Advancing AI 2026 | Steve Berg, AMD

Clips
More from Advancing AI 2026

Steve Berg

CVP & GM Server CPU Cloud Business Group

AMD

play_circle_outline Hybrid cloud and rapid rise of neo-clouds/distributed compute

play_circle_outline CPUs Powering Agentic AI Inference: Heterogeneous SKUs, Orchestration, and GPU Optimization

play_circle_outline Venice EPYC Zen6: 1.7x Performance, Up to 256 Cores — Driving Cloud TCO Savings, Better Performance‑per‑Dollar and Vendor Neutrality

play_circle_outline Helios rack-scale architecture for hyperscale and neo-cloud deployments

play_circle_outline Globalized Sustainable Supply Chains and Onshoring: Tackling Memory Shortages That Hamper Delivery Speed

Info

Steve Berg, AMD

Steve Berg

CVP & GM Server CPU Cloud Business Group AMD

search

(INTRO)

John Furrier

>> Welcome back everyone to theCUBE's live coverage here in San Francisco, California for AMD's Advancing AI event. All the global leaders are here, industry leaders, as well as the free events, so a lot of practitioners and technologists checking out all the future AI infrastructure coming out of the oven from AMD, and of course the ecosystem. Steve Berg here, the Corporate Vice President and General Manager of the Cloud Business Group at AMD. Steve, thank you for coming on theCUBE.

Steve Berg

>> Cube, good to see you, thanks for spending the time.Yeah, thanks for having me, I appreciate it.

John Furrier

>> Cloud is dead, long live the cloud. cloud continues to thunder away, but the data center is booming, so hybrid cloud has not only won, we've been saying that on theCUBE for a long time, but more importantly, this sets the table for what compute will look like as inference and agents come in across the board, for the hyperscalers, neo -clouds, or enterprises, you're starting to see this distributed compute, connected intelligence layer emerging. What's your thoughts? You saw the growth of the cloud.

Steve Berg

>> Oh yeah.

John Furrier

>> Now you have eight more clouds, call them Neo -Clouds. They're going to be hyperscalers.

Steve Berg

>> It's really been phenomenal to see what's happened over even just the last, it's not even 18 months old that the agentic AI world is, right? The era that's existing now. But I've been doing this in the cloud for about 15 years and watching the explosion that's happening is amazing. and I'm really excited about what's happening with the GPU and how the GPU is becoming much more relevant again. Actually probably the most relevant element of the infrastructure in the agentic AI era today.

John Furrier

>> If you remember back in the day, because we are students of history being one of the first Amazon customers before they even had domains, long EC2 strings. But around 2012, 2013 we saw the explosion. It was basic building blocks and they had higher level services. And AI is in a similar place where Amazon beat the competition, HP tried to have a cloud, Dell tried to have a cloud, everyone tried to have a cloud, but that went to the main three. Now in AI you're seeing the rise of the neo-cloud, so a very rapid acceleration of the new hyperscalers. Enterprises are going to be their own clouds because they have their own sovereignty. They have to have their own hybrid strategy. And the inference is the killer app, so training will be there, but inference is becoming very much a compute -centric architecture.

Steve Berg

>> Yes.

John Furrier

>> This is key to all the stuff being talked about here and in the industry. This is a sea change. Compare the cloud growth and how that translates, if any, similarities to what's going on in AI, the emergence of basically what I call phase two of AI, where you're starting to see production workloads, higher level services, you hear compute routing.

Steve Berg

>> Yeah.

John Furrier

>> Cost is now on the table.

Steve Berg

>> That sounds familiar. It does sound familiar. And for my business, it's really important because all of the agents, all these agents that are being set up to orchestrate the workloads that are happening, right, and to route the data, and to go pre-process or post-process it, get it out, basically make sure that those really expensive GPUs are not waiting for information to run, right? And so the explosion of demand for GPUs in the last 18 months and what we're modeling in a forecast coming for the future is just amazing to me. And I think the new agentic era is driving this heterogeneous need for a lot of GPU compute with a range of different types of SKUs that we're seeing from all the cloud service providers asking us whether it's a storage workload or whether it's an HPC workload or whether it might be an inference even at the GPU level. We're seeing a need for lots of different types of GPU compute. It's not one GPU to rule them all, it's many. And that's where AMD is really well positioned.

Dave Vellante

>> So John and I both dropped a post this weekend after doing quite a bit of research this week. We've been following AMD for decades, but the team and John and I put together some research. I dropped like 3,500 words on Saturday. People say, John, nobody reads that many. You know who reads it? LLMs. Yeah. At any rate. And they summarize it really well.

Steve Berg

>> Exactly. And then I can pick it up.

Dave Vellante

>> Populating the intelligence. But at any rate, it was a deep dive, but I didn't have time to get into how you won the day in the cloud. And I was flying out here with Ashish from IDC, who follows this stuff, that's what IDC does. They count markets. And he said, AMD dominates in the cloud. You got well over, I think you got almost 50 % of the market in GPUs.

