Advancing AI 2026 | Art Baudo, AWS

Clips
More from Advancing AI 2026

Art Baudo

Principal Product Marketing Manager & Head of EC2 Product Marketing

AWS

play_circle_outline AWS at 20: Cloud Growth, CapEx-Driven Infrastructure Shifts, and Rising CPU Demand for AI Token and Agent Workloads

play_circle_outline AMD EPYC Adoption Since 2018 and 2026 Turin Launch: High-Frequency X8aedz and Memory-Optimized M8azn Instances

play_circle_outline EDA and simulation workloads (e.g., Rivian crash simulation) moving to cloud

play_circle_outline Instance migration and choice: seamless transitions between generations, architecture compatibility

play_circle_outline Workload/model routing across CPU and GPU tiers for optimized cost and performance

Info

Art Baudo, AWS

Art Baudo

Principal Product Marketing Manager & Head of EC2 Product Marketing AWS

search

John Furrier

>> Welcome back, everyone, to theCUBE's live coverage here in San Francisco of AMD's Advancing AI event. All the leaders are here in the industry, AMD, and also practitioners. It's a free event, so it's a packed house, completely sold out. I'm John Furrier with Dave Vellante, my co-host. We're trying to bring all the AI knowledge to you. We got a great guest, CUBE alumni Art Baudo, Principal Product Marketing Manager, head of EC2 product marketing at AWS, Amazon Web Services. Art, great to see you again. Thanks for coming back on in person.

Art Baudo

>> Yeah, it's nice to see you, John and Dave. It's great to be here.

John Furrier

>> It's very rare that we get the two of us in one spot, but this is an epic event, pun

Art Baudo

>> intended.I love it.

John Furrier

>> So obviously we've been following AWS. Everyone knows the history. We actually started with AWS in 2007 when I first started using it. They didn't have domains at the time. I used to joke with, Matt Garman and Dave Brown, that it was a long EC2 string. Then they had domains. But really with theCUBE since 2013 and cloud made the market, cloud native, cloud, AWS made SaaS. Now we're in another era where SaaS is moving to projects. AI is now involved. You can bring in all kinds of artifacts into the AI, into the compute, email it, can be scanned, Slack messages. You got tools developing Q and Kiro you guys built. You got all kinds of really kind of forward deployed engineering. So it's really a renaissance that feels early days of AWS in the AI realm because the interface is changing. All this new value is coming in, but it's built on cloud.

Art Baudo

>> Yeah.

John Furrier

>> So you guys have been doing well. You did a debt financing. I saw Andy Jassy announce that you're spending a lot of CapEx, you're building out more and more horsepower. So give us a state of the compute landscape because it seems to be with agents coming in, AgentCore a year ago, I saw Swami do an announcement on that yesterday, one year anniversary, the growth. Give us a perspective inside AWS, what's it like?

Art Baudo

>> Well, first, one of the things I think you bring up 2007, but the interesting thing to me is, well, this is our 20th year of EC2 and also the 20th year of S3. And I think those two anniversaries this year mark kind of some pretty major milestones in terms of where we have seen customer growth and customers change over that period of time. we've added thousands of instances now, instance types, and we see customers on the AI side expanding what they're using. We offer a broad depth of AI platforms that you can use, but also a lot of customers are also using, with the additional performance on the CPU side, we're seeing customers do agentic, AI on the CPU as well. We did some demos as well here showing how people can send tokens into the CPU as well. So now with AI taking off, we're seeing people use it in multiple different spaces across the industry and across the instance types as well. And I've never seen so many things start to move into the cloud, things that I think we hadn't thought originally would move or other people hadn't thought initially would move over to the cloud. today I was with Marvell giving an example how they're doing electronic design automation for silicon development. They're moving that into the cloud and how much they've been able to expand. So because of the resources of the cloud, which is to me very exciting.

