NVIDIA's Jensen Huang on Reasoning Models, Robotics, and Refuting the "AI Bubble" Narrative
00:00:05Hson, thanks so much for joining ustoday.>> So great to have you guys. What anamazing year.>> What a year.>> Happy Hanukkah, merry Christmas,>> happy new year coming up. Yep. Happyholidays.>> So, uh, with everything that's happenedin 2025,um, and you know, being in the middle ofthe vortex with it, what do you reflecton and say like this surprised you mostor this is the biggest change?>> Let's see. There there's some thingsthat didn't surprise me like for examplethe scaling laws didn't surprise mebecause we already knew about that. Thetechnology advancement didn't surpriseme. I was pleased with the improvementsof grounding. I was pleased with theimprovements of reasoning. I was pleasedwith uh uh the connection of all of themodels to to to search. I'm pleased thatit that uh there are now routers thatare in front of these models so that itcould depending on the confidence of theanswers go off and do necessary researchand and just generally improve thequality and the accuracy of answers.>> I'm hugely proud of that. I think thewhole industry addressed one of thebiggest skeptical responses of AI whichis hallucination and um generatinggibberish and all of that stuff. I Ithought that this year the wholeindustry everything from
00:01:21every everyfield from language to vision torobotics to self-driving cars the theapplication of reasoningand the grounding of the of of of theanswers. Um big big leaps would you guyssay this year?>> Huge. I mean things like open evidencetoo for medical information wheredoctors are now really using that as atrusted resource like you Harvey forlegal you're really starting to see AIemerge as one of these things that'sbecome a trusted tool or counterpartyfor you know experts to actually be ableto do what they do much better.>> That's that's right. And so so in a lotof ways I was expecting it but I'm stillpleased by it. I'm proud of it. I'mproud of all of the industry's work inthis area. I'm really pleased and and uhuh and probably a little bit surprisedin fact that token generation rate forinference especially reasoning tokensare growing so fast several exponentialsat the same times it seems and uh andI'm so pleased that that these tokensare now profitable that people aregenerating I heard somebody hurt todaythat that open evidence speaking of them90% gross margins I mean those are veryprofitable tokens.>> Yeah.>> And so they're obviously
00:02:36doing veryprofitable, very valuable work. Cursor,their margins are great. Uh Claude'smargins are great for the enterprise useof OpenAI. Their margins are great. Umso anyways, it's really terrific to seethat that um we're now generating tokensthat are sufficiently good, so good invalue that that people are willing topay good money for. And so I I thinkthese are are really great grounding forthe year. I mean some of the things thatthe narrative that that um uh of coursethe conversation with China reallyreally you know occupied a lot of my mytime this year. Geopoliticsuh the importance of technology in eachone of the countries. Uh I spent moretime traveling around the world thisyear than just about any time in the hall of my life combined. You know myaverage elevation this year is probablyabout 17,000 ft. You know so so it'snice to be here on the ground with youguys. Um and so so I think uhgeopolitics the importance of AI to allthe nations uh all worth talking aboutlater. You know of course I spent a lotof time on expert control and and makingsure that our strategy is nuanced and uhreally grounded and um uh promotesnational security but recognizing theimportance of various uh various facetsof national security. Um a
00:03:51lot ofconversations around that. Um, you know,of course, of course, uh, lots ofconversation about jobs, the impact ofAI, uh, energy,>> um, uh, labor shortage. I mean, boy, wecovered everything, did we? Yeah.Everything was AI.>> Everything was AI. Yeah, it wasincredible.>> Yeah, AI was definitely the center ofthe storm for like every one of thosethemes. Maybe one we can start withactually um is jobs because or therejobs and employment because when I lookat the traditional AI community evenbefore things were scaling and evenbefore AI was really working there was astrong sort of doomsday component in thepeople working on AI oddly enough rightthe people who were most trying to pushthe field forward were often the peoplewho are most pessimistic which is veryodd why would you do both at once>> and I feel like that narrative has takenover some subset of media or some set ofother things despite all the things thatwe think are very positive about what AIhas done That's going to help withhealthcare, with education, withproductivity, with all these otherareas.>> And in in general, whenever we have atechnology shift, you have a shift interms of the jobs that are important,but you still have more jobs.>> That's right.>> Could you talk about how you think aboutemployment and jobs and sort of whatpeople are saying and what you think thereal narrative is there?>> Maybe what I'll do is I'll I'll groundit on uh three points in space, threepoints in time. now.>> Mhm.>> Uh maybe uh
00:05:06uh very near future and thensome some point out out in the distanceand and maybe maybe somecounternarratives.>> Um something else to think about withrespect to jobs in the near term.>> Uh one of the most important things isthat that AI is not just AI is software>> but it's not pre-recorded software asyou know. For example, Excel was writtenby several hundred engineers. Theycompiled it. It's pre-recorded and thenthey distribute it as is for severalyears. In the case of AI, because ittakes into the context, what you askedof it, what's happening in the world,right? Contextual information, itgenerates every single token for thefirst time, every time.>> Which means every time you use thesoftware and and everything that we do,AI is being generated for the first timeever. Just like intelligence, ourconversation today relies on some, youknow, ground truth and some knowledgeand but it's every single word is beinggenerated for the first time here. Thething that's really really quite uniqueabout AI is that it needs thesecomputers to generate these tokens everysingle time. I call them AI factoriesbecause it's producing tokens that willbe, you know, used all over the world.Now,
00:06:21some people would say it's alsopart of infrastructure. The reason whyit's infrastructure is because obviouslyit affects every single application.It's used in every single company. It'sused in every single industry everysingle country. Therefore, it's partinfrastructure like energy and andinternet. Now, because of that and theamount of computers that's necessary togenerate these tokens and it's neverhappened before and because we needthese factories, three new industrieshave emerged. Number one, well, threenew type of plants have to be created.Number one, we have to build a lot morechip plants.>> Mhm. TSMC is building, right? SKH Highixbuilding a lot more plants and so weneed more chip plants. We need morecomputer plants. These computers arevery different. These are supercomputersthat the world's never seen before.Right. Grace Blackwell looks like a verydifferent type of computer than anythingthat's ever been made. And entire rackis one GPU.>> And so we need new supercomputer plants.And then we need new AI factories. Thesethree plants are currently being metbeing built in the United States at verylarge scale quite broadly all over theUnited States for the very first time.>> The number of construction workers,plumbers, electricians, technicians,network engineers, you know, right? Thethe number of the
00:07:37skilled labor that'snecessary to support this new industryin the near term,>> it'll be enormous. Let's just face it.Uh I'm [clears throat] so excited tohear that electricians are seeing theirpaychecks double. They're being they'rebeing paid to travels like like us. Wego on business trips. They're going onbusiness trips. And so it's reallyterrific to see, you know, that thisthese three industries are now threetypes of plants, factories are justcreating so much so much jobs. The nextpart is the the near-term impact of AIon jobs. And one of my favorites is um Ilove Jeff Hinton. uh he said uh you knowsome five six seven years ago that infive years time uh AI will completelyrevolutionize radiology that everysingle radiology application will bepowered by AI and that radiologistsuh will no longer be needed and that hewould advise this the first professionnot to go into is radiology and he'sabsolutely right 100% of radiologyapplications are now AI powered. That'scompletely true and in some eight yearstime it is now
