NVIDIA CEO Jensen Huang Leaves Everyone SPEECHLESS (Supercut)

our TCO is so good that even when the competitor's chips are free it's not cheap enough in the last 10 years we reduced the cost of computing by 1 million times the cost of deep learning by 1 million times if we could reduce the marginal cost of computing down to approximately zero we might use it to do something insanely amazing large language models to literally extract all of digital human knowledge from the internet and put it into a computer and let it go figure out what the WIS what the knowledge is that idea of scraping the entire internet and putting it in one computer and let the computer figure out what the program is is an insane concept but you wouldn't ever consider doing it unless the marginal cost of computing was Zero we made that breakthrough and now we've enabled this new way of doing software imagine you know for for all the people that are still new to artificial intelligence we figured figured out how to use a computer to understand the meaning not the pattern but the meaning of almost all digital knowledge and anything you can digitize we can understand the meaning so let me give you an example Gene sequencing is digitizing genes but now with large language models we can go learn the meaning of that Gene we can understand what's the meaning of a cell a whole bunch of genes that are connected together and this is from a computer's pers perspective no different than a whole page of words and you asked it to summarize what did it say summarize it for me what's the meaning this is no different than a long page of genes what's the meaning of that big long page of proteins what's the meaning of that that AI which was enabled by this new form of computing we call Accelerated Computing that took three decades to do uh is probably the single greatest invention of the technology industry this will likely be the most important thing of the 21st century the GPU chip that is behind uh artificial intelligence right now is your h100 and I know you're introducing an h200 and I think I read that you plan to upgrade that each year and so could you think ahead five years March 2029 you're introducing the H700 right what will it allow us to do that we can't do now let me first say something about the chip that John just described um as we say a chip all of you in the audience probably you imagine there's a chip kind of like you know like this um the chip that John just described uh weighs 70 lbs it consists of 35,000 Parts eight of those parts came from tsmc that one chip replaces a data center of old CPUs like this into to one computer the savings of that one computer is incredible and yet it's the most expensive computer the world's ever seen it's it's a quarter of a million dollars per chip we sell the world's first quarter million dollar chip but the system that it replaced the cables alone cost more than the chip this h100 the cables of connecting all those old computers that's the that's the incredible thing that we did we took this entire data center We Shrunk get into this one chip the way that this chip works it works not just at the chip level but it works at the chip level and the algorithm level and the data center level it doesn't do all of its work by itself it works as a team and so you connect a whole bunch of these things together and networking is part of it and it computes at data center scales and together what's going to happen in the next 10 years say John um will increase the computational capability for deep learning by another million times what happens when you do that um today we we kind of learn and then we apply it we go train inference we learn and we apply it in the future we'll have continuous learning the the training process and the inference process the training process and the deployment process application process will just become one well that's exactly what we do you know we don't have like between now and 7:00 in the morning I'm going to be doing my learning and then after that I'll just be doing inference you're learning and inferencing all the time and that reinforcement learning Loop will be continuous and that reinforcement learning will be grounded with real world data through interaction as well as synthetically generated data that we're creating in real time so this computer will be imagining all the time it'll do synthetic data generation it'll do reinforcement learning it'll continue to be grounded by real world experiences um it'll imagine some things it'll test it with real world experience it'll be grounded by that and that entire Loop is just one giant Loop that's what happens when you can compute for a million times cheaper than today and so as I as I'm saying this notice what's what's at the core of it when you can drive the marginal cost of computing down to zero then there are many new ways of doing something you're willing to do this is no different than I'm willing to go further places because the marginal cost of Transportation has gone to zero I can from here to New York relatively cheaply if it would have taken a month you know probably never go and so it's exactly the same in transportation and just about everything that we do and so we we're going to take the marginal cost of computing down to approximately zero as a result we'll do a lot more computation as you probably know there have been some recent stories that Nvidia will face more competition in the inference Market than it has in the training Market Market but what you're saying is it's actually going to be one market I think is there going to be a separate training chip market and inference chip Market or it sounds