Nvidia CEO: Customers are buying now, not waiting for next chip
Nvidia reported first quarter results that topped Wall Street expectations on both the top and bottom lines. The tech giant also announced a 10-for-1 stock split and is raising its dividend.
For Q1 of fiscal 2025, Nvidia (NVDA) reported revenue rose 262% to $26.0 billion, with its Data Center being the biggest contributor. Revenue for that unit soared 427% year-over-year to $22.6 billion.
There have been concerns that with the company’s new Blackwell platform coming later this year, customers may be holding off on purchasing Nvidia’s Hopper products. In an exclusive interview, Nvidia founder and CEO Jensen Huang said that’s not the case. “Hopper 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 that demand is just so strong,” Huang says.
One point Huang made is just how big inference is. Nvidia describes AI inference as “the process of using a trained model to make predictions on never-seen-before data.” Huang tells Yahoo Finance’s Julie Hyman and Dan Howley Nvidia is in a “great position” regarding inference because of how complicated the problem is. Inference is going to be “a giant market opportunity for us,” Huang adds.
Watch the video above to hear how Huang says automakers like Tesla (TSLA) are using his company’s products to power the future of autonomous vehicles.
This post was written by Stephanie Mikulich.
For more Yahoo Finance coverage of Nvidia:
Nvidia stock pops 4% after earnings beat forecasts, announces stock split and dividend hike
Nvidia CEO Jensen Huang is the ‘man of the year’: Investor
Why this analyst says Nvidia is not a stock to buy
How Nvidia earnings are impacting the chip market
Beyond the Ticker: Nvidia
Video Transcript
I’m Julie Hyman, host of Yahoo Finance’s market domination here with our tech editor Dan Howley NVIDIA has done it again.
The chip giant blowing past analysts expectations in its strong fiscal first quarter data center revenue alone soaring by 427% year over year.
And the company also gave another bullish sales forecast which shows that A I spending momentum continues apace on top of all that, the company also announced a 10 for one forward stock split and raised its dividend.
Joining us now NVIDIA founder and Ceo Jensen Wang, 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 wanna 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 uh but what I, what I could tell you about, about Blackwell is this, this is, this is uh uh a giant leap in, in um uh in A I and it was designed for trillion parameter A I 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 uh 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 uh designed uh also for incredibly fast infer and inference used to be about recognition of things.
But now infer as you know, is about generation of information, generative A I.
And so whenever you’re talking to Chad GP T and is 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 Infer technology is really, really complicated and requires a lot of performance.
And so Blackwall is designed for large models for generative A I and we designed it to fit into any data center.
And so it’s air cool liquor cool X 86 or this new revolutionary processor we design called Grace Grace Blackwell 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 A I to Ethernet data centers.
So the number of ways that you could deploy Blackwell is way way higher than the than hopper generation.
So I’m excited about that.
II I wanna talk about the, the Infer Jensen, you know, some analysts have brought up the idea that as we move over towards infer from the, the training that there may be some in house companies uh uh processors from companies that those made from Microsoft Google Amazon may be more suited for the actual Infer I guess.
How does that impact NVIDIA then?
Well, infer it used to be easy, you know, when people started talking about inference, uh generative A I didn’t exist and now generative A I 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 A I is also used for um understanding the con 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 a 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 reason about a particular task so that it can make a plan.
And so the the the the genera generative A I 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 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 A is and then now black wall is coming.
So in other words, do you think you still have a competitive advantage even as the market sort of shifts to infer, we have a great position in inference because inference is just a really complicated problem, you know, and the software stock is complicated.
The type of models that people use is complicated.
There’s so many different types.
It’s just gonna be a giant market market opportunity for us.
The vast majority of the world’s infer today as as people are experiencing in their data centers and on the web, a vast majority of the infer 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 GP US 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, it’s really important to take a step back and realize that what we build is not a GP U chip, we call it Blackwell and we call it GP U.
But we’re really building A I factories, these A I factories have CP US and GP us and really complicated memory.
The systems are really complicated, it’s connected by MV link, there’s an MV link switch, there’s in Finan switches in Finan ni.
And then now we have Ethernet switches and Ethernet N and all of this connected together with this incredibly complicated spine called MV link.
And then the amount of software that it takes to build all this and run all this is incredible.
And so these A I 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 stags and our architecture can now deeply integrate into theirs.
But everybody is a little different.
So we build it as an A I factory, we then disaggregate it so that everybody can have A I 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 is the most complex computer the world’s ever made.
And so it’s sensible that almost everything is constrained.
J I wanna ask about the uh cloud providers versus the the other industries that you said are are getting into the, the JA I game or, or getting NVIDIA chips.
You, you had mentioned that uh in uh comments in the actual release.
And then we heard from uh CFO collect cress 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?
Will, 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 uh Mark is investing in uh Llama two was a breakthrough.
Llama three was even more amazing.
Uh They’re creating models that, that are, that are activating uh large language model and generative A I 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 Tesla’s uh full self driving is that it’s an end to end generative model and it learns from watching videos surround video and it, it learns about how to drive uh end to end and generate using generative A I uh uh predict the next, the path and the, and the uh how to steer the uh how to understand and how to steer the car.
And so the, the tech technology is really revolutionary and the work that they’re doing is incredible.
So I gave you two examples.
Uh a start up company that we work with called recursion has built up a super computer 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 A I model that can understand and be learned, 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.
Uh 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 A I 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 A is need to be grounded in physical A I need to be under needs to understand the physical world.
And the on the best way to teach these A is how the physical world behaves is through video, just watching tons and tons and tons of video.
And so the the combination of this multimodality training capability is going to really require a lot of uh 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.