00:08:53completely pervaded uh uhradiology. However, what's interestingis that the number of radiologistsincreased>> and so now the question is why and thisis where the difference between taskversus purpose of a job. A job has tasksand has purpose. And in the case of aradiologist, the task is to study scans,but the purpose is to diagnose disease>> and to research>> and and that exactly and they're doingresearch. And so in the case in theircase, the fact that they're able tostudy more scans more deeply,um they're able to uh request morescans, do a better job diagnosingdisease, the hospital's more productive,they can have more patients, whichallows them to make more money, whichallows them to want to hire moreradiologists. And so the question iswhat is the purpose of the job versuswhat is the task that you do in yourjob? And and as you know, I spend mostof my>> day typing. [snorts] That's my task, butmy purpose is obviously not typing. Andso the fact that somebody could use AIto automate a lot of my typing, and Ireally appreciate that, and it helps alot.>> Um, it hasn't really made
00:10:08me, if youwill, less busy. In a lot of ways, Ibecome more busy because I'm able to domore work. So, I think that that's thesecond part to consider is the taskversus the purpose of the job. Thisexample really strikes home because mymy sister-in-law Erin actually leads umin nuclear medicine at Stanford, right?So, she's in radiology and with all thetechnology advancements that are coming,>> these doctors really welcome it and theyare working 20 hours a day trying to domore research and serve more patients.Exactly. And and I think one thing thatis often missed beyond the sort of um uhdiversity of jobs being created by thisinvestment in infrastructure is actuallyhow much latent demand there is fordifferent goods that we we need insociety like better healthcare. I don'tthink anybody feels like you know whatwe have reached the the tiptop uhmountaintop of like what Americanhealthcare or global healthcare could beand um the more we can make these peopleproductive the more demand there will be>> that's exactly right if I if Nvidia wasmore productive it doesn't result inlayoffs it results in us doing more morethings>> I met your new hire class today you seemto be hiring every week anyway yeah>> that's exactly right right the the moreproductive we are the
00:11:23more uh ideas wecan explore uh the more growth as as aresult the more profitable we becomewhich allows us to pursue more ideas andso I think you're you're absolutelyright that that if if the job if if yourif your life if the world the problemsis literally already specified andthere's no other problem to solve thenproductivity would actually reduce theeconomy but it's clearly going toincrease the e economy I think that theNext part that I would consider is, youknow, people say, gosh, all of theserobots that we're talking about, it'sgoing to take away jobs. As as we knowvery clearly, we don't have enoughfactory workers. Our economy is actuallylimited by the number of factory workerswe have. Most people are are having avery hard time retaining their workers.Um, we also know that the number oftruck drivers in the world is severelyshort. And the reason for that is peopledon't want those jobs where you have totravel across the country and live indifferent parts of the world, differentparts of the country, you know, everysingle night. So people want to stay intheir town, stay with their families.And so I think that I think the firstpart is that having robotic systems isgoing to allow us to cover the labor
00:12:38shortage gap which is really reallysevere and getting worse because ofaging population. This is this is notonly United States, it's all over theworld as you guys know.>> And so we're going to cover the laborshortage. But the second part thatpeople forget and and as a result we'llgo there are shortages as well in otherplaces that people talk about AI beingrelevant. Accounting would be an examplewhere there's shortages there. Nursingis another example. So you know you canyou can go through multiple otherindustries and say okay there's gapsright>> and AI is trying to help fill thosegaps.>> That's exactly right. And so so umautomation is going to help us increaseand solve the the the the labor gap. Nowpeople also don't don't remember thatwhen we have cars, we need mechanics totake care of our cars.>> And if you look at the robo taxis thatare that are even on the streets today,it's taken 10 years for that to happen.Look at all the maintenance crews andall of the the the various, you know,hubs that they're in where you have totake care of these robo taxis and justimagine we have a billion robots.>> Mhm.>> It's going to be the largest repairindustry on the planet. So I I think alot of people don't they they just haveto think through>> and this is the part where you said umwhen we create this type of automation,
00:13:54we create this other job. Right now lookat AI is creating so many jobs. Mhm.>> The AI industry is creating a boom ofjobs.>> I think one of the core challenges hereis it's very easy to draw a straightline of extrapolation from like oh youknow uh there are tools that helplawyers be more productive. It's goingto replace the lawyers but it's actuallyit takes like a step of incrementalreasoning to say there's a sucking soundin the economy for everything in AIinfrastructure. there's actually asucking sound toward all of this demandthat is latent in the places where wehave gaps where um I think a lot ofpolicy makers have focused on you knowwe can't replace or reduce what we havewhen it's really there's there's farmore demand in what we actually are not>> and in the case of lawyer what's thewhat's the purpose of the lawyer versusthe task of the lawyer>> reading a contract writing a contract isnot the purpose of the lawyer thepurpose of the lawyer is to help youresolve conflictAnd that's more than reading a contract.It's more than writing a contract. Thepurpose is to protect you. That's morethan reading a contract. It's more thanwriting a contract. And so I think justit's really really important to go backto what is the purpose of the job versus
00:15:09the task that we use,>> you know, to perform that job thatchanges over time.>> Yeah. The other big theme of the yearthat you mentioned that I think isreally important to touch upon is bothuh China is sort of in the rise ofChinese open source in particular whereyou know some of the highest scoringmodels against benchmarks now areChinese models on the open source sideon the closer side it's still a lot ofthe US models but things like QuinnDeepseek etc>> are doing very well you've long been aproponent for open source in generalcould you could you share views aboutboth China emerging for AI for opensource and what the US should be doingin terms of both open source as well asits own industries>> when you Think about these complicatedinterconnected dependentum networks of problems. These this youknow big goop of a mesh of problems it'salways good to to go back and find aframework for what it is that we'retalking about. In the case of AI um whatis AI?Well, of course, the technology of AIand the capability, the capabilities ofAI is about automation. It's aboutautomation of intelligence for the veryfirst time. And you could combine itwith megatronics technology to embodythat megatronics and and
00:16:24make it performtasks.>> So that's what's AI automation. But whatwhat is the stack that makes AIpossible? What's the technology stack?functional stack. And of course the ethe easiest way to think about that isit's kind of like a fivey year five yearfive layer cake which is at the lowestlevel is energy.>> Um it transforms energy to the outputthat I just described. The next layer ischips. The next layer is infrastructureand that infrastructure is both hardwaresoftware right this is where land powerand shell this is where construction isdata centers are the software stack>> you know for orchestrating the so it'ssoftware and hardware the layer abovethat is where everybody thinks aboutwhich is AI which is the models>> we know this but it's really helpful tounderstand that AI is a system of models>> and AI is a um a techn technology thatunderstands information and there'shuman information and so we often timesthink about AI as a chatbot>> but remember there's biologicalinformation there's chemical informationthere's physical information of allkinds there's financial informationthere's healthcare information there's fthere's information of all
00:17:39modalitiesall kinds AI is really really broad andof course human language is at thefoundation of of many things but it'snot the essence of everything because asyou know you know biology moleculesdon't understand English>> they understand something else rightproteins don't understand English theyunderstand something else I think thenext layer the important thing is is uhthat's where the AI models are butthere's a whole the AI is very verydiverse and then the the layer abovethat is is applications and it dependson the industry and you alreadymentioned open evidence there youmentioned Harvey there's cursor there'sall kinds of right there's all kinds ofapplications full self-driving is reallyan application, an AI application thatis embodied into a mechanical car>> and figure is a AI application that hasbeen embodied into a mechanical human.And so, so you got all these differentapplications. Well, this five layerstack is one way of thinking about it.And then the next way of thinking aboutI just mentioned is AI is reallydiverse. When you now have thisframework of what the the technologycapabilities are, how to how to buildthe technology and how diverse it is,then you can come back and think aboutokay, let's ask the question, howimportant is open source?