John Furrier

>> we've been covering the compute side of EC2. We're here at the AMD event. Just talk about the transition of AMD as a company because they have been a supplier to AWS. What we're hearing here today is a systems architecture shift for the company. Obviously, their stock's doing great because there's a huge demand curve for infrastructure, but it seems to be a different game for the semiconductor companies because they've got to build out systems because there's a lot of cloud and on-prem and edge growth. What's your relationship with AMD now? And how has it changed?

Art Baudo

>> Just so I don't— I think that, when we started, we introduced AMD EPYC processors back in 2018, and we've introduced every generation of EPYC processor to date with Turin. Some of our Turin instances actually just launched in 2026. So very prescient to that discussion. But when we first started, the EPYC instances were using multithreading. And once we moved to the 7th generation, we moved to single-threaded instances, with the performance going from 6th generation to 7th generation, that is some pretty good improvement in performance, raw performance that it's been able to deliver. And customers in the AI space in particular and other spaces in the EDA space are able to take advantage of those performance improvements. We also have high frequency instances as well that are available on the AMD platform. Our X8aedz and our M8azn ones have the high 5 GHz frequency as well. So that provides performance frequency, memory as well. So we have been expanding that overall portfolio since we started that introduction in 2018.

John Furrier

>> What use cases are emerging that you could share? Because a lot of people have reined in cloud spend. Now they've got tokenomics. They're starting to see real value being created by deploying on cloud, some really great use cases. How has that changed what they do and how has that changed the price performance opportunities? what's the pattern recognition?

Art Baudo

>> My take is that as we've added more performance, we've tried to continuously make sure we optimize on price performance for customers. So some of our most recent instances continue to see a reduction in price performance. That is a continuous focus from the AWS side, and that is no different on the AMD instances. As that performance has gone up, so has the price performance that we've delivered to customers. We're seeing— we continue to see a wide variety of workloads that are happening. But one of the things I also mentioned, we're seeing a need to do more simulation for customers, that those simulations that you can do in the cloud— Rivian today was talking about how they're simulating vehicle crashes using the cloud. They need the breadth of the CPUs, the scaling of the CPUs in the cloud to be able to do that. They see the same on the EDA side as the complexity of design is increasing. There is a continuous need for continued improvement in performance that exists in the cloud.

Dave Vellante

>> I want to ask you, you mentioned you started adopting EPYC in 2017. I think that's the same year you launched Nitro, roughly, I think.

Art Baudo

>> Yeah. So we adopted EPYC in 2018 and we launched Nitro in 2017.

Dave Vellante

>> Okay, there you go.

Art Baudo

>> But obviously all of the EPYC instances are off of Nitro.

Dave Vellante

>> Right, so I wanted you to talk about the value, customer value of Nitro. We wrote a piece, our research team wrote a piece years ago on AWS's secret weapon.

John Furrier

>> Dave loves Nitro.

Dave Vellante

>> Well, Nitro was like, yeah, I love Nitro too. It really was, is AWS's secret weapon. But at any rate, the customer value was it gave you the ability to provide optionality and security that you can't go in and see, and enclaves and the like. So I wonder if you could discuss that, both the optionality and the security angle.

Art Baudo

>> Yeah, so first on the optionality, the Nitro System has allowed us, we believe, we credit the Nitro System with being able to deploy so many instances for customers. It has allowed us to faster deploy the instances that are out there. And we believe that has also contributed to performance as well for customers. The Nitro System has allowed us to offload that, some of the functions over into hardware and software. Our hypervisor has been offloaded over there onto the Nitro System. And also with the Nitro System, we have zero operator access, something that over time we've also worked to showcase even more to customers. We did in 2023, we worked with the NCC Group to have our architecture analyzed and the NCC Group Talked about how our architecture is designed so that no one can access that architecture. We've introduced attestation for software-based instances as well. So security as well, that's been in there. So it's been something that evolution has paired with our evolution overall desire to deliver security and performance at scale to customers.

Dave Vellante

>> So how does it fit vis-à-vis AI? What are the, are there trade-offs, or is it just a sort of a natural extension into AI? How are you thinking about that?