Steve Berg

>> Our target, on a revenue basis, that's our target by the end of the year.

Dave Vellante

>> Okay, so you're just under 50%, impressive. And he says well over 50 % in the cloud.

Steve Berg

>> Yeah.

Dave Vellante

>> So, my question is, can you take us through sort of how you won the day, number one, and number two is, how should we think about Venice and how will it extend that lead?

Steve Berg

>> Yeah, that's a great question. And it's been a long road. As prior to AMD starting to win in the cloud, there was a dominant provider. But we did it with products. All the way back to starting with Naples and Milan. getting into those design wins and teaching the cloud.

Dave Vellante

>> That's like 2017.

Steve Berg

>> Yeah, yeah, long time ago, but it's been a long road. But the products have led to being able to win in the cloud world, right? The design wins, we've been able to educate the customers, teach them where the TCO is, it's a TCO play. It's always a TCO play when it comes to cloud. And ensuring that they have the right performance per dollar for their workloads is how we've been able to become dominant in the cloud. Venice is going to significantly extend that, right? We've got 1.7x performance improvement on Venice. So we're going to extend our play there. We've got the most cores. Agentic plays really well for Venice because of all the higher core counts. and we have a whole platform, a whole stack of what we call OPNs or SKUs that will be able to meet all their different workload needs.

Dave Vellante

>> You know what's interesting, John, is when you think about the history, let's take AWS, you obviously know AWS deeply. they started working, of course, with Intel, and then they started working with AMD, and then they said, okay, we want to do our own thing. They started working with Annapurna, then they bought Annapurna, you know that. But despite all that, Steve, AMD continued to gain share throughout. I think a lot of people don't really understand it. They say, oh, Graviton, that's why, no, but why, it's software compatibility, it's ecosystem, it's affinity.

Steve Berg

>> Yeah, and it's the enterprises that run in the cloud that want to have that open ecosystem, and they don't want the lock-in. lock-in is not a preference for providers that are enterprises that are going to cloud or others that are going to cloud. They want to be able to run the same software and ecosystem that they do on-prem in a hybrid environment as they do in the cloud. So yeah, we've done a phenomenal job with getting more market segment share. The build your own cloud providers are going to have their own silicon strategy and they're always going to want that because of the economics associated with it. But we're going to do very well with whatever's left.

Dave Vellante

>> But the trick is, my experience,customers will choose value and risk lock-in, but the trick that you're pulling out of your hat is you're able to produce substantial value and preserve the openness. And that's oftentimes a balance that many in the industry can't strike. Usually it's a trade -off. And I wonder if you could comment on that. Is that a DNA thing? Is that a roadmap? Is it, what's the secret sauce that enables that?

Steve Berg

>> Well, I think first, it starts with AMD was an HPC company. We strive for the absolute best, highest performance. And you see that in the AWS infrastructure, right? When you look across AWS offerings of hardware, AMD's absolutely at the top. We're at the top of the stack. We also operate though within the value chain because we can skew our material down to areas that offer value to the customers at the price point that they need to adopt. And then Graviton and our competitors sit in between all of that. But we have the ability to differentiate our products within the high performance element, and we'll always strive for that, the best performance possible.

John Furrier

>> Yeah, we were talking on the last segment, a couple segments around the changing user interface. The cloud was a SaaS game.

Steve Berg

>> Yeah.

John Furrier

>> Those were applications. We talked about the GUI and how natural language is replacing the GUI. That changes, you work backwards from that, that changes the underlying thing, I think that's what AI's doing. If you look at AI, there's a lot of similarities with cloud And AI, because two things are happening. You still need the hyperscaler -like scale, but whether it's hyperscalers or neoclouds, they want massive data centers, low latency, AI factory -like things. But also you're hearing things like it's not the models, it's the system, which you guys are proposing, which we agree, we've been writing about that. But in the AI world, production are replacing the apps. If you think about it, because SaaS had records, you upload your address book maybe, but not like AI, you get your emails, you get all your work tickets, you can connect to anything and it brings all that in. That's essentially work.

Steve Berg

>> Yeah.

John Furrier

>> So the conversation interface, natural language, whether it's text or speech, gets into these domain production, but they're bringing in so much more data than say an application. Applications can do emails, read your Slack, and reason and do the work on your behalf.

Steve Berg

>> You're talking about all the tasks that happen.

John Furrier

>> Yeah, so you have the big infrastructure that you sell into, but also you have this emerging AI layer.

Steve Berg

>> Yes.

John Furrier

>> How do you rationalize that? Is this two separate groups? Is it like, okay, we'll still provide the servers and compute?

Steve Berg

>> It's not two separate groups. It's just what we look at as a cut of the demand for our compute, right? And we're starting to now model out the differences between the general traditional general purpose LLM, machine learning, GPU related demand, but there's also, what we call AI agentic type general purpose. Workloads that are tuned to task operations or tooling, right, and that's a sliver of the pie, but it's the same, it's still demanding and driving the same compute that was prior to this agentic era.

