:54) bit more empathetic about the depth of (00:53:55) the complexity of the work (00:53:57) >> and and tr try to truly understand the (00:53:59) purpose of the work. Often times the the (00:54:01) technology addresses the task, it (00:54:04) doesn't address the purpose. So I guess (00:54:07) one of the other narratives from we're (00:54:09) looking at narratives that are true (00:54:10) versus not true, you know, for 25. One (00:54:13) other narrative that's come up has been (00:54:14) more about energy and energy utilization (00:54:16) and will we have enough energy to (00:54:18) support AI. How do how do you think (00:54:20) about that? On the first week of (00:54:22) President Trump's administration, he (00:54:23) said drill, baby drill. He got so much (00:54:25) flack for that. (00:54:27) If not for this entire change in in (00:54:31) sentiment about energy growth in our (00:54:33) country, (00:54:34) >> we can all concede now we would have (00:54:38) handed this industrial revolution to (00:54:40) somebody else. (00:54:42) >> And we're still power constrained. (00:54:43) >> We're still power constrained. Yeah. (00:54:45) >> Without energy, there can be no new (00:54:48) industry. (00:54:49) >> Mhm. And of course, we've been energy (00:54:52) starved now for what, a decade. If not (00:54:54) for the fact that President Trump (00:54:56) reversed that narrative, we would be (00:54:58) completely screwed. (00:55:00) >> Mhm. (00:55:00) >> Without energy, you can't have (00:55:02) industrial growth. Without industrial (00:55:04) growth, the the nation can't be more (00:55:06) prosperous. Without being more (00:55:08) prosperous, we can't take care of (00:55:09) domestic issues. We can't take care of (00:55:11) social issues. You know, on and on and (00:55:13) on. And so, the fact of the matter is, (00:55:15) we need energy to grow. We need every (00:55:17) form of energy. We need, you know, (00:55:19) natural gas. We need to be, of course, (00:55:21) we need more energy on the grid. We need (00:55:23) more energy behind the meter. Uh we're (00:55:25) going to need nuclear. Uh wind is not (00:55:28) going to be enough. Solar is not going (00:55:29) to be enough. Let's just all acknowledge (00:55:31) that we'll take it. We'll take (00:55:32) everything we can. Um but the fact that (00:55:34) matters, I think, for the for the next (00:55:36) decade, (00:55:37) >> natural gas, you know, is probably the (00:55:40) the only way to go forward. What's (00:55:42) really interesting is I I agree the (00:55:43) timeline is too far out to address (00:55:46) people's um you know power generation (00:55:48) issues in 27 and 28 where uh you know (00:55:51) large players building clusters are very (00:55:53) concerned but the the biggest drivers of (00:55:57) like climate innovation in the US have (00:56:00) actually been as a result of this AI (00:56:02) infrastructure problem right because (00:56:04) people look at the demand (00:56:05) >> finally that's right demand (00:56:07) >> they look at the demand and the demand (00:56:09) is driving people to create massive of (00:56:11) new battery companies, solar (00:56:13) concentrators. It's put new energy be (00:56:15) new energy like you know willpower (00:56:18) behind (00:56:18) >> SM the AI industry is driving all of (00:56:22) that sustainable energy industry. (00:56:24) >> Yeah. (00:56:24) >> Um because people see that there is (00:56:26) going to be demand for it right so even (00:56:28) if and I think there is no practical (00:56:30) answer in the small number of years time (00:56:32) frame versus uh large gas right um uh it (00:56:37) still drives climate innovation. Yeah, (00:56:38) no question about it. No question about (00:56:40) it. And I I think that's exactly right (00:56:42) that that you know doomer messages um (00:56:46) causes policy and that policy may may (00:56:50) affect the industry in some way. But (00:56:52) there's nothing more powerful than (00:56:53) demand. Look at all the jobs that's (00:56:55) being created. Look at all the the (00:56:56) industries that's being formed around (00:56:57) it. um sustainable energy likely and (00:57:00) when history rewrites it as Sarah, I (00:57:02) think