Art Baudo

>> So I see it as a natural extension, frankly. So customers are, with the additional performance that the CPUs have been getting, we're seeing customers use the CPUs for AI. The Nitro System is something that's deployed across our entire ecosystem. So it's available for customers. I know we're here at an AMD event, But the Nitro System is deployed across our entire portfolio, from our CPU portfolio to the GPU portfolio and to our own Trainium and Inferentia and Graviton-based products as well.

John Furrier

>> On migration, this comes up a lot. We're hearing a lot of people trying to migrate in the past year or so, maybe 18 months to start here. The— now we're hearing with AI, migration is going faster. You guys have published a lot of examples of migrations for this, that platform, but When you get down to migrating to infrastructure, what's that look like? What are people doing or is that abstracted away completely? If I want to move from Graviton to AMD or how do I mix and match the best elements in the design?

Art Baudo

>> One of the things we try to do is give customers choice. Customers are able to make that decision and migration from an instance type to another, we try to make as seamless as possible. Specifically, if you're going from, 7th generation to 8th generation instances, the architecture is not changing. If you're sticking, let's say, with AMD instances or a Graviton-based instance, so you're able to migrate very quickly, achieve the performance benefits. And if you're someone that needs both that performance and, you're looking for the price performance to decline, that upgrade affords you that ability to do that. And we're seeing customers make that trade-off.

Dave Vellante

>> We were talking earlier, I'd love your opinion on this, Art. About, the old, the Gartner term of the split IT.

John Furrier

>> Bimodal.

Dave Vellante

>> Bimodal IT, remember that? It was old stuff here, new stuff here. You've got a massive install base of applications running general purpose computing, x86, slash, infrastructure. How do you see those two worlds, the AI and the general purpose world, Are they coming together in a new fused architecture? Will they be bimodal? How do you see that playing out?

Art Baudo

>> I see them as complementary. They can exist in the same space. And I think that depending upon the industry also depends upon how they're using some of those AI skill sets. As I mentioned before, with, folks such as Rivian doing and now exploring more AI to help with their simulation and modeling, they're pairing that with some of their knowledge and their existing software and their existing databases. And I see that to be something that customers are doing. They're adding this AI as an additional layer to support some of the fundamentals that they have that exist right now.

John Furrier

>> Has Agentic changed your view on EC2, all the different elements on the infrastructure side? How has that changed, if at all, how you guys operate?

Art Baudo

>> I wouldn't say that it has changed how we operate. From the perspective of, Amazon is one of the most customer-centric companies that's out there. So we listen to what customers are looking for, and that's how these instances come about. And I would say that it's only made me actually more optimistic about EC2. The reasons I think to go to the cloud and use and leverage the performance that exists are even greater now today. And I think the fundamentals of the cloud, the ability to deliver security, the ability to deliver performance on demand for customers, especially in some of these workloads where you have a really, really heavy burst and a need for lots of CPU performance, and then suddenly you may not need it for a period of time. The cloud is phenomenal at doing that.

Dave Vellante

>> I want to ask you about that customer obsession. John, you and I have asked, Andy Jassy this question years ago about why the focus on primitives, and his answer was always pretty consistent, because it allows us to change quickly as the market changes. And my question is, as you focus on the customer within your domain, one of the hard parts about covering Amazon is you do so much. And if you're a generalist analyst I am, I'll go to Amazon, which part of database do you cover? Is it operational or analytic? I'm like, I cover database and data and Infrastructure, the whole gamut. And so as you focus on customers and their needs within your domain, how do you assess feedback from your colleagues that might be further up the stack or adjacent, and how do you bring that together to really nail the holistic customer need? Because you guys have seemingly done a great job of that. Is that because you're so sort of developer-centric, is it because you have the flexibility of those primitives. Does that question, make sense to you? And how do you bring it all together so that it nails the customer need?