00:18:54Well, withoutopen source, you know, today, of course,the frontier models, the the the leadinglabs have chosen to to use a closedsource um application approach, which isjust fine. you know what people decideto do with their business models is isreally in the final analysis. It's theirbusiness and they have to they have tocalculate what is the best way for themto get the return on investment so thatthey could scale up and and make betteradvances. Um however they they made thatcalculus is fantastic. On the otherhand, uh without open source, as youknow, startups would be challenged, uhcompanies that are in in uh uh differentindustries, whether it's manufacturingor transportation or um it could be inhealthcare. Without open source today,all of that AI work would be suffocated.>> And so, they just need to have somethingthat's pre-trained. They need to havesome fundamental technology aboutreasoning. from that they could alladapt, fine-tune, you know, train theirAI models into exactly the domain andapplication they want. And so whatpeople really really miss is just theincredible pervasiveness and theimportance of open source to
00:20:10all ofthese industries. large companies uhwithout without open source some of someof 100-year-old companies that I workwith>> in in industrial spaces in healthcarespaces they would be suffocated theywouldn't be able to do that>> open source at this point is driving allof our data centers is driving a bigchunk of telefan in the world in termsof Android or other devices it's drivingexactly>> you know to your point a lot of theindustrial applications so it's alreadypervasive and I think the big questionis>> open source without open source highered>> higher ed wouldn't happen>> education research>> startups I mean the list goes on, youknow, and so so>> we talk we talk all day long about thetip but the most visible part of thatthe most the part that's most newsworthymaybe but underneath that is such animportant space of open source AI andwhatever we decide to do with policiesdo not damage that innovation flywheel.So I spent a lot of time uh educatingeducating uh uh policy makers to helpthem understand whatever you decidewhatever you do don't forget opensource. Whatever you decide whatever youdo don't forget biology.I think the counternarrative here thatis worth addressing is that essentially
00:21:25like you know there should be amonolithic vertical player andmonolithic asset in the like one modelthat does it all and that we can't giveaway that crown jewel to other countriesor non-American companies and and youryour argument is like we actually needthis huge diversity of AI applicationsand and the American advantage isactually or any any sovereign advantageis in the whole stack right? Thecapability to deliver any piece of it.>> I guess someday we will have God AI.>> But when is that day?>> But but that someday that someday isprobably on biblical scales, you know, Ithink galactic scales. Um I I think it'sit's not helpful to go from where we aretoday to God AI.>> And um I don't think any companypractically believes they're anywherenear God AI. And nor nor do I do I seeany researchers having any reasonableability to create god AI. The ability toh understand human language and genomelanguage and molecular language andprotein language and amino acid languageand physics language all supremely well.That god AI just doesn't exist.>> And and yet we have a
00:22:40lot of industriesthat need AI. Mhm.>> AI is if if you will at the simplisticlevel, it's just the next computerindustry.>> And give me an example of a company, anindustry, a nation who doesn't needcomputers.>> Mhm.>> And we all don't have to wait around forGod AI for us to advance, right? So GodAI is not showing up next week. I'mfairly certain of that. Okay. And God[clears throat] AI god AI is not notgoing to show up next year, but thewhole world needs to move forward nextweek, next year, next decade. I thinkthat that the idea of a monolithicgiganticcompany,>> country, nation, state that has got AIis just>> it's unhelpful.>> It's unhelpful. It's too extreme.>> Then in fact, if you want to take it tothat level, then we ought to just allstop everything.What's the point of having evengovernments? I mean, why why why arethey doing policies? God AI is going tobe smart enough to avert, you know, workaround any policy. And so, what's thepoint? And so, I I think that that weought to bring things back to the groundground level and start thinking aboutthings practically and and use commonsense.>> This seems to be
00:23:55like a big theme ingeneral in terms of this conversationwhere there's been a lot that's beenkind of put out there that seems veryextreme if you actually think about it.It's the jobs and employment. Nobody'sgoing to be able to work again. It's GodAI is going to solve every problem. It'swe shouldn't have open source for XYZreason despite open source powering muchof our industries already.>> That's right.>> And so it seems like in general maybeone of the themes of 2025 was there's alot of extremes that were sort ofpainted in the public with AI that ifyou look at them very closely don'treally follow a logical change in termsof happening anytime soon.>> Yeah. And so it's it's it sounds likeit's really important to have thisconversation.>> Extremely hurtful frankly. And I I thinkwe've done a lot of damage uh with verywellrespected people um who have whohave painted a doom doomer narrative umend of the world narrative sciencefiction narrative and um you know and Iand I appreciate that that many of usgrew up in and enjoyed science fiction.>> Um but I but it's not helpful. It's nothelpful to people. It's not helpful tothe industry. It's not helpful tosociety. It's not helpful to thegovernments. Mhm.>> There are a lot of many people in thegovernment who obviously aren't asfamiliar with as as comfortable with thetechnology>> and when PhDs of this and CEOs
00:25:10of that>> goes to governments and explain anddescribe these end of the worldscenarios and extremely extremelydystopian future the future. Um, youhave to ask yourself, you know, what isthe purpose of that narrative and whatis their what are their intentions andwhat do they hope? Why are they why arethey talking to governments about thesethings to create regulations tosuffocate startups? [clears throat]>> For what reason would they be doingthat, you know, and so>> and do you think that's just regulatorycapture where they're trying to preventuh new startups from showing up andbeing able to compete effectively orwhat do you think is the goal of some ofthese conversations? you know, I I can'tI can't um uh guess what they what theyhave in mind. I know that the concern isregulatory capture. As a policy, as apractice, I don't think companies had togo toum governments to advocatefor the regulation on other companiesand other industries. just in practicetheir their intentions are clearlydeeply conflicted and and uh theirintentions are clearly you know notcompletely in the best interest
00:26:25ofsociety. I mean they're obviously CEOsare obviously companies and obviouslythey're advocating for themselves>> and so so I think if we can all>> come back to where are we today>> and think about where the technology isgoing to be. I mean look lit literallyin one year's time as we were talkingabout in the beginning uh some of themost proud moments is when the industrywas able to invest very aggressively inadvancing AI technology instead of beingslowed down.>> Remember just two years ago people weretalking about slowing the industry down>> but as we advanced quickly what did wesolve? We solved grounding, we solvedreasoning. We solved research. All ofthat technology was applied for goodimproving the functionality of the AInot you know>> yet the end has not come>> yet the end has not come it's becomemore useful it's become more functionalit's become able to do what we ask it todo you know and so the first the firstpart of the safety of a product is thatit perform as advertised>> the first part of safety is performancethat it's is
00:27:41supposed like the firstpart of safety of a car isn't that someperson is going to jump into the car anduse it as a missile. The first part ofthe car is it works as advertised.>> Mhm.>> 99.999%of the time working as advertised. Andso it takes a lot of technology to makethat car or make that AI work asadvertised. And I'm really glad that inthe last couple two three years theindustry has invested so much inenhancing the functionality of the AI asadvertised. And I think if if we were toto look at the next 10 years, we have somuch work to do to make it work asadvertised. Meanwhile, as as you know,you both of you invest so much in in thein the ecosystem, you see so manycompanies being built for um syntheticdata generation so that the AIs could bemore grounded uh more diverse uh lessbiased more safe uh you're investing ina whole bunch of companies in cybersecurity using AI for cyber security youright people think that there's this AIum the marginal cost of AI is going togo go down significantly and it is>> and therefore the AI is going to bedangerous. It's exactly the opposite.