you you're going to be absolutely (00:57:03) right that that if not for AI, well AI (00:57:08) was is probably the biggest driver for (00:57:10) sustainable energy ever. (00:57:11) >> Yeah. A friend of mine has a saying that (00:57:13) uh doomers are the people who sound (00:57:14) smart at dinner parties and optimists (00:57:16) are the people who drive humanity (00:57:17) forward. And I think that's very true (00:57:18) for for all these things we've talked (00:57:20) about. Yeah. So (00:57:21) >> yeah, it's really true. (00:57:22) >> Yeah. Well, that that's one of the big (00:57:24) big um takeaways for for uh this last (00:57:27) year, the battle of narratives. (00:57:30) >> And it's too simplistic (00:57:33) um to say that everything that the (00:57:35) doomers are saying are irrelevant. (00:57:38) That's not true. A lot of very sensible (00:57:40) things are being said. Um it is too (00:57:42) simplistic to say that when somebody is (00:57:44) optimistic that they're just naive. (00:57:46) >> It needs to be grounded in reality. (00:57:48) Yeah, that optimistic people are just (00:57:50) naive, you know, (00:57:51) >> and that that's obviously not true. (00:57:54) >> Um, but I think we just have to be (00:57:55) mindful of the balance of it. (00:57:59) >> When 90% of the messaging is all around (00:58:02) the end of the world and doom and the (00:58:04) pessimism and you know, I think we we're (00:58:07) scaring people (00:58:08) >> from making the investments in AI that (00:58:11) makes it safer, more functional, more (00:58:13) productive (00:58:14) >> and more useful to society. And so we (00:58:16) just, you know, more secure. We, you (00:58:18) know, all of that takes technology. (00:58:20) Security takes technology. Safety takes (00:58:22) technology. I appreciate that my car is (00:58:24) safer today because it has better (00:58:26) technology than a car 50 years ago. (00:58:28) >> And so so I I think it takes technology (00:58:30) to be safe, technology to be secure. And (00:58:33) so I I'm I'm I'm delighted to see that (00:58:36) the the advancement of technology is (00:58:38) still accelerating and ongoing. And so (00:58:41) we just have to make sure that the the (00:58:42) policy makers around the world, the (00:58:44) governments um are able to are are (00:58:47) thinking about balancing these two (00:58:49) ideas. (00:58:50) >> How do you So I guess we've talked a lot (00:58:52) about 25 (00:58:53) >> and the narratives of 25. How do you (00:58:54) think about 26? What are you excited (00:58:56) about? What do you see coming? What do (00:58:58) you think are big changes that we should (00:58:59) be aware of? (00:59:00) >> I am optimistic that that um our (00:59:04) relationship with China will improve. (00:59:05) Mhm. [clears throat] (00:59:06) >> that President Trump and the (00:59:08) administration um has a really really (00:59:11) grounded and common sense um attitude (00:59:14) about um and philosophy around around (00:59:17) how to think about China that that (00:59:20) they're an adversary (00:59:22) >> um but they're also also a partner in (00:59:24) many ways and that the idea of (00:59:27) decoupling is naive and the idea of (00:59:30) decoupling um for whatever reason (00:59:33) philosophical reasons or national (00:59:34) security reasons It's just not not it's (00:59:37) not based on any common sense and the (00:59:39) more you the more deeply you look into (00:59:41) it the more the two countries are (00:59:44) actually highly coupled. (00:59:47) >> Um both countries ought to ought to (00:59:48) invest in their own independence. Um I (00:59:52) you know when you depend too much on (00:59:53) someone the relationship becomes too (00:59:55) emotional uh as you know [laughter] and (00:59:58) so it's good to have some independence (01:00:00) or as much independence as either either (01:00:02) would like but to recognize that there's (01:00:04) a lot of coupling a lot of dependence (01:00:06) between the two countries and and I (01:00:08) think there's a there needs to be a (01:00:10) nuanced strategy a nuanced attitude (01:00:13) about how to how to how to manage this (01:00:15) relationship in a productive way for all (01:00:17) of