Nvidia CEO Jensen Huang talks blowout quarter, AI, inferencing, ongoing demand, and more

I'm Julie Heyman host of Yahoo finances Market domination here with our Tech editor Dan Howley Nvidia has done it again the chip giant blowing past analyst expectations in its strong fiscal first quarter data center Revenue alone soaring by 4 127% year-over-year and the company also gave another bullish sales forecast which shows that AI spending momentum continues a pace on top of all that the company also announced a 10 for one forward stock split and ra its dividend joining us now Nvidia founder and CEO Jensen Wong fresh off the conference call Jensen welcome thank you so much for being with us I'm happy to be here nice to see you guys you too I want to start uh with Blackwell which is your next Generation chip it's shipping this year you said on the call you also said on the call we will see a lot of Blackwell Revenue this year so if we're looking at about $28 billion in Revenue in the current quarter and Blackwell is a more expensive product than Hopper the chip series out now what does that imply about Revenue in the fourth quarter and for the full year well it should be significant yeah Blackwell Blackwell and and as you know we guide one quarter at a time and but what I what I could tell you about about Blackwell is this this is this is um a giant leap in in um uh in Ai and it was designed for trillion parameter AI models and this is as you know we're already at two trillion parameters uh models sizes are growing about doubling every six months and the amount of processing uh between the size of the model the amount of data is growing four times and so the ability for uh these data centers to keep up with these large models really depends on the technology that we bring bring to them and so the Blackwell is is designed uh also for incredibly fast inferencing and inference used to be about recognition of things but now inferencing as you know is about generation of information generative Ai and so whenever you're talking to chat GPT and it's generating information for you or drawing a picture for you or recognizing something and then drawing something for you that generation is a brand new uh inferencing technology is really really complicated and requires a lot of performance and so Blackwell is designed for large models for generative a I and we designed it to fit into any data center and so it's air cooled liquid cooled x86 or this new revolutionary processor we designed called Grace Grace blackwall super chip and then um uh you know supports uh infinite band data centers like we used to but we also now support a brand new type of data center ethernet we're going to bring AI to ethernet data centers so the number of ways that you could deploy Blackwell is way way higher than than Hopper generation so I'm excited about that I I I want to talk about the the inferencing Jensen you know some analysts have brought up the idea that as we move over towards inferencing from the the training that there may be some inhouse companies uh uh processors from companies that those made from Microsoft Google Amazon maybe more suited for the actual inferencing I guess how does that impact Nvidia then well inferencing used to be easy you know when people started talking about inference uh generative AI didn't exist and now generative AI is is uh uh of course is about prediction but it's about prediction of the next token or prediction of the next pixel or prediction of the next frame and all of that is complicated and and generative AI is also used for um understanding the cont in order to generate the content properly you have to understand the context and what what is called memory and so now the memory size is incredibly large and you have to have uh context memory you have to be able to generate the next token really really fast it takes a whole lot of tokens to make an image takes a ton of tokens to make a video and takes a lot of tokens to be able to uh reason about a particular task so that it can make a plan and so gener the the the gener generative AI um era really made inference a million times more complicated and as you know the number of chips that were intended for inference uh kind of kind of fell by the wayside and now people are talking talk about building new Chips you know the versatility of invidious architecture makes it possible for people to continue to innovate and create these amazing new Ai and then now black Wall's coming so in other words you think you still have a competitive Advantage even as the market sort of shifts to inferencing we have a great position in inference because inference is just a really complicated problem you know and the software stack is complicated the type of models that people use is complicated there's so many different types it's just going to be a giant market market opportunity for us the vast majority of the world's inferencing today as as people are experiencing in their data centers and on the web vast majority of the inferencing today is done on Nvidia and so we we I expect that to continue um you said on the call a couple of times that you'll be Supply constrained for both Hopper and then Blackwell uh chips well until next year because of the vast demand that's out there um what can you do about that are there any sort of levers you can pull to help increase Supply copper demand grew throughout this quarter after we announced Blackwell and so that kind of tells you how much demand there is out there people want to deploy these data centers right now they want to put our gpus to work right now and start making money and start saving money and so so that that demand is just so strong um you know it's really important to take a step back and realize that that what we build is not a GPU chip we call it Blackwell and we call it GPU but we're really building AI factories these AI factories have CPUs and gpus and really complicated memory the systems are really complicated it's connected by MV link there's an MV link switch there's infiniband switches infiniband Nicks and then now we