00:28:56Ifthe marginal cost of AI is going to godown significantly, that one AI is goingto be monitored by millions of AIS.>> Mhm.>> And more and more AI is going to bemonitoring monitoring each other. Peopledon't can't forget that an AI is notgoing to be an agent by itself. It'slikely the AI is going to be surroundedby agents monitoring it. And so it's nodifferent than if the if the marginalcost of of keeping society safe waslower. We have police in every corner.>> So one thing that that we were talkingabout a little bit earlier was just thecost of AI and how it's been comingdown. And so>> I I think um in 2024 the the cost ofGPT4 equivalent models if you look at amillion tokens it came down over 100x.Um you know somebody in my team did thisanalysis to show that. Uh so the costsare dropping pretty dramatically andvery rapidly and part of it is all theadvancements you all have been drivingon and the Nvidia level but also justacross the stack getting big efficiencygains.>> Yeah.>> Um at the same time model companies aretalking about how the costs are risinghow there's enormous sort of capitalmodes to building these things out. Howdo you think about cost of training andcost of inference over time and whatthat means for the average end user orthe average startup company trying tocompete or people trying to do more inthis industry? I forget the statisticthat but but you know Andre
00:30:12AndreCararpathy um estimated the cost ofbuilding the first chatbt I think>> versus now I think you could do that onthe PC now.>> Yeah. Yeah. It's probably tens ofthousands of dollars at this point ormaybe even less.>> Right. And so it costs nothing.>> Mhm.>> And and>> he has an open source project that youcan do in a weekend.>> Oh, is that right? Okay. That'sincredible. Right. We're talking aboutthree years. Mhm.>> Mhm.>> What people people said cost billions ofdollarsum supercomputers built raising billionsof dollars in order to do all that now>> cost you know something that you can doon a weekend on a PC. And so that tellsyou something about how quickly we'remaking making AI more cost effective>> or Spark sorry probably not quite a PC.>> Okay. Not quite a PC. Yeah. We'reimproving our architecture andperformanceum every single year. The first GBTU Ithink was trained on Voltus.>> Mhm.>> And then uh Ampearum you know and and it wasn't I thinkthe first breakthroughs none of itincluded Hopper.>> Mhm.>> And um of course Hopper the last coupletwo three years and um uh we're off inBlackwell for the last year and a halfor so. And um every single one
00:31:27of thesegenerations the architecture improvesand of course the number of transistorsgo up and uh the capacity goes up everysingle generation very easily everyevery single year from a computingperspective. The combination of all thatgetting 5 to 10x every single year>> is not unusual. And here comes Reubenjust around the corner. And so we'reseeing 5 to 10x every single year. Wellcompounded it's incredible. Moore's lawwas two times every year and a half>> and over the course of five years is 10xover the course of 10 years is 100x>> in the in the in the case of AI over thecourse of 10 years is probably 100,000to a millionx okay and that's just thehardware>> then the next layer is the algorithmlayer and the model layer thecombination of all that the fact that ifyou were to tell me that in the cost inthe in the in in the span of you know 10years we're going to reduce the cost oftoken generation about a billion times.I would not be surprised.>> Mhm.>> Okay. And so that's the tokconomicsof of of AI. On the training side, it'snot quite as aggressive in in costreduction, but it's close. If you wereto say that that every single year we'reincreasing by two or 3x over the courseof 10 years, incredible.
00:32:42But theimportant idea is when somebody says itcost $und00 million to train somethingor half a billion dollars to trainsomething.Well, next year it's 10 times less. Nextyear it's 10 times.>> For people to scale these things up,though, right? So the counter argumentis, well, we'll just get bigger everyyear by 10x or 100x or, you know, we'lltry to offset that decrease in cost byscale>> and others can't keep up.>> Yeah. But really what's happening is isyou're and and this is where come in asyou know the scale went up by a factorof 10 but the computational burden didnot go up by a factor of 10 becauseyou're getting the compounded benefitsof all three things. The hardware isgoing up the the algorithms of thetraining models are going up and ofcourse the model architecture is goingup and we're getting the benefit oflearning from each other. This is, youknow, let's face it, Deep Seek wasprobably the single most important paperthat most Silicon Valley researchersread from in the last couple years.>> It was the only thing that felt frontierthat was open.>> That's right.>> In years,the value of open source again puttingout these papers.>> Literally, Deep Seek>> benefited American startups and AmericanAI labs all over>> and infrastructure companies>> and infrastructure company
00:33:57all over.probably the single greatestcontribution to American AI last year.>> And so if you said this out loud, ofcourse, you know, people>> kind of shudder um that we're uhAmerican AI is actually getting learningfrom and benefiting from uh AI fromother nation. But why would that besurprising? You know, AI researchers inall over America, all over America areuh Chinese natives and come fromdifferent countries. We benefit fromevery country. become benefit from everyresearcher and no all of the world'sideas don't have to come from the UnitedStates and so I I think um back to youryour original question it is the casethatyou know some of the narratives aroundaround the cost of AI is about scaringeverybody out of the market you knownobody ought to do pre-training but usnobody should do you know training thesefrontier models but us because thebecause of innovation of modelsalgorithmsand the computing stack, the cost of AIis actually decreasing well more than10x every single year. And so if you'rejust one year behind or even six monthsbehind, you could you could really stayclose.>> And I think one thing that felt verydifferent to me about 2025
00:35:12is um Ilia uhsaid recently that uh you know we're inthe age of research again versus an ageof scaling. I think both things arehappening by the way. Everybody is alsotrying to scale on multiple dimensions.>> Yeah, exactly. both are happening.>> You know, being 6 months behind or beingat 100 versus a 200k cluster, I thinkmatters if you are competingsymmetrically, but now you have peoplefrom Frontier Labs or um at the very topof the game who have very differentideas about how to progress from here orwho are working on diversity ofproblems, right? Uh and and I I thinkthat felt different from 24 maybe wherethere was a lot of energy focused onjust pre-training scale and LLM.>> Yeah. And several several otherdynamics. Um, as the market grows, eachone of these models could choose to haveverticals>> or segments where they want todifferentiate.>> Somebody could decide to be a bettercoder. Somebody could decide to be justbetter at being easier to be accessibleso that it could be a greater consumerproduct.>> You know, the diversity of these models.As a result, you could you couldprobably make a niche leap withouthaving to be great at everything elseand still be super valuable to themarket.>> It's no longer necessary to boil the
00:36:27entire ocean. The f two years ago,because it was called pre-training pre,you know, people people said, well, youknow, pre-training is over. First ofall, pre-training is not over. But thepoint of pre-training is to trainyourself for training. That's why it'scalled pre-training to prepare yourselfto do the real training. And now we callit post-training. It's kind of weird. II think it's just training, butpre-training is pre-training andtherefore it's training. Training as youas as we all know is is where uh computescaling directly translates tointelligence. You you've you've largelynow now this the the data the the datanecessary to train a model is actuallypretty small. Maybe it's just theverifiable results. Now it's reallyalgorithmic, very compute intensive andso and you don't have to be good ateverything in life as you know just likeall of us we don't we could decidebecause we don't have time to learneverything equally well. We decide tochoose a specialty and focus all of ourenergy on it and we become superhuman orincredibly good at something that otherpeople are not. And so I think AI labsare going to start doing the same.They're going to start bifurcating intovarious segments and over