John Furrier

>> And you guys are doing the orchestration, that's another big theme here, that compute orchestration everything.

Steve Berg

>> It absolutely does.

John Furrier

>> So in that case, you're an operating system, Cloud's not the product, GPT's not the product, Llama's not the product, they're inputs.

Steve Berg

>> Yeah, it's like the GUI for driving all of these Agentic tasks and everything that happens on the hardware itself.

Dave Vellante

>> Well, it's interesting, Cloud is like the new legacy, right? Everybody talks about that.

Steve Berg

>> Modern legacy.

Dave Vellante

>> Modern legacy, you call it. Google today, just announced earnings like an hour or so ago, I guess,

Steve Berg

>> Yeah, I haven't looked at it yet.

Dave Vellante

>> I looked at it quickly, it looked like a blowout, while its stock's down after hours, but their cloud revenues, which everybody was predicting, the whisper number was 70%, it grew 82%, which means GCP, which always grows meaningfully faster than Google Cloud overall, probably grew, I don't know, 85 to 90%, amazing. CapEx, 20.5 billion. They increased their CapEx forecast, so that's probably why the stock's down, right, because everybody's freaking out, but to me, that's great news. All right, this continues.

John Furrier

>> They got 51.4 billion of backlog of contracted work that has not yet been recognized as revenue. So, how's the stock for AMD doing?

Steve Berg

>> It's quite simply great news, right? It's fantastic news. It's fantastic news. And I don't know when that's going to end, but we're reaping the rewards of those CapEx demand needs across the industry.

Dave Vellante

>> And AMD's able to deliver. It's not this quarter or next quarter. No, no, no. You look at the memory shortages.

Steve Berg

>> The difference between this and any other bubble we've seen before is there's real demand and real need behind all this.

John Furrier

>> there's real demand for build out, hence the backlog at Google. All the large scale providers, and the neo-clouds emerged super fast. So these services, SaaS was the cloud, now production are neo-clouds.

Steve Berg

>> It's real economic demand behind the...

John Furrier

>> All right, so take us through the enablement. Behind the scenes, you guys are working hard. We're seeing all the results here on display. we've been tracking it, extracting the signal out of this, which is you guys are building a new architecture to enable this, because it's not yesterday's cloud game with just SaaS, it's cloud plus hybrid and all these new use cases. What are some of the requirements you guys are thinking through? Obviously, compute orchestration, memory prices are high, supply chain, open source, distill this down, explain the strategy, explain that in common sense. what does it mean?

Steve Berg

>> Again, I think when you break down what the cloud customers are doing with their infrastructure, it's not really different than what it's been for the past few years, other than there are now these agents, there are now these autonomous capabilities that are driving just a massive amount of compute. So we're basically navigating what our architecture needs to look like to support these task -driven, I'll call them workloads, although that's the same work as a task that you were doing when you may call up Salesforce or do a query there, now can be done autonomously. So we're figuring out what the architecture needs to look like to drive that the best. And Venice is really well set up for that because again, 256 cores and the capabilities that it has is really going to help us capitalize on this.

John Furrier

>> What's the coolest thing you've seen in the past 12 months from your perspective? Okay, okay. And how does that fit into the picture? because that is rack scale deployment. It's hard to fit that into the telecom closet in the enterprise, but for the large scale AI factories.

Steve Berg

>> We will see cloud adoption and neo-cloud adoption for Helios without a doubt.

Dave Vellante

>> Hyper scale and neo-cloud.

Steve Berg

>> Yeah, yeah, for sure. Hyper scale and neo-cloud.

Dave Vellante

>> Because people think, oh no, hyper scale, they're going to build that themselves. But they will adopt Helios because of the ecosystem and all the things that we talked about before, right?

Steve Berg

>> Yeah, absolutely. And because AMD can provide all of the value from every piece of ingredient, GPU, GPU, NIC, all of the, on an open ecosystem, on something that all of the enterprises and users will be able to adopt easily. So we prefer non -lock-in versus...

Dave Vellante

>> Sorry, I cut you off. You were saying that we will see cloud adoption in the hyperscale cloud.

Steve Berg

>> We'll see cloud, yeah. From Helios. Absolutely. And it may be a version of Helios, but it'll be Helios architecture as a reference for sure.

Dave Vellante

>> So you customize that as necessary. How were you able to do that? Are you open?

John Furrier

>> Talk about the openness, because we've been following the Open Compute Project since it was formed, we watched that evolution, a lot of goodness came out of that. Talk about the heterogeneous nature. Okay, I have Helios, of course, that can be big centers. But distributed computing, you're going to have smaller footprints, you guys are playing well there. When you connect all this together, take us through that because a lot of people are saying, okay, where's the value for me? And then the applications go to production.

