NVIDIA GTC 2025 | Sarbjeet Johal, Stackpane

Clips
More from NVIDIA GTC 2025

Sarbjeet Johal

Founder Stackpane, Analyst CUBE Collective

Stackpane

play_circle_outline Embracing Open Source and AI Infrastructure: NVIDIA's Developer-Focused Partnerships at GTC Super Bowl

Info
Transcript

Sarbjeet Johal, Stackpane

Sarbjeet Johal

Founder Stackpane, Analyst CUBE Collective Stackpane

Nvidia GTC 2025: An In-Depth Analysis of Emerging AI and Infrastructure Trends

Sarbjeet Johal of theCUBE provides an analysis of the pivotal developments from Nvidia's GTC 2025 event. This video captures significant insights and key highlights from Jensen Huang's keynote address, as well as broader implications for AI and next-generation infrastructure technologies.

In this episode, Johal, a seasoned analyst for theCUBE, brings expertise to the forefront by exploring Nvidia's strategic direction at GTC 2025. Joined by hosts from theCUBE Research... Read more

explore Keep Exploring

What were some of the key takeaways from the recent keynote presentation? add

bolt Powered by CUBE AI

Sarbjeet Johal, Stackpane

search

>> Hello, welcome back to theCUBE. We are here at GTC for coverage of NVIDIA's annual conference. Jensen Huang gave the keynote, always a bellwether, always a trendsetter, it's like I'm back to school learning, taking notes because it's all the stuff we talk about. Everything that we love on theCUBE and SiliconANGLE, he talks about. So it really is the Super Bowl and I think we coined that term, I know you used it. I think I said on theCUBE pod a few weeks ago. He actually said the Super Bowl is the event, I think we said that.

>> Yes.

>> Sarbjeet Johal's here, theCUBE Contributor Analyst. Sarbjeet, this is amazing. He said that, you tweeted that last night I think.

>> Yeah, I said -

>> I said it on theCUBE pod.

>> Yes.

>> I don't think anyone else has said this is the Super Bowl of tech but us. So Jensen shouts us out.

>> That's good.

>> That's good for us.

>> Yeah.

>> We love, this is our show.

>> This is it, your face glows up when you see all this infrastructure applications, developers, it's a mashup of all the technology. We heard the TELECO announcements around radios and whatnot. So yeah, let's dig into this.

>> This is an ultimate show, it's got infrastructure, AI infrastructure, which we predicted is going be surging onto the scene. And then we then said in one of our last interviews that it was going to open up into massive tsunami of applications with data, the agents are coming. And robotics is a great north star because that's also an application of data, right? So we got all these things coming together and then he weaves in enterprise compute. So I have to ask you, what is your hot take from this keynote? What's your assessment? What's your takeaway?

>> I think the major takeaway is like when things get steady, that's a good state to be in. I think the NVIDIA's mode is still good and they are getting into a steady state, so all these new announcements are incremental. I believe they are Nex-Gen infrastructure, they are going to release that. And the two main things actually, I tweeted earlier and I will give my longer take later in the evening we'll keep it short here, one is that they seem to be taking the open source route for making them even stronger with developers. So if they open source some stuff, it works only on their hardware, it gives them more power, right? That's one thing. And the second is that they seem to be sort of embracing more partnership and alliances based product development because there are quite a few announcements from today's keynote, which were with many storage vendors, including Dell, where we're sitting in their area right now, and IBM, and NetApp. So, all the storage vendors are getting together and they're working with NVIDIA to sort of address the last frontier of AI, if you will. So network got stronger, of course compute first, then network, then storage is also getting sort of face up, if you will.

>> Storage is a great example, they spend a lot of time on storage and they talked about pre-filled decode reasoning versus chat response, so performance is critical. Token speed, token for factory. The thing that I like the most, this is kind of really getting in the nerd weeds here, but the Dynamo, he says it's the operating system for the AI factory. I have been screaming at the top of my lungs for the past year, "What operating system runs on these factories?" So this is going to be a major battleground, Sarbjeet, they have their own AI factory called Dynamo operating system.

>> They open source it, they open sourced Dynamo.

>> They open sourced it. And so if that gets traction, that can be like the Linux of AI factories. So we need an operating system to run these factories. And also we talked about photonics because now he said data centers are like stadiums, I'll use copper for short-

>> He talked about serti and copper, and copper can go to one to two meters, but after that you need something better.