00:37:42time you'regonna and startups will do the same.>> They'll find a micro niche and they'lltake something open and then beincredibly good at it.>> Well, I think one of the most optimisticviews here is uh actually that thesemicroniches are quite valuable, right? Iwas talking to Andre um because I'vebeen talking to a lot of people abouttheir predictions for next year. We'llask you yours as well of course. Um umbut he asked you know what is a what'san example of a prediction that wouldhave been preient last year uh and myanswer everything's easy in retrospectis that coding would be the firstapplication level business that gets toa billion of AR as an AI native appright and I I think if you taken an oldworld view of this>> um you would have believed like one oftwo narratives right one is uh singlemodel does everything and it'll all justbe subsumed into something monolithicMhm.>> And two is that developer tools neverget very big, right? Well, kind ofdepends on how valuable the developertool is. Now, I think many more peopleunderstand software engineering is in aniche and there's more demand than everfor it,>> but I think we'll see more like that.>> Also interesting, uh we are using we weuse cursor here and we use cursorpervasively here. Every engineer uses itand the number of engineers, you justmentioned it, the number of people we'rehiring today is just incredible.>> Yep.>> Right. Monday is
00:38:58come to work at Nvidiaday and and um uh why is that? Uh thisis now the purpose and the task.>> The purpose of a software engineer is tosolve known problemsand to find new problems to solve.Coding is one of the tasks.>> And so if the purpose is not coding, ifyour purpose literally is coding,somebody tells you what to do, you codeit. All right? Maybe you're going to getreplaced by the AI. But most of oursoftware engineers, all of our software,their goal is to solve problems. And itturns out we have so many problems inthe company and we have so manyundiscovered problems. And so the moretime they have to go exploreundiscovered problems, the better off weare as a company. Nothing would give memore joy than if none of them are codingat all. They're just solving problems.>> You see what I'm saying? And so I Ithink that this framework of purposeversus task is really good for everybodyto apply. For example, somebody who's awaiter, their job is to not to take theorder. That's not their job. As it turnsout, their job is so that we have agreat experience. And if somebody ifsome AI is taking the order, their jobor even delivering the food, their jobis still helping us have a greatexperience. They they would reshape
00:40:13their jobs accordingly. And so so Ithink the um the question about aboutcost of compute um uh is reallyimportant. Let's let let me come back toone the the reason why we are sodedicated to a programmable architectureversus a fixed architect. Remember along time ago>> uh a CNN chip came along and they saidNvidia is done.>> And then and then a transformer chipcame and Nvidia was done.>> People are still trying that. Yes.>> Yeah. NP and and the benefit of thesededicated AS6 of course it could performa job really really well andtransformers is a much more universal AInetwork but the transformer as you knowthe species of it is growing incredibly>> the attention mechanism>> the attention mechanism how it thinksabout contextdiffusion versus auto reggressive>> these hybrid SSM transformation>> hybrid SSM for example Neotron we justannounced a new hybrid SM SM and and sothe architecture of transformer is infact changing very rapidly and over thenext several years it's likely to changetremendously and so we we dedicateourselves to an architecture that'sflexible for this reason so that we canon the one hand adapt
00:41:28with rememberbecause MOS law is largely overtransistor benefit is only tens 10%maybe a couple of years>> and yet we would like to have hundredsof X every year and so the benefit isactually all in algorithms and anarchitecture that enables any algorithmis likely going to be the best one rightbecause the transistor didn't it didn'tadvance that much and so I I think thethe our dedication to programmability isnumber one for that reason we have somuch optimism for innovation andalgorithms and iteration software thatwe protect our programmability for thatreason the second thing is is byprotecting this architectureour installed base is really large. Whena software engineer wants to optimizetheir algorithm, they want to make surethat it doesn't run on just one this onelittle cloud or this one little stack.They want it to run on as many mo on asmany computers as possible. So the thefact that we protect our architecturecompatibility then flash attention runseverywhere. So SSM run everywhere,diffusion runs everywhere, autoreggression runs everywhere. Justdepending it doesn't matter what youwant to do. CNN still run everywhere.LSTM still runs everywhere. And so thatthis this architecture that is
00:42:43architecturally compatible so that wehave a large installed base programmablefor the future is really important inthe way that we help to advance and as aresult all of this drives the cost down[clears throat] and and I'm super proudthat that um uh our latest innovationMVLink72we're the lowest cost token generationmachine in the world by enormous amountsand the reason for that is because arereally really hard>> and so you know people didn't expectthat um that forees it's probably easierto train but for inference is incrediblyhard to generate tokens on but as ascost drop usually you open up newapplications or new verticals thatbecome more and more accessible>> and we talked a little bit about codinglike cursor and cognition and othercompanies that are really benefitingfrom that in this last year do you haveany thoughts or predictions in terms ofwhat the next breakthrough industrieswill be or new applications or areasthat you're most excited about coming in26 in particular like Are there one ortwo things that you think will>> because of three things I because ofbecause of a couple two three things I Ithink I think several industries aregoing to are going to experience theirchat moment. Um I believe thatmulti-modality
00:43:57and um very long context is going toenable of course really really cool chatbots. Um but the basic architecture thatin combination with breakthroughs insynthetic data generation is going tohelp create the chat GPT moment fordigital biology.>> That moment is coming.>> And by digital biology, do youspecifically mean other aspects of likeprotein folding or protein binding orprotein diagnosis? I see proteins.>> I think we're good at proteinunderstanding. Mhm. Now multi-proinunderstanding is coming online and werecently created a model called LAprina. It's open. Um it's formulti-proin>> um understanding and and representrepresentation learning and generation.Uh so so I think that the proteinunderstanding is is advancing veryquickly. Now protein generation is goingto advance very quickly. Chat GPD momentproteins.>> Yeah. There are a lot of interestingcompanies working on molecule design inendtoend way like chai.>> Exactly. And then and then of coursechemical understanding and chemicalgeneration and then protein chemical>> confirmation understanding andgeneration. Is that right? And so thatcombination the chat GBT moment thegenerative AI moment all
00:45:12of that stuffis coming together for for um digitalbiology>> and to your to your point about like newindustries or you know the way I thinkabout it is like investing in the inputsfor this AI as well. All of these thingsaround biology and chemistry andmaterial science, they require realworld data generation andexperimentation, right? And that's newinfrastructure too.>> New infrastructure, uh, synthetic datais going to be really important becausethey just have such sparse, right? Sparsparity of data and they just don't haveas much as human language. And there thethe real breakthrough is going to bewhen we can train a a world foundationmodel, a foundation model for proteins,a foundation model for cells. I'm I'mvery excited about both of those things.Once we have a a foundation model, ourunderstanding capability, our generativecapability, that data flywheel is reallygoing to take off.>> The this this the second area that I'mexcited about, um, of course, reasoningmade huge breakthroughs in language, butbecause of reasoning, cars are going tobe able to perform better. So, insteadof just perception cars and planningcars, they're going to be reasoningcars. So, these cars are going to bethinking all the time. And when theycome up they come up to a circumstancethey they've never en encountered beforethey can break it down intocircumstances they have encountered itbefore and construct a reason