Art Baudo

>> I think it's very much about continuing to listen to what customers are doing, talking to and understanding customers. And there are a lot of us that talk to customers and we do try to take into account different customer viewpoints as we deliver different products. But also that sometimes results in the development of different products just to cover different customer needs at the same time. And I think we'll see a continuation of that as we go forward.

John Furrier

>> What's your favorite customer principle? Amazon principle? Bias for Action?

Dave Vellante

>> Favorite LP?

Art Baudo

>> Favorite leadership principle? I say Earn Trust.

John Furrier

>> Earn Trust. Okay. So on the agent piece, I want to get back to this because one of the things I've been watching very carefully is how fast you guys move to meet the need with Bedrock. And SageMaker and getting closer to the primitives or the machine code. I asked Matt Garman last year, all the Anthropic engineers at that time, Perplexity, this is when that was in the heyday, early kind of breakout, they were getting closer to the silicon. So the code was getting closer to that. How has that changed? And SageMaker became that layer and Bedrock became more for the developer, pick your model. But you guys have been very accurate on that. Choice is key. We hear that all the time. But on the developer side, they want to get closer to the processors. They want to get closer to the instances. What are you guys doing for the ecosystem? Because one of the themes here is open source. That's huge. You guys are part— we see you guys at the Linux Foundation all the time, so we know you're active. What tools are you building for these new AI-native developers, if at all, to interact with? You guys think about that software? Piece of it, because the software with the hardware seems to be the big advantage. What's the update? What kind of stuff are you building for the ecosystem from a tooling standpoint for optimizing and tuning?

Art Baudo

>> I have to turn over part of this answer to my other colleagues who are working on that space a little bit more than I am, but I can say that we're working on that tooling and trying to make sure we are working hand in hand with developers to make sure that we meet those needs. As for the tools and other updates that are happening, I'll have to leave it to some of them to offer.

John Furrier

>> My next question, that would be a little more internal to Amazon. So last year we did something new.

Art Baudo

>> Okay.

John Furrier

>> We had what we call a re:Invent halftime report.

Art Baudo

>> Okay.

John Furrier

>> So in June, I went up and sat down with all the leaders because it seemed the first half of 2025 was so fast. We got an update. We didn't have time to go up this year. It was so busy. You guys seem to be super busy. Everyone's cranking. What's been the velocity change of product releases, cadence of the culture? I thought last year was fast. We've got re:Invent coming up in December. Yeah, it seems this year is probably 2 to 3x faster. How would you describe the pace of play in your world and what things do you see at re:Invent coming down the pike?

Art Baudo

>> Well, I would say that this has been an incredible time to be in compute and also in the cloud space. I've seen, this year so far we've introduced a series of the AMD instances early in the first quarter as well. We're seeing even more customers use those instances that I was privileged enough to talk about today. And then we've also added our next generation of Graviton instances have deployed, and I'm excited. re:Invent, as you mentioned, is going to be here before we all know it. And there's going to be even more acceleration of those deliverables. But that's also in direct response to customers accelerating the need for more and more compute as well. And I see that happening as we deliver.

John Furrier

>> We saw— we were talking about the news on our last interview when the cloud leader came on from AMD, Peter, and Google had earnings out. You guys have talked about the CapEx. The forecasts for CapEx keep getting bigger. We're seeing model routing as a big theme when you talk about CPU in relation to the GPU, but workload migration, workload routing, we're starting to see kind of tier 1 workloads might want the best of the best infrastructure. Do you guys see the CapEx growth changing the quality of service in terms of being smarter around, okay, these are great instances for that, let's move that workload there.

Art Baudo

>> So I think we have, as we have added more instances, we've been even more targeted for the workloads that those instances target. So, and that has been as customers have had some specific needs for specific workflows, workloads, excuse me. And one of the things that we had was, we designed some instances around electronic design automation, particularly on the X8aedz side of the house. And that's an example of a workload that has been delivered over there. Not that those instances aren't used for other things, but I see that more workloads have contributed to even more instances, more performance that's needed, and more memory as well.