Steve Berg

>> Yeah, again, I think it'll be based on, the value will be in the customer needs. And our value is again providing a performance per dollar, a TCO per dollar story that is going to be interesting for the end users and for the cloud providers to get away from egregious type pricing and egregious type monetization. And we have all the pieces of the puzzle. That's the beauty of Helios, is it all comes together, and it can be with a partner that they trust. AMD has been able to deliver year over year, gen over gen with our Zen architecture. We're going to do the same thing with Instinct. And I think customers are going to appreciate having a vendor that has the capabilities that we do and the performance that we do at reasonable economics.

John Furrier

>> So two areas that come up a lot in conversations with customers we talk to are supply chain and software. They want to have diversity, and there's a little bit of risk management obviously with the price of, say, memory, but when they look at the cost of running their operations, this is now becoming, not just token economics, tokenomics, what do you want to call it, about token maxing, which has burned a lot of budgets. They are looking at a long-term view of how do I architect a generational shift? They have to look at the operating factors. How important is that open heterogeneous perspective?

Steve Berg

>> I think it's very important. Again, it's about providing the choice and providing that open ecosystem and software stack, ROCm and everything that we have that allows them to innovate on their own business plans.

John Furrier

>> You must get a lot of phone calls hey I need that rack immediately. So there's a lot of pressure, we saw some of the backlog. And we know that all the hyperscalers have this backlog and they're doing debt financing, they're ramping up, the demand is huge, the demand curve's looking good, so really not a lot of bubble. Maybe inflation on price and expectations, but there's demand for, how fast can you deliver that? Yeah. What's the speed of deployments?

Steve Berg

>> I am only constrained by the memory market at this point. And yes, the phone is ringing off the hook for more GPUs and more compute. And if you guys have any memory you can provide, bundle it. It's going to be a side business for us. We'd be selling it on eBay if we did. Yeah, yeah. That's the biggest constraint.

Dave Vellante

>> So, I'm reading this great article. I think it's out of SemiAnalysis. You guys officially announced Venice in May.

Steve Berg

>> Yeah.

Dave Vellante

>> Sixth generation.

Steve Berg

>> Zen

Dave Vellante

>> 6. Zen 6, on TSMC's N2 process. And so, I think you're, and they're talking about how you're going to talk about it at Advancing AI, so it's all.

Steve Berg

>> It'll be, tomorrow you'll see a lot more information.

Dave Vellante

>> Okay, great, so we can't talk too much about that, but it's here, it's coming, it's ramping, yields are up. It's like go time.

Steve Berg

>> Manufacturing, it's go time, yeah. It's go time, and we've been doing design wins with all of our cloud service providers over the course of the last year, so you're going to see a lot of, you've seen Microsoft's announcements on it just recently too. They have two new instances that are Venice-based and we're really excited about that. And you'll probably hear a lot more in the coming months on other cloud service providers that have been designing in Venice.

John Furrier

>> Final question on agentic, how would you peg the progress of adoption? We've seen over the last year some great use cases, but not full scale, but definitely a lot of tire kicking, a lot of deployments in production, but not massive. What's the pace or status, in your opinion, of where agentic is and where we'll be?

Steve Berg

>> It's irrational how much it's growing. And I can't tell you enough about how excited we are about the growth and the demand that it's creating for EPYC-based compute. Again, we're positioned so well to capitalize on this because Venice will be the best EPYC, will be the best agentic processor on the planet. And I think everyone's going to put it into their infrastructure to run agentic workloads.

John Furrier

>> And why is that? What would you say they're being used for?

Steve Berg

>> The diversity of the stack. we have 256 cores. You're going to see a whole lot of other compute that becomes available to run every single workload fathomable out there. And agents are running on typical workloads from the past. The workloads haven't changed. It's autonomously driven. and the diversity of our portfolio lends really well to it. It's not a one SKU fits all.

John Furrier

>> It fits nicely into the price points that people want to be at, but also helps with the GPU utilization.

Steve Berg

>> Exactly.

John Furrier

>> It feeds the GPUs at the right time, we've got routing of models, the intelligence starts to come together.

Steve Berg

>> It can scale the stack to meet the individual workload needs, whatever they might be using agentic for, and make it available basically immediately. What are we doing?

Dave Vellante

>> Some of the production for Venice is happening, some percentage is happening onshore, correct?

Steve Berg

>> There is, we are running today, Turin in Arizona.

Dave Vellante

>> Yeah, right, so at some point, that'll evolve.

Steve Berg

>> Exactly.The supply chain's going to be very, global and sustainable for Venice. Yeah, demand and. Our supply chain is very strong.

John Furrier

>> Yeah, invention comes from a lot of those constraints. strange to work around, congratulations. Well Steve, great to have you on, and I guess my final question is, what's your focus for the rest of the year? You got supercomputing coming out, you got the HPC, I mean high performance computing is now called supercomputing.

Steve Berg

>> Right, my focus is to ensure that all the cloud service providers are ready to launch this product when it becomes generally available, and on time to market, and being the, again, the world's best GPU for AI, agentic and ubiquitous across all the clouds and neo -clouds and anybody who needs it. That's my focus.

John Furrier

>> Well we'll see you around, Supercomputing, re:Invent. Great to see you, thanks for coming on.