00:46:27reasoningsystem for how to navigate through it.And so the out of domain out of you knowout of distribution>> part of AI is going to very much be beaddressed by reasoning systems or and asa result we could do more things than wewere taught to do between uh generativeAI uh and um multimodal uh you knowvision language action models andreasoning systems. I think we're goingto see big breakthroughs in human robotsor multi-mbodiment robots. you know does>> what do you think what do you think is atime frame for that because if you lookat the self-driving analog and obviouslyself-driving technologies were based onvery different types of neural networksthan what we're using today in terms ofyou know there's been a big swap overthe last two three years>> in terms of how we do a lot there>> we started too soon>> self-driving cars really had four erasera was smart sensors>> connected into a car>> the mobile [clears throat] eye era>> the mobile eye era and even even thevery earliest days of ofYeah.>> Yeah. Even the earliest days of Whimo,>> the the um you're talk you're usingsmart sensors um a lot of humanengineered algorithms>> and education severe mapping as far>> extreme mapping>> mapping and then different systems forplanning
00:47:42and perception.>> Exactly. And so so you're essentiallycreating a car that is driving ondigital rails, right? It's no differentthan than the rails at Disneyland. Thereare digital rails. And so that's thefirst generation. the second generation.Um and during that generation you haveperception, world model and planning.>> Mh.>> And and the these modules um and eachone of these modules have the limits oftheir technology and and perception wasfirst imple was was first affected bydeep learning uh first and then and thenuh and then it propagated through thepipeline.>> And so that but that system was toobrittle>> and it only knows how to perform whatyou taught it. And now where we are areendtoend models and then and then wherewe're going to go next are end to endmodels. There you go. So that those arekind of the four eras in a lot of ways.If we would have started self-drivingcars probably three years ago,>> we would probably be exactly the sameplace.>> All our poor friends who were working inself-driving. Yeah.>> And and I don't I don't mind it. I'vebeen working on on it for 10 years.Nvidia's self-driving car stack, by theway, number one rated safety in theworld today.>> Number one, we just got we just got thatrating today uh last week. And numbertwo is Tesla. So,
00:48:57I'm very proud thattwo American companies are up on the>> Are you um So, from a roboticsperspective, you think because we'vealready built all these sorts oftechnologies in the modern era, roboticswon't have the same 10, 15 years. That'sright.>> I'm much more optimistic with roboticsbecause we we've kind of>> advanced foundational technology.>> Now, you know, people are thinking abouthuman robotics. Human robotics has a lotof challenges. I mean, there's all themegatronics challenges there. You know,like for example,>> it's not helpful if the robot weighs 300lb>> and what happens if it falls over andinteracting with kids and so on soforth. And so, so you got all kinds ofchallenges to deal with. I'm certainthat we're going to we're going to solvethose. But remember the fundamentaltechnology that goes into a human robotrobot can go into a pick and placerobot.>> Um it could be it could be um how do youthink about one thing I've been curiousabout for robotics in particular is if Ilook at who won or who who who'sperceived as winning in self-driving.>> It's largely incumbents, right? It'sWhimo, it's Tesla. You mentioned uh thesafety rating Nvidia's gotten. And soit's people who've been working on thisfor a long time. It took a lot ofcapital. It was really intensive to getthere. You have supply chain, you havehardware, you have all this extracomplexity. Do you think the same thingwill be true in robotics? Are thewinners basically going to be Tesla withOptimus and other people
00:50:12who have bothbeen in the industry for a while butalso have all those sort of incumbenteffects? Do you think there's room forstartups?>> They will be one of the leader one ofthe one of them and and andsurely a major one. Um but everythingthat moves will be robotic.>> Everything that moves will be robotic.And everything that moves is a verylarge space. It's not all human orrobot. And yet every AI will bemulti-mbodiment meaning you know justlike just like a human with our m ourmulti-mbodimentAI ourselves>> we could sit in a car>> and embody that>> we could pick up a tennis racket embodythat we could pick up a chopstick embodythat>> and so we could embody the>> people are general purpose right theycan do all these things>> exactly and so AIS are going to becomegeneral purpose so you have one arm pickand place maybe it's two arms pick andplace could be six arms pick and place,you know. So, so I think you're going tohave all kinds of different sizes andshapes. It could be a caterpillar. Itcould be, you know, it could be anexcavator. It could be all kinds ofstuff. And so AI will embody those justas a just as a a construction workerembodies an excavator embodies atractor. You know, they you know,>> could there be a small number ofcompanies then that do the embodimentfor everything
00:51:27or are you saying morethere's going to be niche applications?You should definitely see a lot ofsoftware companies and then those thatsoftwarecompany could serve a lot of a lot ofdifferent>> verticals but each one of the verticalswill still have solution providers thatthen grounds it all turns it intosomething that works perfectly. Does itmake sense? Because in the case of AIfor consumers if it works 90% of thetime you're delighted you you're youknow you're mind blown. If it works 80%of the time you're satisfied. In thecase of most industrial and physicalAIs, if it works 90% of the time, nobodycares about that. They only care aboutthe 10% that it fails. Basically, youknow, 100% dissatisfaction. And so, yougot to take it to 99.99999.So, the core technology might be able toget get you to 99%.>> And then a vertical solution providerlike a Caterpillar or somebody, theycould take that core technology and makeit 99.999%great. Do you think that's what happenslike earliest on because in in marketsthat are this immature it seems one ofthe fastest paths to market could befull verticalization right because youjust have control of iteration speed>> the different the the difficultydifficulty of of verticalization fortechnology that that is general purposeis that you don't have the R&D scale tobuild a general purpose technology.
00:52:43Now,of course, open source helps thattremendously,>> which is the reason why you're going tosee a, you know, a a big surge ofverticalopportunities in AI in the next severalyears.>> My my prediction would be over thecourse of the next five years, theexcitement is going to beverticalization.>> Notice we we're excited about OpenEvidence, we're excited about Harvey,we're excited about Cursor. cursor is isa horizontal but it's kind of ahorizontal vertical>> you know and so um I'm I'm super excitedabout all the verticals>> you know a lot of people said yeah AI isgonna get so god AI is going to get sogood that all these rapper companies aregoing to be obsolete it's just it missesthe big point>> you know the reason why you could talkabout the reason why somebody can talktalk about somebody is creatingtechnology could talk about the life ofa surgeon is because they've never beena surgeon the reason why somebody whobuilds at AI and talk talks about thelife of a accountant and a tax, youknow, a tax expert because they've neverbeen a tax expert, you know, and so so II think they just the reason whysomebody could talk about being a busboy without being a bus boy is theynever been a bus boy. And so so I Ithink you you you've got to be a littlebit more empathetic about the depth ofthe complexity of the work>> and and tr try
00:53:58to truly understand thepurpose of the work. Often times the thetechnology addresses the task, itdoesn't address the purpose. So I guessone of the other narratives from we'relooking at narratives that are trueversus not true, you know, for 25. Oneother narrative that's come up has beenmore about energy and energy utilizationand will we have enough energy tosupport AI. How do how do you thinkabout that? On the first week ofPresident Trump's administration, hesaid drill, baby drill. He got so muchflack for that.If not for this entire change in insentiment about energy growth in ourcountry,>> we can all concede now we would havehanded this industrial revolution tosomebody else.>> And we're still power constrained.>> We're still power constrained. Yeah.>> Without energy, there can be no newindustry.>> Mhm. And of course, we've been energystarved now for what, a decade. If notfor the fact that President Trumpreversed that narrative, we would becompletely screwed.>> Mhm.>> Without energy, you can't haveindustrial growth. Without industrialgrowth, the the nation can't be moreprosperous. Without being moreprosperous, we can't take care ofdomestic issues. We can't take care ofsocial issues. You know, on and on andon.