Dave Vellante

>> Maybe another way of asking that question is, 2020, 2021, remote work, everybody went to the cloud, huge explosion in demand. And then as we entered 2022 and got through the year, cost optimization became a big thing. So AWS came up with a number of tools and was very aggressive actually about cost optimization. I remember there was a lot of talk in the market, oh, they're going to hurt revenue. And, ultimately it showed you did the right thing for your customers and they bought more. And we come into, this era. Do you see a similar sort of discussion going on with customers on sort of token cost optimization? And if so, are there things that you're thinking about, how are you helping them with that?

Art Baudo

>> So I think we spent a lot of time on cost optimization, as you mentioned, and I think we've continued to try to do that for customers. And I mentioned before that price performance number, we've tried to make sure that we've continued to provide the best price performance that we can to customers. That helps as, your workloads are increasing and you're needing more performance and you're sending in more tokens to try to optimize that cost. We've also initiated additional tools for customers. on the Graviton side, we issued a Graviton Savings Dashboard for customers to see. We've issued— we've been trying to help customers get more information about Spot Instances as well so they can take advantage of these cost-saving vehicles as well within their ecosystem to do the best that they possibly can to save money within the infrastructure that we have and also offer them and show them where instances are available and give that information to them so they can use the spot side as well even more.

Dave Vellante

>> All right.

John Furrier

>> Great to see you. We want to catch up with you at re:Invent for sure. I guess my final question for you, I'm just curious because you're at AWS, In 2022, Amazon held a conference called re:MARS.

Art Baudo

>> Yes,

John Furrier

>> re:MARS was one of my favorite events. They only happened for one year. It was really machine learning, automation, robotics, and space. It was really a— even I saw James Hamilton running around and he was just there for fun. They had everything there. AMD showing a lot of robotics, seeing physical AIs here. Is there a perspective on AWS with physical AI being really kind of an edge kind of thing.

Dave Vellante

>> How do—

John Furrier

>> and I know Amazon has a ton of experience with robotics, so it's not new to you guys. How do you see the robotics cloud thing playing out? How do you guys view that? I'm just curious because this is right after agents, physical AI is going to be right around the corner. And we talk to these developers, they want to develop in the cloud, they want to have edge capability. Any thoughts around how you guys think about it? Because a lot of those learnings could be really moved into

Art Baudo

>> Yeah, I'll say

John Furrier

>> that with re:Invent coming up, we will talk about physical AI there.

Art Baudo

>> We will even deliver sessions talking about physical AI. So I want to be clear to people out there that we do see this as incredibly important. And as you mentioned, we do have experience in robotics at Amazon, as well. And we see this being complementary to the overall usage in the cloud here. And I look forward to giving more information about how we do some of that as well.

John Furrier

>> all the engineering you guys have done in AWS, the

Art Baudo

>> learnings.

John Furrier

>> Yeah.All right. You're not going to say anything, but I tried my best.

Dave Vellante

>> Well, but the point is, you have a huge, huge amount of experience with

Art Baudo

>> this.

Dave Vellante

>> Yes. even things that go back to, remember, Saildrone? that's still Saildrone.

John Furrier

>> Yeah. Really? We did an interview with them last year, as an example. And re:MARS was a visionary event. It didn't happen. That was 4 years ago. It didn't happen. But It was such a great collection of research and scientists and commercial. And now we're seeing next year probably going to be a breakout year with robotics, certainly in the US. China's a little bit ahead. But all those hardware, software engineering problems that you guys work on every day to power cloud and AI, they're in the— basically a robot is a supercomputer with all the latency challenges, all the right compute, all the right things have to happen. Safety.

Art Baudo

>> Yeah, it is a sweet spot in combining lots of different elements of compute. Somebody took a picture of one of the robots here today and there's the most— there's a head, we all see it and it looks a cute human, but behind it is a $1,000 camera type thing and memory that's basically behind it. It's a big walking giant computer that's videoing you and processing that information. Computer

John Furrier

>> vision. It's great to have you onWish we could go longer and we'll see you at re:Invent. Thanks for coming on.