like you're going to be continuously uh training and switching to inference maybe within one chip I I don't I don't know why don't you explain some well today today whenever you uh prompt uh an AI it could be chat GPT or it could be co-pilot whenever you prompt it's doing inference so it's it's generating information for you whenever you do that what's behind it 100% of them is idious gpus and so nvidias most of the time you engage our our our platforms are when you're inferencing and so we are 100% of the world's inferencing today is NVIDIA now is inferencing hard or easy the the reason why people are picking on inference is when you look at training and you look at Nvidia system doing training when you just look at it you go that looks too hard I'm not going to go do that I'm a chip company that doesn't look like a chip a lot of competitors tend to say we're not into training we're into inference inference is incredibly hard the hard part of inference is the goal of somebody who's doing inference is to engage a lot more users to to apply the software to a large install base inference is an install based problem this is no different than somebody who's writing a an application on on a on an iPhone um the reason why they do so is because iPhone has such an large install base almost everyone has one and so if you wrote an application for that phone it's going to be able to benefit everybody well in the case of Nvidia our accelerated Computing platform is the only accelerated Computing platform that's literally everywhere and because we we've been working on it for so long if you wrote an application for inference and you take that model and you Deploy on Nvidia architecture it literally runs everywhere and that takes enormous patience and years and years of success and dedication to architecture compatibility you know so on so forth you make completely state-of-the-art chips is it possible though that you'll face competition that is claims to be good enough not as good as Nvidia but good enough and and much cheaper is that a threat we have more competition than anyone on the planet has competition uh not only do we have competition from competitors we have competition from our customers and I'm the only competitor to a customer um fully knowing they're about to design a chip to replace ours and I show them not only what my current chip is I show them what my next chip is and I'll show them what my chip after that is if a customer can can um do something specifically that's more cost effective uh quite frankly I'm even surprised by that and the reason for that is this when you see it when you see computers these days it's not a computer like like a laptop it's a computer it's a Data Center and you have to operate it and so people who buy and sell chips think about the price of chips people who operate data centers think about the cost of operations our time to deployment our performance our utilization our flexibility across all these different applications in total allows our operations cost they call total cost of operations TCO our TCO is so good that even when the competitor's chips are free it's not cheap enough and that that is our goal to add so much value that the alternative um is not about cost I was hoping not to sound competitive but John asked a question that kind of triggered a competitive Gene and I came AC I I want to say I want to apologize I came across you know if you will a little competitive he surprised me with a competive I I I I thought I was on an economic Forum could you just dive a little deeper into what you see as ai's role in drug Discovery the first role is to understand understand the meaning of the digital information that we have the question is now what is the meaning of that protein what is the meaning of this protein what is this function as you guys know uh there's you you can chat with a PDF my favorites are you take a PDF file of a of a a research paper and you load it into chat GP it's like talking to the researchers what what inspired this this research what problem does it solve you know what was the Breakthrough what was the what was a state of art before then in the future we going to take a protein put it into chat GPT just like PDF what are you for what what enzymes activate you you know what makes you happy and so so that that's that's one of the most profound things we can do is to understand the meaning of biology does that make sense if we can understand the meaning of biology as you guys know once we understand the meaning of almost any information that it's in the world of computer science in the world of computing amazing engineers and amazing scientists know exactly what to do with it but that's the Breakthrough boy Oregon State and Stanford are really proud of you Stanford has a lot of aspiring entrepreneurs students that are entrepreneurs and maybe they're computer science Majors or or engineering majors of some sort please don't build gpus what what advice would you give them uh to improve their chances of success I think one of my my great advantages is that I have very low expectations most of the Stanford graduates have very high expectations you you and you deserve to have have high expectations because you came from a great school um uh you were very successful you're surrounded by other kids that are just incredible you you naturally have very high expectations um people with very high expectations have very low resilience and unfortunately resilience matters in success I don't know how to teach it to you except for I hope suffering happens to you and and to to this day I use the word the phrase pain and suffering inside our company with