the people of two countries and for (01:00:20) all of the people around the world, (01:00:21) everybody depends on a productive, (01:00:24) constructive relationship of the two (01:00:26) most important nations and the single (01:00:29) most important relationship for the next (01:00:30) century. And so we have to find that (01:00:32) answer. And I'm I'm I I'm just really (01:00:35) delighted uh that President Trump is (01:00:37) looking for a constructive answer. And (01:00:40) so I I think that next year uh will be a (01:00:42) much better better better year than the (01:00:44) last several. I'm happy with the (01:00:46) administration was able to to to suggest (01:00:49) a a an export control um policy that is (01:00:54) grounded on national security (01:00:56) recognizing that they already make so (01:00:59) many chips themselves and they they can (01:01:01) depend on Huawei themselves for their (01:01:03) military for their national security. (01:01:05) they got ample technology to do that. (01:01:08) And so that American technology, (01:01:10) although general purpose um is unlikely (01:01:13) to be used by their military because (01:01:14) their military is too smart, just as our (01:01:16) military is too smart to to use their (01:01:18) technology. And so it's grounded on (01:01:20) national security. It's grounded on on (01:01:23) uh technology leadership. It's grounded (01:01:25) on national prosperity. You know, one of (01:01:28) the things that that we just always have (01:01:29) to remember is that the world's (01:01:31) mightiest military uh is supported by (01:01:34) the world's mightiest mil economy. And (01:01:37) so the wealth that we generate um brings (01:01:40) jobs home, creates prosperity in the (01:01:42) United States, um provides for tax (01:01:44) revenues, and ultimately funds the (01:01:47) mightiest military on the planet. And so (01:01:49) that circular system, that (01:01:51) interconnected system requires a nuanced (01:01:54) strategy. and and um uh and and and and (01:01:57) I'm I'm I'm pleased to to to to see some (01:02:00) of the progress in that area that allows (01:02:03) American technology companies to keep (01:02:05) America first and keep America ahead (01:02:08) >> and to to support American technology (01:02:10) leadership on the one hand um to win (01:02:13) globally (01:02:15) >> and and then and then China of course is (01:02:17) sorting itself out you know I mean not (01:02:19) sorting but they're sorting out the (01:02:21) attitude about how to think about (01:02:22) American technology and there (01:02:24) >> because historical argument there has (01:02:25) been that if if you look for example at (01:02:27) the internet um there was what was known (01:02:29) as a great firewall right China (01:02:31) basically (01:02:31) >> prevented US competition into China (01:02:33) while the opposite wasn't as true (01:02:36) >> um there's been mass expatriation of US (01:02:38) jobs and industry to China as sort of (01:02:40) part of the development of the 90s and (01:02:42) 2000s and so I think a lot of the things (01:02:43) that people have brought up from a China (01:02:45) US policy perspective besides just the (01:02:47) military adversarial relationship um or (01:02:50) spheres of influence or you know all the (01:02:51) various things like that is also just (01:02:53) the economic imbalances that have (01:02:54) perceived to exist between the two (01:02:56) countries. The way that I would think (01:02:58) through that is go back to the first (01:03:00) principles of technologies again (01:03:03) >> and and let's say the internet you have (01:03:05) the chip industry you have the systems (01:03:07) industry the software industry you have (01:03:09) the services industry on top remember (01:03:11) China's internet growth has been a boon (01:03:15) for Intel and AMD selling CPUs (01:03:18) >> Micron selling DRAM skinex and Samsung (01:03:21) selling DRAM (01:03:23) >> it is the second largest internet (01:03:26) market for American technology industry (01:03:29) >> and so so maybe maybe it wasn't helpful (01:03:32) to some layer of the stack (01:03:33) >> the Googles of the world (01:03:35) >> but don't