Dave Vellante

>> We got a cloud partner coming up next.

John Furrier

>> All right, Steve Berg, Corporate Vice President, General Manager of the Cloud Business Group. Again, cloud computing set the table. You see Kubernetes, cloud native hybrid powering all this AI infrastructure. Compute is a perfect storm for all the agents coming in to create more enablement, more agency for the users. That's what we're hearing here. I'm John Furrier with Dave Vellante. Thanks for watching, we'll be right back after this short break.

Steve Berg, AMD

search

(INTRO)

John Furrier

>> Welcome back everyone to theCUBE's live coverage here in San Francisco, California for AMD's Advancing AI event. All the global leaders are here, industry leaders, as well as the free events, so a lot of practitioners and technologists checking out all the future AI infrastructure coming out of the oven from AMD, and of course the ecosystem. Steve Berg here, the Corporate Vice President and General Manager of the Cloud Business Group at AMD. Steve, thank you for coming on theCUBE.

Steve Berg

>> Cube, good to see you, thanks for spending the time.Yeah, thanks for having me, I appreciate it.

John Furrier

>> Cloud is dead, long live the cloud. cloud continues to thunder away, but the data center is booming, so hybrid cloud has not only won, we've been saying that on theCUBE for a long time, but more importantly, this sets the table for what compute will look like as inference and agents come in across the board, for the hyperscalers, neo -clouds, or enterprises, you're starting to see this distributed compute, connected intelligence layer emerging. What's your thoughts? You saw the growth of the cloud.

Steve Berg

>> Oh yeah.

John Furrier

>> Now you have eight more clouds, call them Neo -Clouds. They're going to be hyperscalers.

Steve Berg

>> It's really been phenomenal to see what's happened over even just the last, it's not even 18 months old that the agentic AI world is, right? The era that's existing now. But I've been doing this in the cloud for about 15 years and watching the explosion that's happening is amazing. and I'm really excited about what's happening with the GPU and how the GPU is becoming much more relevant again. Actually probably the most relevant element of the infrastructure in the agentic AI era today.

John Furrier

>> If you remember back in the day, because we are students of history being one of the first Amazon customers before they even had domains, long EC2 strings. But around 2012, 2013 we saw the explosion. It was basic building blocks and they had higher level services. And AI is in a similar place where Amazon beat the competition, HP tried to have a cloud, Dell tried to have a cloud, everyone tried to have a cloud, but that went to the main three. Now in AI you're seeing the rise of the neo-cloud, so a very rapid acceleration of the new hyperscalers. Enterprises are going to be their own clouds because they have their own sovereignty. They have to have their own hybrid strategy. And the inference is the killer app, so training will be there, but inference is becoming very much a compute -centric architecture.

Steve Berg

>> Yes.

John Furrier

>> This is key to all the stuff being talked about here and in the industry. This is a sea change. Compare the cloud growth and how that translates, if any, similarities to what's going on in AI, the emergence of basically what I call phase two of AI, where you're starting to see production workloads, higher level services, you hear compute routing.

Steve Berg

>> Yeah.

John Furrier

>> Cost is now on the table.

Steve Berg

>> That sounds familiar. It does sound familiar. And for my business, it's really important because all of the agents, all these agents that are being set up to orchestrate the workloads that are happening, right, and to route the data, and to go pre-process or post-process it, get it out, basically make sure that those really expensive GPUs are not waiting for information to run, right? And so the explosion of demand for GPUs in the last 18 months and what we're modeling in a forecast coming for the future is just amazing to me. And I think the new agentic era is driving this heterogeneous need for a lot of GPU compute with a range of different types of SKUs that we're seeing from all the cloud service providers asking us whether it's a storage workload or whether it's an HPC workload or whether it might be an inference even at the GPU level. We're seeing a need for lots of different types of GPU compute. It's not one GPU to rule them all, it's many. And that's where AMD is really well positioned.

Dave Vellante

>> So John and I both dropped a post this weekend after doing quite a bit of research this week. We've been following AMD for decades, but the team and John and I put together some research. I dropped like 3,500 words on Saturday. People say, John, nobody reads that many. You know who reads it? LLMs. Yeah. At any rate. And they summarize it really well.

Steve Berg

>> Exactly. And then I can pick it up.

Dave Vellante

>> Populating the intelligence. But at any rate, it was a deep dive, but I didn't have time to get into how you won the day in the cloud. And I was flying out here with Ashish from IDC, who follows this stuff, that's what IDC does. They count markets. And he said, AMD dominates in the cloud. You got well over, I think you got almost 50 % of the market in GPUs.

Steve Berg

>> Our target, on a revenue basis, that's our target by the end of the year.

Dave Vellante

>> Okay, so you're just under 50%, impressive. And he says well over 50 % in the cloud.

Steve Berg

>> Yeah.

Dave Vellante

>> So, my question is, can you take us through sort of how you won the day, number one, and number two is, how should we think about Venice and how will it extend that lead?