great Glee and the reason and I mean that you know boy this is going to cause a lot of pain and suffering and I mean that in a happy way um because you want to refine the character of your company you want greatness out of them and greatness is not intelligence as you know greatness comes from character and character isn't isn't informed out of smart people it's formed out of people who suffered you know for all of you Stanford students I I wish upon you you know ample doses of pain and suffering I have a couple question what's a story about your leather jacket and the second the second is according to your projection and calculation in 5 to 10 years how much more semiconductor manufacturing capacity is needed to support the growth of AI okay uh I appreciate two questions the first question is this is what my wife bought for me and this is what I'm [Laughter] wearing I do 0% of my own shopping uh as soon as she finds something that doesn't make me itch because she knows she's known me since I was 17 years old and the way I say I don't like something is it makes me itch and so as soon as she finds me something that doesn't make me itch if you look at my closet the whole closet is a shirt because she doesn't want to shot for me again and so so that's why uh this is all she bought me and this is all I'm wearing and if I if I don't like the answer I can go shopping otherwise I can wear it and it's good enough for me okay we second question on this the forecast is actually very this is very I'm horrible at forecasting but I'm very good at first principled reasoning of the size of the opportunity and so let me first reason for you um uh I have no idea how many Fabs but here's here's the thing that I do know the way that we do Computing today all the words all the videos all the sound everything that we do is retrieval based it was pre-recorded does that make sense as I say that every time you touch on a phone remember somebody wrote that and stored it somewhere it was pre-recorded okay every modality that you know in the future most of computing will be generative in the today 100% of content is pre-recorded if in the future 100% of content will be generative the question is how many how does that change the shape of computing how much more networking do we need more less of that do we need memory of this and and the answer is we're going to need more Fabs however it's not as if the efficent of computing is what it is today and therefore the demand is this much in the meantime I'm improving Computing by a million times every 10 years while demand is going up by a trillion times and that has to offset each other does that make sense that's just a matter of time but it doesn't change the fact that one day all of the computers in the world will be changed 100% every single data center will be all of those general purpose Computing data centers 100% of the trillion dollars worth of infrastructure will be completely changed and then there'll be new infrastructure built on even on top of that okay next question right here Ben remember the last question has all big pressure you guys agree with that do you can we all agree right here the the person who La asked the last question don't don't leave us all depressed I'm going to invoke your commandment to have low expectations at this juncture um you you mentioned you're competing with your customers and I'm wondering you know given the advantages that you have why they're doing that and I'm wondering if in the future you see yourself building more customized solutions for customers of a certain scale um as opposed to the solutions that you have now which are more horizontal are we willing to customize the aners yes now why is it that the bar is relatively High because each generation of our our platform first of all there's a GPU there's a CPU there's a networking processor there are two types of switches I just built five chips for one generation people thinks is one chip but it's five different chips each one of those chips are hundreds and hundreds of millions of dollars to do then you've got to put them into a system you got networking stuff you got C transceiver stuff you got optic stuff you got a mountain of software to do it takes a lot of software to run a computer as big as this room and so if the customization is so different then then you have to repeat the entire R&D however if the customization Leverage es everything and adds something to it then it makes it's makes a great deal of sense maybe it's a it's a proprietary security system maybe it's a confidential Computing system maybe it's a a a new way of doing a numerical processing um that that could be extended uh we're very open-minded to that our customers know that I'm willing to do all that and recognizes the the the if you change it too far you've basically reset and and you've squandered you know the the nearly hundred billion dollar that's taken us to get here um uh to to redo it from from scratch and so they want to leverage our ecosystem to the extent that that that that will be done I'm very open to it yeah and they know and they know that yeah okay so with that I think we need to wrap up thank you so much to John and Jensen

Channel: Ticker Symbol: YOU

Views: 652,734

Keywords: jensen huang stanford, nvidia stanford, nvda, nvda stock, nvidia stock, top stocks, best stocks, ai stocks, chatgpt, openai, growth stocks, tech stocks, pltr, pltr stock, palantir stock, amd stock, smci, smci stock, sora, nvidia gtc, jensen huang, jensen huang keynote, best ai stocks, best stocks to buy now, top ai stocks, nvidia keynote

Id: oNwoA5akBlg

Length: 18min 24sec (1104 seconds)

Published: Tue Apr 30 2024