exclude every layer of the (01:03:38) stack always come back every single one (01:03:40) of these things take a step back and (01:03:41) look at the whole stack (01:03:43) >> maybe that's a theme for today as well (01:03:44) and it makes sense that you would you (01:03:46) would send this message but you know (01:03:47) technology is actually not just the the (01:03:50) sort of internet software application (01:03:52) layer that's been very dominant for two (01:03:54) decades (01:03:54) >> it's the whole stack and Remember as as (01:03:57) as Intel and AMD prospered (01:04:01) >> uh with the internet industry uh in (01:04:03) China growth the China industry growth (01:04:05) don't forget China also contributed (01:04:07) tremendously to open source. No country (01:04:10) in the world contributes more to open (01:04:11) source than China. And look at all the (01:04:13) startups here in America that were able (01:04:15) to benefit from that open source to (01:04:17) create the the new startups that are (01:04:18) here. And so you can't look at one area (01:04:23) in isolation. You have to look at the (01:04:25) whole life cycle of the technology and (01:04:28) look at every layer of the stack. Does (01:04:29) it make sense? When you take a look at (01:04:31) that from that lens, (01:04:33) >> China's internet industry (01:04:36) generated enormous prosperity for (01:04:39) America. (01:04:40) >> Mhm. (01:04:41) >> Just not at the internet company per se. (01:04:44) >> Jensen, my other investor friends will (01:04:45) not forgive me if I don't ask you about (01:04:47) 2026 um uh on the business side. Uh are (01:04:52) we in an AI bubble? AI bubble. Yeah, (01:04:54) there's a lot of ways to reason through (01:04:55) that. (01:04:56) >> And so, so again, um, you know, when (01:05:00) when asked that question, my mind goes (01:05:02) to what is AI and where are we in that? (01:05:07) There's AI, (01:05:09) then there's computing. You know, as you (01:05:11) know, Nvidia invented accelerated (01:05:14) computing. Accelerated computing does (01:05:16) computer graphics and rendering. AI (01:05:17) doesn't. Um, accelerated computing does (01:05:20) data processing, SQL data processing. AI (01:05:22) doesn't. (01:05:23) >> Um, accelerated computing does molecular (01:05:25) dynamics and quantum chemistry. AI (01:05:27) doesn't. You know, all these are all (01:05:29) things that people could say someday AI (01:05:31) will, but it doesn't today. Accelerated (01:05:33) computing is really essential for uh (01:05:35) classical machine learning, XG boost, (01:05:38) recommener systems, the whole process of (01:05:40) uh feature engineering, extract, load, (01:05:42) and transform. That entire data science, (01:05:46) machine learning life cycle, accelerated (01:05:48) computing is used for all of that. The (01:05:50) first thing to go to is in the context (01:05:52) of Nvidia. (01:05:54) What we see is the the the dynamic is (01:05:57) the shift from general purpose computing (01:05:59) to accelerated computing because MOS (01:06:01) laws largely ended. You can't use CPUs (01:06:03) for everything anymore like you used to. (01:06:06) And so it's just no longer productive (01:06:09) enough. It's not deflationary enough. (01:06:12) >> And so so we have to move towards a new (01:06:14) computing model. And that's where (01:06:15) accelerator comes in. If you if (01:06:17) generative AI well excuse me if chatbots (01:06:20) let's just go you know open AI and (01:06:22) Anthropic and Gemini if none of that (01:06:24) existed today Nvidia would be a (01:06:27) multiundred billion dollar company and (01:06:29) the reason for that is because as you (01:06:31) know the foundation of computing is (01:06:33) shifting to accelerated computing (01:06:36) >> that's the first thing to to realize is (01:06:38) is to take a step back and ask yourself (01:06:40) what is actually happening now the next (01:06:43) layer up the question about AI now (01:06:45) becomes What is AI? Now, we ask that we (01:06:48) ask the AI bubble question and we always (01:06:51) go back to OpenAI's revenues 100%. Don't (01:06:54) we? (01:06:54) >> Mhm. (01:06:54) >> You ask somebody, hey, is there an AI (01:06:56) bubble? Everybody goes directly to (01:06:59) OpenAI's revenues. First of all, if (01:07:01) OpenAI currently has twice the capacity, (01:07:04) their revenues would double. You guys (01:07:05) know that if they have 10 times the (01:07:07) capacity, their I really believe their (01:07:09) revenues would 10 times. And so, they (01:07:11) need capacity. This is no different than (01:07:13) Nvidia needs wafers from TSMC. Just (01:07:16) because you know Nvidia exists and and (01:07:18) we're doing great doesn't mean we don't (01:07:19) need capacity. We need capacity. We need (01:07:21) capacity of DRAM. We need and so in our (01:07:23) world it's sensible to everybody. We (01:07:25) need capacity. Well, in their world they (01:07:26) need factories (01:07:27) >> and if they don't have factory capacity (01:07:29) how they generate tokens, which is where (01:07:30) we started our conversation today and so (01:07:33) they need factory capacity in order to (01:07:35) increase their revenue growth. But (01:07:37) nonetheless, we also said that AI is (01:07:41) more than chatbots. It includes all (01:07:43) these different fields of science. Um, (01:07:45) Nvidia's AV business is coming up on 10 (01:07:48) billion dollars. Nobody ever talks about (01:07:50) that. And you have to train world (01:07:52) models. You have to train these AI AVs (01:07:54) and it's happening robo taxis happening (01:07:56) all over the world. Our AI work with uh (01:07:59) digital biology, our AI work in (01:08:01) financial services. The whole industry (01:08:04) of quants, quantitative trading is (01:08:07) moving towards Yeah, exactly. They used (01:08:10) to be classical machine learning. A (01:08:12) whole bunch of human featured they call (01:08:14) quants, right? These these specialized (01:08:17) mathematicians were trying to figure out (01:08:19) what the predictive features are. Now we (01:08:21) use AI to figure it out. And so in order (01:08:23) to have instead of having quants, you (01:08:25) need a lot of supercomputers. Financial (01:08:27) services is one of our fastest growing (01:08:28) segments. billions of dollars in in (01:08:31) quants, you know, in financial services, (01:08:33) billions of dollars in AV, billions of (01:08:36) dollars in robotics coming up, billions (01:08:38) of dollars in digital biology. And so (01:08:41) how big can that all that be? Well, (01:08:43) simple logic is this simple math. (01:08:45) Whether you you think that AI is going (01:08:47) to replace shortage, labor shortage or (01:08:50) workforce shortage in any kind, um, (01:08:53) let's ignore that for a second. The (01:08:55) world is at hundred trillion dollars in (01:08:57) GDP. out of that let's just say 2% 2% (01:09:02) annually is R&D and let's just go back (01:09:05) in time five years ago if you were to (01:09:07) take the largest drug discovery company (01:09:09) in the world drug company in the world (01:09:10) and where's all of their R&D wet labs (01:09:14) >> today what are they do doing building (01:09:17) supercomputers (01:09:19) >> and so there's a fundamental shift in (01:09:22) how they think about that $2 trillion (01:09:25) >> it used to be $2 trillion for the old (01:09:27) way of doing things. It's now going to (01:09:28) be $2 trillion in the AI way of doing (01:09:31) things. Well, $2 trillion is going to (01:09:33) need $2 trillion of R&D is going to be (01:09:36) powered by a whole bunch of (01:09:37) infrastructure. And that's the reason (01:09:39) why we're building supercomputers (01:09:41) everywhere around the world. And so so I (01:09:43) think if if you reason about it from the (01:09:45) outside in, you know, either from the (01:09:48) foundation up, from the outside in, you (01:09:50) come to the conclusion that what we're (01:09:52) experiencing, what all three of us are (01:09:54) experiencing, which is the amount of (01:09:56) computing demand is insane. (01:09:59) >> Give me an example of a startup company (01:10:01) that goes, "No, we're good." (01:10:03) >> They are all dying for computing (01:10:04) capacity. Give me an example for a (01:10:07) researcher in any university, a (01:10:09) scientist in any company who says got (01:10:11) plenty of capacity. Everybody is dying (01:10:14) for capacity. And so we have a global (01:10:17) multi- company, multi-industry shortage. (01:10:21) It's not just about open AI even though (01:10:22) open AI could use a lot more capacity as (01:10:24) well. So I think I think how we think (01:10:26) about this what with the narrative the (01:10:29) narrative is not helpful and it's a (01:10:31) little bit too superficial to say how do (01:10:34) you prove there's an AI bubble$12 (01:10:36) billion of revenues (01:10:39) hundreds of billions of dollar (01:10:40) infrastructure being built is a little (01:10:42) bit too simplistic. (01:10:44) >> Yeah. The other thing people um tend to (01:10:45) point out is the MIT study. There (01:10:47) there's some study that I think came out (01:10:48) of MIT that claimed that most enterprise (01:10:50) deployments of AI weren't that useful. (01:10:52) And you're like, well, did you do the (01:10:53) change management? Did you do a reorg? (01:10:55) Did you integrate into tooling? Did you (01:10:57) like how long did it even take to (01:10:58) implement it? If a planning cycle in an (01:10:59) enterprise is a year and is something in (01:11:01) six months and so it feels like there's (01:11:03) a lot of these kind of again overstated (01:11:05) things that get a lot of attention, but (01:11:07) then you map it against what's actually (01:11:08) happening. (01:11:09) >> Yeah. (01:11:10) >> And the growth of these companies using (01:11:11) AI and it's just a completely different (01:11:12) world. And and and if you want to find (01:11:14) out where the world's innovation's (01:11:17) happening, I would not go find out at an (01:11:19) enterprise. (01:11:20) >> Would you guys agree? (01:11:22) >> Yeah. (01:11:22) >> Enterprise is like the slowest adopters (01:11:25) of new technologies. I would go talk to (01:11:27) all of the startups, the 30, 40,000 (01:11:30) startups that are currently doing this (01:11:32) stuff. I would go talk to Open Evans. (01:11:34) How how's it working? I would go go talk (01:11:36) to cursor. How's coding working by the (01:11:37) way? You know, I would just go talk to (01:11:39) these people. (01:11:40) >> I think it's really interesting that you (01:11:41) see that. Um, of course you do have (01:11:43) companies making, you know, hundred (01:11:45) million plus, multiund million plus (01:11:47) progress of AR in enterprise sales, (01:11:50) Harvey, Sierra, etc. But some of the (01:11:52) fastest growing companies have been (01:11:54) enduser adopted even in conservative (01:11:57) industries, right? Like healthcare, you (01:11:59) know, skeptical industries like (01:12:00) engineering, (01:12:01) >> healthare, the most right, the most (01:12:03) conservative of all. But guess what? (01:12:05) They are so concerned about getting the (01:12:08) right answer (01:12:10) >> that the ability to have something like (01:12:12) open evidence. (01:12:12) >> Yeah. (01:12:13) >> To do grounded research, high quality (01:12:15) research and get that get that research (01:12:18) as information to you. Nobody wants to (01:12:20) do research. They want answers. Nobody (01:12:22) wants to do search. They want answers. (01:12:23) Is that right? (01:12:24) >> A bridge is a great example of that too (01:12:25) where they're basically making it really (01:12:26) easy to do the physician knows instead (01:12:28) of the physician sitting there and doing (01:12:29) it. Back to your point on task versus (01:12:31) >> task versus purpose. Exactly. And I (01:12:32) think a different way to think about the (01:12:33) demand is like there are so many jobs (01:12:35) where you're asking the the work is (01:12:38) actually like an impossible ask right of (01:12:40) a doctor or a radiologist keep up with (01:12:43) the world's biomedical knowledge in R&D (01:12:46) which is accelerating you know computing (01:12:47) and otherwise um and then (01:12:49) >> like archive papers (01:12:51) >> there was a time you s you and I read (01:12:54) >> you and I both both used to do I don't (01:12:56) do that anymore but here now now I just (01:12:59) load it all into chat (01:13:00) >> GBTh you Now I just load it all in with (01:13:03) all of the the ones that are interesting (01:13:05) and and I make it learn it (01:13:07) >> and then I you know make it summarize (01:13:09) and another summary and I I interact (01:13:11) with it. But but the point is uh we used (01:13:14) to do search. We don't do it that (01:13:15) anymore. I don't do search. We used to (01:13:17) do research. You know the goal is to get (01:13:20) answers. The goal is to get smarter. And (01:13:22) these AIs allow us to help us do all (01:13:24) that. And I think all of it all of it (01:13:28) comes back with it. It's all more (01:13:31) helpful if you come back to the (01:13:32) framework that says AI is a multi-layer (01:13:36) cake (01:13:37) >> and that AI is not just a chatbot. AI is (01:13:41) very very diverse in all of the (01:13:43) industries and modalities and (01:13:45) information and applications that it (01:13:47) addresses. When you think about wanting (01:13:50) to win (01:13:51) >> that America should win AI, it should (01:13:54) not just be America should have this (01:13:57) company win AI, but it we should try to (01:13:59) win across the board (01:14:01) >> and across domains. (01:14:02) >> Across domains. Exactly. And when we (01:14:04) think about open source, all of a sudden (01:14:06) this this is a helpful framework. When (01:14:08) we think about winning, it's a helpful (01:14:09) framework. When we think about uh energy (01:14:11) is a helpful framework that because we (01:14:13) need factories. Factories need energy. (01:14:16) And without energy, we have no factory. (01:14:17) without factories we have no AI that's a (01:14:19) helpful framework and so I think if if (01:14:22) um if we if we have a better (01:14:25) understanding a system a framework for (01:14:28) understanding what AI is I think the (01:14:30) narratives will become more common sense (01:14:32) the narratives will become more (01:14:33) pragmatic (01:14:35) >> become more balanced we want to keep (01:14:37) people safe (01:14:38) >> but one of the best ways to keep people (01:14:40) safe is advancing advancing of (01:14:42) technology quickly (01:14:43) >> and and I think the industry is doing (01:14:45) that and I'm very proud of the industry (01:14:47) for doing that. (01:14:48) >> No one wants to drive a car from, you (01:14:50) know, the first decade of cars. And so I (01:14:53) I think uh (01:14:54) >> ABS is a really good thing. (01:14:56) >> Yes, (01:14:56) >> ABS is a really good thing. Lane keeping (01:14:58) is a really good thing. There's no (01:15:00) question FSD is a really good thing. (01:15:02) >> And I think people will be excited about (01:15:04) the, you know, third or fourth year of (01:15:06) AI. (01:15:07) >> Yeah. No, no doubt. And and I I say with (01:15:10) great pride that the industry made (01:15:14) tremendous strides this last year. (01:15:17) all the technologies we've mentioned. (01:15:19) Um, and that the scaling laws are so (01:15:21) intact that we we now know that more (01:15:26) compute, more intelligence (01:15:28) >> and and um uh gosh, the the the the (01:15:34) innovations in one in in one sector (01:15:37) diffuses and spreads across all of the (01:15:39) other sectors so fast. I'm so happy to (01:15:42) see all that. And so I think the next (01:15:44) five years it's going to be (01:15:45) extraordinary. No, no doubt about it. (01:15:46) And I think next year is going to be (01:15:48) incredible. (01:15:48) >> Amazing. Well, we're excited to talk to (01:15:50) you at the end of next year, too. (01:15:51) >> Yeah. Looking forward to it. Thank you (01:15:53) guys for all the work that you guys do. (01:15:54) Congratulations. What a great year. (01:15:55) >> Wow. Amazing year. (01:15:57) >> Yeah. A lot. Thank you. (01:15:58) >> Yeah. Thank you. Happy New Year. Happy (01:16:00) New Year. (01:16:03) >> Find us on Twitter at No Prior Pod. (01:16:05) Subscribe to our YouTube channel if you (01:16:07) want to see our faces. Follow the show (01:16:09) on Apple Podcasts, Spotify, or wherever (01:16:11) [music] you listen. That way you get a (01:16:13) new episode every week. And sign up for (01:16:15) emails or find transcripts for every (01:16:17) episode at no-briers.com.