>> He spent a lot of time, and he had to burn minutes later by blowing through the slides for the partners, so he kind of short ended the ecosystem a bit. But you see where he spent his time, scale up, scale out, okay? And then the packaging, he went and spent a lot of time on how they're engineering the system with the different processors and the GPUs. So, he's getting into the weeds saying this is the innovation, the bigger systems, of course they want to do that because they make more money, buy more, save more, make more money.

>> Yeah, he said that, buy more, save more. Yeah, that's his mantra.

>> And he revised it, buy more, make more money.

>> Make more money, yeah.

>> Well that's what he said.

>> Build more actually he said.

>> Build more.

>> Yeah.

>> All right, anything else you saw there that you liked?

>> Actually the one on Sunday evening, I came here and shot some video and gave my anticipation of what they will talk about and they totally talked about the same things that I anticipated. They talked about the libs, they talk about the SDKs. They focus on the developers of the AI applications as well as the developers of the new models, so they have to cater to both parties, inference as well as the model trainers from scratch, right? So there were some concerns from the market about deep-seq and do you need more GPUs or can you do with less? But that, I think fear has been eliminated to a certain extent or alleviated.

>> So I'm going to give you a rundown of my notes and I want you to react to it just for folks watching, in case you missed a keynote. He started out talking about GeForce, it's 25 years old, took off growing San Jose to be bigger because they stayed here in San Jose, they now took over the whole streets. He called it the Super Bowl of the year or event. Last year was Woodstock, CUDA libraries by industry, he saw a major adoption. He introduced Cisco, NVIDIA, and T-Mobile for a 6G full stack radio-

>> Yeah, that was a huge .

>> That's huge, GM announced a self-driving car fleet deal with NVIDIA. Talked about air-cooled to liquid-cooled, 60K components per computer to 600K components per rack. 120 kilowatts full racks with liquid cooling. One exaflop computers in one rack, the computer node in the 'blade" my word blade, he didn't say blade. Blackwell systems shipping, Grace Blackwell NVLink 72, and a roadmap to like whatever, a thousand, whatever it was, 5,000. And then he went into the whole reasoning piece, set the table for what I was hearing about NVLink, a lot of discussion around the efficiencies around performance and how CUDA's adopting there. Of course, Dynamo and WrenchNet. He even threw Hopper under the bus.

>> Yes. Yeah.

>> He said, you can't give them away. Hopper should say don't buy those, buy the faster Blackwells with NVLink 8. Floating point 4 with Dynamo Blackwell was killer. And then he brought back this idea about revenue, he really made it a business imperative with energy, energy-bounded success. So we're bounded by energy now in the new model, that was a huge point.

>> And the latest chips will be more energy efficient, we know that.

>> And so that energy was a big topic, that's a scarce resource. The Vera Rubin chip, hat tip to her.

>> Rubin Ultra is in 2027.

>> The Ultra's coming out. The NVL72, that's in the second half. Blackwell, Ultra, second half, not this half.

>> Yes, second half of 2026. And in 2027, that'll be Rubin Ultra.

>> No the Vera Rubin, you said second half of 2026. The Ultra's is the 2027.

>> Ultra is 2027, yes.

>> That's 600 kilowatt rack.

>> That is .

>> That's going to be amazing. Sees the word extreme scale up, I love that word.

>> Speed is the new scale, actually we have been saying that for a couple of years now. But actually, I mean now it's growing to a giant, speed is the new scale. What is actually the whole idea behind all the computing? We want to go to the real time, we are not real time yet. We can't do real time inference, real time video analysis, audio analysis. NLP, can't do real time, there's a delay and then more and more compute will come into the picture and we'll do real time analysis of these things. And that's when we say, okay, now we are total digital on the factory floor.

>> He actually pointed out the big clouds too. A surge in AI infrastructure is coming, we are here on the ground with theCUBE bringing all the action. AI infrastructure is going to continue to accelerate at massive speeds now, it's coming to the finish line that's going to open up a next wave of innovation in software and data. You're going to start to see agents come on super fast, next 12 to 24 months will be spectacular. Of course, stay with theCUBE, we've got you covered from Silicon Valley to Wall Street. Join our trust network theCUBE network with NYSE Wired. We'll get more coverage here from the show floor after this short break, stay with us. Thanks for watching.