Steve Berg

>> Yeah, that's a great question. And it's been a long road. As prior to AMD starting to win in the cloud, there was a dominant provider. But we did it with products. All the way back to starting with Naples and Milan. getting into those design wins and teaching the cloud.

Dave Vellante

>> That's like 2017.

Steve Berg

>> Yeah, yeah, long time ago, but it's been a long road. But the products have led to being able to win in the cloud world, right? The design wins, we've been able to educate the customers, teach them where the TCO is, it's a TCO play. It's always a TCO play when it comes to cloud. And ensuring that they have the right performance per dollar for their workloads is how we've been able to become dominant in the cloud. Venice is going to significantly extend that, right? We've got 1.7x performance improvement on Venice. So we're going to extend our play there. We've got the most cores. Agentic plays really well for Venice because of all the higher core counts. and we have a whole platform, a whole stack of what we call OPNs or SKUs that will be able to meet all their different workload needs.

Dave Vellante

>> You know what's interesting, John, is when you think about the history, let's take AWS, you obviously know AWS deeply. they started working, of course, with Intel, and then they started working with AMD, and then they said, okay, we want to do our own thing. They started working with Annapurna, then they bought Annapurna, you know that. But despite all that, Steve, AMD continued to gain share throughout. I think a lot of people don't really understand it. They say, oh, Graviton, that's why, no, but why, it's software compatibility, it's ecosystem, it's affinity.

Steve Berg

>> Yeah, and it's the enterprises that run in the cloud that want to have that open ecosystem, and they don't want the lock-in. lock-in is not a preference for providers that are enterprises that are going to cloud or others that are going to cloud. They want to be able to run the same software and ecosystem that they do on-prem in a hybrid environment as they do in the cloud. So yeah, we've done a phenomenal job with getting more market segment share. The build your own cloud providers are going to have their own silicon strategy and they're always going to want that because of the economics associated with it. But we're going to do very well with whatever's left.

Dave Vellante

>> But the trick is, my experience,customers will choose value and risk lock-in, but the trick that you're pulling out of your hat is you're able to produce substantial value and preserve the openness. And that's oftentimes a balance that many in the industry can't strike. Usually it's a trade -off. And I wonder if you could comment on that. Is that a DNA thing? Is that a roadmap? Is it, what's the secret sauce that enables that?

Steve Berg

>> Well, I think first, it starts with AMD was an HPC company. We strive for the absolute best, highest performance. And you see that in the AWS infrastructure, right? When you look across AWS offerings of hardware, AMD's absolutely at the top. We're at the top of the stack. We also operate though within the value chain because we can skew our material down to areas that offer value to the customers at the price point that they need to adopt. And then Graviton and our competitors sit in between all of that. But we have the ability to differentiate our products within the high performance element, and we'll always strive for that, the best performance possible.

John Furrier

>> Yeah, we were talking on the last segment, a couple segments around the changing user interface. The cloud was a SaaS game.

Steve Berg

>> Yeah.

John Furrier

>> Those were applications. We talked about the GUI and how natural language is replacing the GUI. That changes, you work backwards from that, that changes the underlying thing, I think that's what AI's doing. If you look at AI, there's a lot of similarities with cloud And AI, because two things are happening. You still need the hyperscaler -like scale, but whether it's hyperscalers or neoclouds, they want massive data centers, low latency, AI factory -like things. But also you're hearing things like it's not the models, it's the system, which you guys are proposing, which we agree, we've been writing about that. But in the AI world, production are replacing the apps. If you think about it, because SaaS had records, you upload your address book maybe, but not like AI, you get your emails, you get all your work tickets, you can connect to anything and it brings all that in. That's essentially work.

Steve Berg

>> Yeah.

John Furrier

>> So the conversation interface, natural language, whether it's text or speech, gets into these domain production, but they're bringing in so much more data than say an application. Applications can do emails, read your Slack, and reason and do the work on your behalf.

Steve Berg

>> You're talking about all the tasks that happen.

John Furrier

>> Yeah, so you have the big infrastructure that you sell into, but also you have this emerging AI layer.

Steve Berg

>> Yes.

John Furrier

>> How do you rationalize that? Is this two separate groups? Is it like, okay, we'll still provide the servers and compute?

Steve Berg

>> It's not two separate groups. It's just what we look at as a cut of the demand for our compute, right? And we're starting to now model out the differences between the general traditional general purpose LLM, machine learning, GPU related demand, but there's also, what we call AI agentic type general purpose. Workloads that are tuned to task operations or tooling, right, and that's a sliver of the pie, but it's the same, it's still demanding and driving the same compute that was prior to this agentic era.

John Furrier

>> And you guys are doing the orchestration, that's another big theme here, that compute orchestration everything.

Steve Berg

>> It absolutely does.

John Furrier

>> So in that case, you're an operating system, Cloud's not the product, GPT's not the product, Llama's not the product, they're inputs.

Steve Berg

>> Yeah, it's like the GUI for driving all of these Agentic tasks and everything that happens on the hardware itself.

Dave Vellante

>> Well, it's interesting, Cloud is like the new legacy, right? Everybody talks about that.