00:55:14And so, the fact of the matter is,we need energy to grow. We need everyform of energy. We need, you know,natural gas. We need to be, of course,we need more energy on the grid. We needmore energy behind the meter. Uh we'regoing to need nuclear. Uh wind is notgoing to be enough. Solar is not goingto be enough. Let's just all acknowledgethat we'll take it. We'll takeeverything we can. Um but the fact thatmatters, I think, for the for the nextdecade,>> natural gas, you know, is probably thethe only way to go forward. What'sreally interesting is I I agree thetimeline is too far out to addresspeople's um you know power generationissues in 27 and 28 where uh you knowlarge players building clusters are veryconcerned but the the biggest drivers oflike climate innovation in the US haveactually been as a result of this AIinfrastructure problem right becausepeople look at the demand>> finally that's right demand>> they look at the demand and the demandis driving people to create massive ofnew battery companies, solarconcentrators. It's put new energy benew energy like you know willpowerbehind>> SM the AI industry is driving all ofthat sustainable energy industry.>> Yeah.>> Um because people see that there isgoing to be demand for it right so evenif and I
00:56:29think there is no practicalanswer in the small number of years timeframe versus uh large gas right um uh itstill drives climate innovation. Yeah,no question about it. No question aboutit. And I I think that's exactly rightthat that you know doomer messages umcauses policy and that policy may mayaffect the industry in some way. Butthere's nothing more powerful thandemand. Look at all the jobs that'sbeing created. Look at all the theindustries that's being formed aroundit. um sustainable energy likely andwhen history rewrites it as Sarah, Ithink you you're going to be absolutelyright that that if not for AI, well AIwas is probably the biggest driver forsustainable energy ever.>> Yeah. A friend of mine has a saying thatuh doomers are the people who soundsmart at dinner parties and optimistsare the people who drive humanityforward. And I think that's very truefor for all these things we've talkedabout. Yeah. So>> yeah, it's really true.>> Yeah. Well, that that's one of the bigbig um takeaways for for uh this lastyear, the battle of narratives.>> And it's too simplisticum to say that everything that thedoomers are saying are irrelevant.That's not true. A lot of very sensiblethings are being said. Um it is toosimplistic to say that when somebody is
00:57:44optimistic that they're just naive.>> It needs to be grounded in reality.Yeah, that optimistic people are justnaive, you know,>> and that that's obviously not true.>> Um, but I think we just have to bemindful of the balance of it.>> When 90% of the messaging is all aroundthe end of the world and doom and thepessimism and you know, I think we we'rescaring people>> from making the investments in AI thatmakes it safer, more functional, moreproductive>> and more useful to society. And so wejust, you know, more secure. We, youknow, all of that takes technology.Security takes technology. Safety takestechnology. I appreciate that my car issafer today because it has bettertechnology than a car 50 years ago.>> And so so I I think it takes technologyto be safe, technology to be secure. Andso I I'm I'm I'm delighted to see thatthe the advancement of technology isstill accelerating and ongoing. And sowe just have to make sure that the thepolicy makers around the world, thegovernments um are able to are arethinking about balancing these twoideas.>> How do you So I guess we've talked a lotabout 25>> and the narratives of 25. How do youthink about 26? What are you excitedabout? What do you see coming? What doyou think are big changes that we shouldbe aware
00:58:59of?>> I am optimistic that that um ourrelationship with China will improve.Mhm. [clears throat]>> that President Trump and theadministration um has a really reallygrounded and common sense um attitudeabout um and philosophy around aroundhow to think about China that thatthey're an adversary>> um but they're also also a partner inmany ways and that the idea ofdecoupling is naive and the idea ofdecoupling um for whatever reasonphilosophical reasons or nationalsecurity reasons It's just not not it'snot based on any common sense and themore you the more deeply you look intoit the more the two countries areactually highly coupled.>> Um both countries ought to ought toinvest in their own independence. Um Iyou know when you depend too much onsomeone the relationship becomes tooemotional uh as you know [laughter] andso it's good to have some independenceor as much independence as either eitherwould like but to recognize that there'sa lot of coupling a lot of dependencebetween the two countries and and Ithink there's a there needs to be anuanced strategy a nuanced attitudeabout how to how to how to
01:00:14manage thisrelationship in a productive way for allof the people of two countries and forall of the people around the world,everybody depends on a productive,constructive relationship of the twomost important nations and the singlemost important relationship for the nextcentury. And so we have to find thatanswer. And I'm I'm I I'm just reallydelighted uh that President Trump islooking for a constructive answer. Andso I I think that next year uh will be amuch better better better year than thelast several. I'm happy with theadministration was able to to to suggesta a an export control um policy that isgrounded on national securityrecognizing that they already make somany chips themselves and they they candepend on Huawei themselves for theirmilitary for their national security.they got ample technology to do that.And so that American technology,although general purpose um is unlikelyto be used by their military becausetheir military is too smart, just as ourmilitary is too smart to to use theirtechnology. And so it's grounded onnational security. It's grounded on onuh technology leadership. It's groundedon national prosperity. You know, one ofthe things that that we just always haveto
01:01:29remember is that the world'smightiest military uh is supported bythe world's mightiest mil economy. Andso the wealth that we generate um bringsjobs home, creates prosperity in theUnited States, um provides for taxrevenues, and ultimately funds themightiest military on the planet. And sothat circular system, thatinterconnected system requires a nuancedstrategy. and and um uh and and and andI'm I'm I'm pleased to to to to see someof the progress in that area that allowsAmerican technology companies to keepAmerica first and keep America ahead>> and to to support American technologyleadership on the one hand um to winglobally>> and and then and then China of course issorting itself out you know I mean notsorting but they're sorting out theattitude about how to think aboutAmerican technology and there>> because historical argument there hasbeen that if if you look for example atthe internet um there was what was knownas a great firewall right Chinabasically>> prevented US competition into Chinawhile the opposite wasn't as true>> um there's been mass expatriation of USjobs and industry to China as sort ofpart of the development of the 90s and2000s and so I think a lot of the thingsthat people have brought up from a China
01:02:44US policy perspective besides just themilitary adversarial relationship um orspheres of influence or you know all thevarious things like that is also justthe economic imbalances that haveperceived to exist between the twocountries. The way that I would thinkthrough that is go back to the firstprinciples of technologies again>> and and let's say the internet you havethe chip industry you have the systemsindustry the software industry you havethe services industry on top rememberChina's internet growth has been a boonfor Intel and AMD selling CPUs>> Micron selling DRAM skinex and Samsungselling DRAM>> it is the second largest internetmarket for American technology industry>> and so so maybe maybe it wasn't helpfulto some layer of the stack>> the Googles of the world>> but don't exclude every layer of thestack always come back every single oneof these things take a step back andlook at the whole stack>> maybe that's a theme for today as welland it makes sense that you would youwould send this message but you knowtechnology is actually not just the thesort of internet software applicationlayer that's been very dominant for twodecades>> it's the whole stack and Remember as asas Intel and AMD prospered
01:04:00uh with the internet industry uh inChina growth the China industry growthdon't forget China also contributedtremendously to open source. No countryin the world contributes more to opensource than China. And look at all thestartups here in America that were ableto benefit from that open source tocreate the the new startups that arehere. And so you can't look at one areain isolation. You have to look at thewhole life cycle of the technology andlook at every layer of the stack. Doesit make sense? When you take a look atthat from that lens,>> China's internet industrygenerated enormous prosperity forAmerica.>> Mhm.>> Just not at the internet company per se.>> Jensen, my other investor friends willnot forgive me if I don't ask you about2026 um uh on the business side. Uh arewe in an AI bubble? AI bubble. Yeah,there's a lot of ways to reason throughthat.>> And so, so again, um, you know, whenwhen asked that question, my mind goesto what is AI and where are we in that?There's AI,then there's computing. You know, as youknow, Nvidia invented acceleratedcomputing. Accelerated computing does
01:05:15computer graphics and rendering. AIdoesn't. Um, accelerated computing doesdata processing, SQL data processing. AIdoesn't.>> Um, accelerated computing does moleculardynamics and quantum chemistry. AIdoesn't. You know, all these are allthings that people could say someday AIwill, but it doesn't today. Acceleratedcomputing is really essential for uhclassical machine learning, XG boost,recommener systems, the whole process ofuh feature engineering, extract, load,and transform. That entire data science,machine learning life cycle, acceleratedcomputing is used for all of that. Thefirst thing to go to is in the contextof Nvidia.What we see is the the the dynamic isthe shift from general purpose computingto accelerated computing because MOSlaws largely ended. You can't use CPUsfor everything anymore like you used to.And so it's just no longer productiveenough. It's not deflationary enough.>> And so so we have to move towards a newcomputing model. And that's whereaccelerator comes in. If you ifgenerative AI well excuse me if chatbotslet's just go you know open AI andAnthropic and Gemini if none of thatexisted today Nvidia would be amultiundred billion dollar company andthe reason for that is because