have ethernet switches and ethernet Nicks and all of this connected together with this incredibly complicated spine called mvy link and then the amount of software that it takes to build all this and run all this is incredible and so these AI factories are essentially what we build we build it as a as a holistic unit as a holistic architecture and platform but then we disaggregate it so that our partners could take it and put it into Data Centers of any kind and every single cloud has slightly different architectures and different stacks and our our stacks and our architecture can now deeply integrated into theirs but everybody's a little different so we build it as an AI Factory we then disaggregated so that everybody can have ai factories this is just an incredible thing and we do this at very hard very high volume it's just very very hard to do and so every every component every every part of our data center uh is the most complex computer the world's ever made and so it's sensible that almost everything is constrained Jess I want to ask about the uh Cloud providers versus the the other industries that you said are are getting into the the gener AI game or or getting Nvidia chips you had mentioned that uh in uh comments in the actual release that we heard from uh CFO CL Crest uh that 40% mid 40% of data center Revenue comes from those Cloud providers as we start to see these other Industries open up what does what does that mean for NVIDIA well will the cloud providers kind of uh shrink I guess their share and then will these other Industries pick up where those Cloud providers were I expect I expect them both to grow uh a couple of different areas of course uh the consumer internet service providers this last quarter of course a big stories from meta the uh the incredible scale that that um Mark is investing in uh llama 2 was a breakthrough llama 3 was even more amazing they're creating models that that are that are activating uh large language model and generative AI work all over the world and so so the work that meta is doing is really really important uh you also saw uh uh Elon talking about uh the incredible infrastructure that he's building and and um one of the things that's that's really revolutionary about about the the version 12 of of uh Tesla's uh full self-driving is that it's an endtoend generative model and it learns from watching video surround video and it it learns about how to drive uh end to end and Jed using generative AI uh uh predict the next the path and and the how distur the uh how to understand and how to steer the car and so the the technology is really revolutionary and the work that they're doing is incredible so I gave you two examples a a startup company that we work with called recursion has built a supercomputer for generating molecules understanding proteins and generating molecule molecules for drug Discovery uh the list goes on I mean we can go on all afternoon and and just so many different areas of people who are who are now recognizing that we now have a software and AI model that can understand and be learn learn almost any language the language of English of course but the language of images and video and chemicals and protein and even physics and to be able to generate almost anything and so it's basically like machine translation and uh that capability is now being deployed at scale in so many different Industries Jensen just one more quick last question I'm glad you talked about um the auto business and and what you're seeing there you mentioned that Automotive is now the largest vertical Enterprise vertical Within data center you talked about the Tesla business but what is that all about is it is it self-driving among other automakers too are there other functions that automakers are using um within data center help us understand that a little bit better well Tesla is far ahead in self-driving cars um but every single car someday will have to have autonomous capability it's it's safer it's more convenient it's more more fun to drive and in order to do that uh it is now very well known very well understood that learning from video directly is the most effective way to train these models we used to train based on images that are labeled we would say this is a this is a car you know this is a car this is a sign this is a road and we would label that manually it's incredible and now we just put video right into the car and let the car figure it out by itself and and this technology is very similar to the technology of large language models but it requires just an enormous training facility and the reason for that is because there's videos the data rate of video the amount of data of video is so so high well the the same approach that's used for learning physics the physical world um from videos that is used for self-driving cars is essentially the same um AI technology used for grounding large language models to understand the world of physics uh so technologies that are uh like Sora which is just incredible um uh and other Technologies vo from from uh uh Google incredible the ability to generate video that makes sense that are conditioned by human prompt that needs to learn from video and so the next generation of AIS need to be grounded in physical AI needs to be needs to understand the physical world and the the best way to teach these AIS how the physical world behaves is through video just watching tons and tons and tons of videos and so the the combination of this multimodality training capability is going to really require a lot of computing demand in the years to come Jensen as always super cool stuff and great to be able to talk to you Dan and I really appreciate it Jensen Wong everybody founder and CEO of Nvidia great to see you guys thank you

Channel: Yahoo Finance

Views: 134,633

Keywords: Yahoo Finance, Personal Finance, Money, Investing, Business, Savings, Investment, Stocks, Bonds, FX, Currencies, NYSE, Equities, News, Politics, Market, Markets, Yahoo FInance Premium, Stock market

Id: ER7iqeYx9HU

Length: 12min 21sec (741 seconds)

Published: Thu May 23 2024