Steve Berg

>> Modern legacy.

Dave Vellante

>> Modern legacy, you call it. Google today, just announced earnings like an hour or so ago, I guess,

Steve Berg

>> Yeah, I haven't looked at it yet.

Dave Vellante

>> I looked at it quickly, it looked like a blowout, while its stock's down after hours, but their cloud revenues, which everybody was predicting, the whisper number was 70%, it grew 82%, which means GCP, which always grows meaningfully faster than Google Cloud overall, probably grew, I don't know, 85 to 90%, amazing. CapEx, 20.5 billion. They increased their CapEx forecast, so that's probably why the stock's down, right, because everybody's freaking out, but to me, that's great news. All right, this continues.

John Furrier

>> They got 51.4 billion of backlog of contracted work that has not yet been recognized as revenue. So, how's the stock for AMD doing?

Steve Berg

>> It's quite simply great news, right? It's fantastic news. It's fantastic news. And I don't know when that's going to end, but we're reaping the rewards of those CapEx demand needs across the industry.

Dave Vellante

>> And AMD's able to deliver. It's not this quarter or next quarter. No, no, no. You look at the memory shortages.

Steve Berg

>> The difference between this and any other bubble we've seen before is there's real demand and real need behind all this.

John Furrier

>> there's real demand for build out, hence the backlog at Google. All the large scale providers, and the neo-clouds emerged super fast. So these services, SaaS was the cloud, now production are neo-clouds.

Steve Berg

>> It's real economic demand behind the...

John Furrier

>> All right, so take us through the enablement. Behind the scenes, you guys are working hard. We're seeing all the results here on display. we've been tracking it, extracting the signal out of this, which is you guys are building a new architecture to enable this, because it's not yesterday's cloud game with just SaaS, it's cloud plus hybrid and all these new use cases. What are some of the requirements you guys are thinking through? Obviously, compute orchestration, memory prices are high, supply chain, open source, distill this down, explain the strategy, explain that in common sense. what does it mean?

Steve Berg

>> Again, I think when you break down what the cloud customers are doing with their infrastructure, it's not really different than what it's been for the past few years, other than there are now these agents, there are now these autonomous capabilities that are driving just a massive amount of compute. So we're basically navigating what our architecture needs to look like to support these task -driven, I'll call them workloads, although that's the same work as a task that you were doing when you may call up Salesforce or do a query there, now can be done autonomously. So we're figuring out what the architecture needs to look like to drive that the best. And Venice is really well set up for that because again, 256 cores and the capabilities that it has is really going to help us capitalize on this.

John Furrier

>> What's the coolest thing you've seen in the past 12 months from your perspective? Okay, okay. And how does that fit into the picture? because that is rack scale deployment. It's hard to fit that into the telecom closet in the enterprise, but for the large scale AI factories.

Steve Berg

>> We will see cloud adoption and neo-cloud adoption for Helios without a doubt.

Dave Vellante

>> Hyper scale and neo-cloud.

Steve Berg

>> Yeah, yeah, for sure. Hyper scale and neo-cloud.

Dave Vellante

>> Because people think, oh no, hyper scale, they're going to build that themselves. But they will adopt Helios because of the ecosystem and all the things that we talked about before, right?

Steve Berg

>> Yeah, absolutely. And because AMD can provide all of the value from every piece of ingredient, GPU, GPU, NIC, all of the, on an open ecosystem, on something that all of the enterprises and users will be able to adopt easily. So we prefer non -lock-in versus...

Dave Vellante

>> Sorry, I cut you off. You were saying that we will see cloud adoption in the hyperscale cloud.

Steve Berg

>> We'll see cloud, yeah. From Helios. Absolutely. And it may be a version of Helios, but it'll be Helios architecture as a reference for sure.

Dave Vellante

>> So you customize that as necessary. How were you able to do that? Are you open?

John Furrier

>> Talk about the openness, because we've been following the Open Compute Project since it was formed, we watched that evolution, a lot of goodness came out of that. Talk about the heterogeneous nature. Okay, I have Helios, of course, that can be big centers. But distributed computing, you're going to have smaller footprints, you guys are playing well there. When you connect all this together, take us through that because a lot of people are saying, okay, where's the value for me? And then the applications go to production.

Steve Berg

>> Yeah, again, I think it'll be based on, the value will be in the customer needs. And our value is again providing a performance per dollar, a TCO per dollar story that is going to be interesting for the end users and for the cloud providers to get away from egregious type pricing and egregious type monetization. And we have all the pieces of the puzzle. That's the beauty of Helios, is it all comes together, and it can be with a partner that they trust. AMD has been able to deliver year over year, gen over gen with our Zen architecture. We're going to do the same thing with Instinct. And I think customers are going to appreciate having a vendor that has the capabilities that we do and the performance that we do at reasonable economics.