01:06:30as youknow the foundation of computing isshifting to accelerated computing>> that's the first thing to to realize isis to take a step back and ask yourselfwhat is actually happening now the nextlayer up the question about AI nowbecomes What is AI? Now, we ask that weask the AI bubble question and we alwaysgo back to OpenAI's revenues 100%. Don'twe?>> Mhm.>> You ask somebody, hey, is there an AIbubble? Everybody goes directly toOpenAI's revenues. First of all, ifOpenAI currently has twice the capacity,their revenues would double. You guysknow that if they have 10 times thecapacity, their I really believe theirrevenues would 10 times. And so, theyneed capacity. This is no different thanNvidia needs wafers from TSMC. Justbecause you know Nvidia exists and andwe're doing great doesn't mean we don'tneed capacity. We need capacity. We needcapacity of DRAM. We need and so in ourworld it's sensible to everybody. Weneed capacity. Well, in their world theyneed factories>> and if they don't have factory capacityhow they generate tokens, which is wherewe started our conversation today and sothey need factory capacity in order toincrease their revenue growth. Butnonetheless, we also said that AI ismore than chatbots. It includes allthese different fields of science. Um,Nvidia's AV
01:07:46business is coming up on 10billion dollars. Nobody ever talks aboutthat. And you have to train worldmodels. You have to train these AI AVsand it's happening robo taxis happeningall over the world. Our AI work with uhdigital biology, our AI work infinancial services. The whole industryof quants, quantitative trading ismoving towards Yeah, exactly. They usedto be classical machine learning. Awhole bunch of human featured they callquants, right? These these specializedmathematicians were trying to figure outwhat the predictive features are. Now weuse AI to figure it out. And so in orderto have instead of having quants, youneed a lot of supercomputers. Financialservices is one of our fastest growingsegments. billions of dollars in inquants, you know, in financial services,billions of dollars in AV, billions ofdollars in robotics coming up, billionsof dollars in digital biology. And sohow big can that all that be? Well,simple logic is this simple math.Whether you you think that AI is goingto replace shortage, labor shortage orworkforce shortage in any kind, um,let's ignore that for a second. Theworld is at hundred trillion dollars inGDP. out of that let's just say 2% 2%
01:09:01annually is R&D and let's just go backin time five years ago if you were totake the largest drug discovery companyin the world drug company in the worldand where's all of their R&D wet labs>> today what are they do doing buildingsupercomputers>> and so there's a fundamental shift inhow they think about that $2 trillion>> it used to be $2 trillion for the oldway of doing things. It's now going tobe $2 trillion in the AI way of doingthings. Well, $2 trillion is going toneed $2 trillion of R&D is going to bepowered by a whole bunch ofinfrastructure. And that's the reasonwhy we're building supercomputerseverywhere around the world. And so so Ithink if if you reason about it from theoutside in, you know, either from thefoundation up, from the outside in, youcome to the conclusion that what we'reexperiencing, what all three of us areexperiencing, which is the amount ofcomputing demand is insane.>> Give me an example of a startup companythat goes, "No, we're good.">> They are all dying for computingcapacity. Give me an example for aresearcher in any university, ascientist in any company who says gotplenty of capacity. Everybody is dyingfor capacity. And so we have a global
01:10:17multi- company, multi-industry shortage.It's not just about open AI even thoughopen AI could use a lot more capacity aswell. So I think I think how we thinkabout this what with the narrative thenarrative is not helpful and it's alittle bit too superficial to say how doyou prove there's an AI bubble$12billion of revenueshundreds of billions of dollarinfrastructure being built is a littlebit too simplistic.>> Yeah. The other thing people um tend topoint out is the MIT study. Therethere's some study that I think came outof MIT that claimed that most enterprisedeployments of AI weren't that useful.And you're like, well, did you do thechange management? Did you do a reorg?Did you integrate into tooling? Did youlike how long did it even take toimplement it? If a planning cycle in anenterprise is a year and is something insix months and so it feels like there'sa lot of these kind of again overstatedthings that get a lot of attention, butthen you map it against what's actuallyhappening.>> Yeah.>> And the growth of these companies usingAI and it's just a completely differentworld. And and and if you want to findout where the world's innovation'shappening, I would not go find out at anenterprise.>> Would you guys agree?>> Yeah.>> Enterprise is like the slowest adoptersof new technologies. I would go talk toall of the startups, the 30, 40,000startups that are currently doing thisstuff.
01:11:32I would go talk to Open Evans.How how's it working? I would go go talkto cursor. How's coding working by theway? You know, I would just go talk tothese people.>> I think it's really interesting that yousee that. Um, of course you do havecompanies making, you know, hundredmillion plus, multiund million plusprogress of AR in enterprise sales,Harvey, Sierra, etc. But some of thefastest growing companies have beenenduser adopted even in conservativeindustries, right? Like healthcare, youknow, skeptical industries likeengineering,>> healthare, the most right, the mostconservative of all. But guess what?They are so concerned about getting theright answer>> that the ability to have something likeopen evidence.>> Yeah.>> To do grounded research, high qualityresearch and get that get that researchas information to you. Nobody wants todo research. They want answers. Nobodywants to do search. They want answers.Is that right?>> A bridge is a great example of that toowhere they're basically making it reallyeasy to do the physician knows insteadof the physician sitting there and doingit. Back to your point on task versus>> task versus purpose. Exactly. And Ithink a different way to think about thedemand is like there are so many jobswhere you're asking the the work isactually like an impossible ask right ofa doctor or a radiologist keep up withthe world's biomedical knowledge in R&Dwhich is accelerating you know computingand
01:12:47otherwise um and then>> like archive papers>> there was a time you s you and I read>> you and I both both used to do I don'tdo that anymore but here now now I justload it all into chat>> GBTh you Now I just load it all in withall of the the ones that are interestingand and I make it learn it>> and then I you know make it summarizeand another summary and I I interactwith it. But but the point is uh we usedto do search. We don't do it thatanymore. I don't do search. We used todo research. You know the goal is to getanswers. The goal is to get smarter. Andthese AIs allow us to help us do allthat. And I think all of it all of itcomes back with it. It's all morehelpful if you come back to theframework that says AI is a multi-layercake>> and that AI is not just a chatbot. AI isvery very diverse in all of theindustries and modalities andinformation and applications that itaddresses. When you think about wantingto win>> that America should win AI, it shouldnot just be America should have thiscompany win AI, but it we should try towin across the board>> and across domains.>> Across
01:14:02domains. Exactly. And when wethink about open source, all of a suddenthis this is a helpful framework. Whenwe think about winning, it's a helpfulframework. When we think about uh energyis a helpful framework that because weneed factories. Factories need energy.And without energy, we have no factory.without factories we have no AI that's ahelpful framework and so I think if ifum if we if we have a betterunderstanding a system a framework forunderstanding what AI is I think thenarratives will become more common sensethe narratives will become morepragmatic>> become more balanced we want to keeppeople safe>> but one of the best ways to keep peoplesafe is advancing advancing oftechnology quickly>> and and I think the industry is doingthat and I'm very proud of the industryfor doing that.>> No one wants to drive a car from, youknow, the first decade of cars. And so II think uh>> ABS is a really good thing.>> Yes,>> ABS is a really good thing. Lane keepingis a really good thing. There's noquestion FSD is a really good thing.>> And I think people will be excited aboutthe, you know, third or fourth year ofAI.>> Yeah. No, no doubt. And and I I say withgreat pride that the industry madetremendous strides this last year.all the technologies
01:15:17we've mentioned.Um, and that the scaling laws are sointact that we we now know that morecompute, more intelligence>> and and um uh gosh, the the the theinnovations in one in in one sectordiffuses and spreads across all of theother sectors so fast. I'm so happy tosee all that. And so I think the nextfive years it's going to beextraordinary. No, no doubt about it.And I think next year is going to beincredible.>> Amazing. Well, we're excited to talk toyou at the end of next year, too.>> Yeah. Looking forward to it. Thank youguys for all the work that you guys do.Congratulations. What a great year.>> Wow. Amazing year.>> Yeah. A lot. Thank you.>> Yeah. Thank you. Happy New Year. HappyNew Year.
01:16:02>> Find us on Twitter at No Prior Pod.Subscribe to our YouTube channel if youwant to see our faces. Follow the showon Apple Podcasts, Spotify, or wherever[music] you listen. That way you get anew episode every week. And sign up foremails or find transcripts for everyepisode at no-briers.com.