John Furrier

>> So two areas that come up a lot in conversations with customers we talk to are supply chain and software. They want to have diversity, and there's a little bit of risk management obviously with the price of, say, memory, but when they look at the cost of running their operations, this is now becoming, not just token economics, tokenomics, what do you want to call it, about token maxing, which has burned a lot of budgets. They are looking at a long-term view of how do I architect a generational shift? They have to look at the operating factors. How important is that open heterogeneous perspective?

Steve Berg

>> I think it's very important. Again, it's about providing the choice and providing that open ecosystem and software stack, ROCm and everything that we have that allows them to innovate on their own business plans.

John Furrier

>> You must get a lot of phone calls hey I need that rack immediately. So there's a lot of pressure, we saw some of the backlog. And we know that all the hyperscalers have this backlog and they're doing debt financing, they're ramping up, the demand is huge, the demand curve's looking good, so really not a lot of bubble. Maybe inflation on price and expectations, but there's demand for, how fast can you deliver that? Yeah. What's the speed of deployments?

Steve Berg

>> I am only constrained by the memory market at this point. And yes, the phone is ringing off the hook for more GPUs and more compute. And if you guys have any memory you can provide, bundle it. It's going to be a side business for us. We'd be selling it on eBay if we did. Yeah, yeah. That's the biggest constraint.

Dave Vellante

>> So, I'm reading this great article. I think it's out of SemiAnalysis. You guys officially announced Venice in May.

Steve Berg

>> Yeah.

Dave Vellante

>> Sixth generation.

Steve Berg

>> Zen

Dave Vellante

>> 6. Zen 6, on TSMC's N2 process. And so, I think you're, and they're talking about how you're going to talk about it at Advancing AI, so it's all.

Steve Berg

>> It'll be, tomorrow you'll see a lot more information.

Dave Vellante

>> Okay, great, so we can't talk too much about that, but it's here, it's coming, it's ramping, yields are up. It's like go time.

Steve Berg

>> Manufacturing, it's go time, yeah. It's go time, and we've been doing design wins with all of our cloud service providers over the course of the last year, so you're going to see a lot of, you've seen Microsoft's announcements on it just recently too. They have two new instances that are Venice-based and we're really excited about that. And you'll probably hear a lot more in the coming months on other cloud service providers that have been designing in Venice.

John Furrier

>> Final question on agentic, how would you peg the progress of adoption? We've seen over the last year some great use cases, but not full scale, but definitely a lot of tire kicking, a lot of deployments in production, but not massive. What's the pace or status, in your opinion, of where agentic is and where we'll be?

Steve Berg

>> It's irrational how much it's growing. And I can't tell you enough about how excited we are about the growth and the demand that it's creating for EPYC-based compute. Again, we're positioned so well to capitalize on this because Venice will be the best EPYC, will be the best agentic processor on the planet. And I think everyone's going to put it into their infrastructure to run agentic workloads.

John Furrier

>> And why is that? What would you say they're being used for?

Steve Berg

>> The diversity of the stack. we have 256 cores. You're going to see a whole lot of other compute that becomes available to run every single workload fathomable out there. And agents are running on typical workloads from the past. The workloads haven't changed. It's autonomously driven. and the diversity of our portfolio lends really well to it. It's not a one SKU fits all.

John Furrier

>> It fits nicely into the price points that people want to be at, but also helps with the GPU utilization.

Steve Berg

>> Exactly.

John Furrier

>> It feeds the GPUs at the right time, we've got routing of models, the intelligence starts to come together.

Steve Berg

>> It can scale the stack to meet the individual workload needs, whatever they might be using agentic for, and make it available basically immediately. What are we doing?

Dave Vellante

>> Some of the production for Venice is happening, some percentage is happening onshore, correct?

Steve Berg

>> There is, we are running today, Turin in Arizona.

Dave Vellante

>> Yeah, right, so at some point, that'll evolve.

Steve Berg

>> Exactly.The supply chain's going to be very, global and sustainable for Venice. Yeah, demand and. Our supply chain is very strong.

John Furrier

>> Yeah, invention comes from a lot of those constraints. strange to work around, congratulations. Well Steve, great to have you on, and I guess my final question is, what's your focus for the rest of the year? You got supercomputing coming out, you got the HPC, I mean high performance computing is now called supercomputing.

Steve Berg

>> Right, my focus is to ensure that all the cloud service providers are ready to launch this product when it becomes generally available, and on time to market, and being the, again, the world's best GPU for AI, agentic and ubiquitous across all the clouds and neo -clouds and anybody who needs it. That's my focus.

John Furrier

>> Well we'll see you around, Supercomputing, re:Invent. Great to see you, thanks for coming on.

Dave Vellante

>> We got a cloud partner coming up next.

John Furrier

>> All right, Steve Berg, Corporate Vice President, General Manager of the Cloud Business Group. Again, cloud computing set the table. You see Kubernetes, cloud native hybrid powering all this AI infrastructure. Compute is a perfect storm for all the agents coming in to create more enablement, more agency for the users. That's what we're hearing here. I'm John Furrier with Dave Vellante. Thanks for watching, we'll be right back after this short break.