Art Baudo

>> Absolutely.

John Furrier

>> All right.

Art Baudo

>> Thank you.Thanks for having me in the house.

John Furrier

>> Again, the Amazon Web Services perspective, continue to see the massive growth. It's a rising tide. Again, tons of CapEx build out the demand for AI infrastructure, the power in this next generation of computing and AI is all being paved by the leaders here. We're doing our part to bring you the content. I'm John Furrier with Dave Vellante. Stay tuned for our next segment after this guest leaves. Thanks.

Art Baudo, AWS

search

John Furrier

>> Welcome back, everyone, to theCUBE's live coverage here in San Francisco of AMD's Advancing AI event. All the leaders are here in the industry, AMD, and also practitioners. It's a free event, so it's a packed house, completely sold out. I'm John Furrier with Dave Vellante, my co-host. We're trying to bring all the AI knowledge to you. We got a great guest, CUBE alumni Art Baudo, Principal Product Marketing Manager, head of EC2 product marketing at AWS, Amazon Web Services. Art, great to see you again. Thanks for coming back on in person.

Art Baudo

>> Yeah, it's nice to see you, John and Dave. It's great to be here.

John Furrier

>> It's very rare that we get the two of us in one spot, but this is an epic event, pun

Art Baudo

>> intended.I love it.

John Furrier

>> So obviously we've been following AWS. Everyone knows the history. We actually started with AWS in 2007 when I first started using it. They didn't have domains at the time. I used to joke with, Matt Garman and Dave Brown, that it was a long EC2 string. Then they had domains. But really with theCUBE since 2013 and cloud made the market, cloud native, cloud, AWS made SaaS. Now we're in another era where SaaS is moving to projects. AI is now involved. You can bring in all kinds of artifacts into the AI, into the compute, email it, can be scanned, Slack messages. You got tools developing Q and Kiro you guys built. You got all kinds of really kind of forward deployed engineering. So it's really a renaissance that feels early days of AWS in the AI realm because the interface is changing. All this new value is coming in, but it's built on cloud.

Art Baudo

>> Yeah.

John Furrier

>> So you guys have been doing well. You did a debt financing. I saw Andy Jassy announce that you're spending a lot of CapEx, you're building out more and more horsepower. So give us a state of the compute landscape because it seems to be with agents coming in, AgentCore a year ago, I saw Swami do an announcement on that yesterday, one year anniversary, the growth. Give us a perspective inside AWS, what's it like?

Art Baudo

>> Well, first, one of the things I think you bring up 2007, but the interesting thing to me is, well, this is our 20th year of EC2 and also the 20th year of S3. And I think those two anniversaries this year mark kind of some pretty major milestones in terms of where we have seen customer growth and customers change over that period of time. we've added thousands of instances now, instance types, and we see customers on the AI side expanding what they're using. We offer a broad depth of AI platforms that you can use, but also a lot of customers are also using, with the additional performance on the CPU side, we're seeing customers do agentic, AI on the CPU as well. We did some demos as well here showing how people can send tokens into the CPU as well. So now with AI taking off, we're seeing people use it in multiple different spaces across the industry and across the instance types as well. And I've never seen so many things start to move into the cloud, things that I think we hadn't thought originally would move or other people hadn't thought initially would move over to the cloud. today I was with Marvell giving an example how they're doing electronic design automation for silicon development. They're moving that into the cloud and how much they've been able to expand. So because of the resources of the cloud, which is to me very exciting.

John Furrier

>> we've been covering the compute side of EC2. We're here at the AMD event. Just talk about the transition of AMD as a company because they have been a supplier to AWS. What we're hearing here today is a systems architecture shift for the company. Obviously, their stock's doing great because there's a huge demand curve for infrastructure, but it seems to be a different game for the semiconductor companies because they've got to build out systems because there's a lot of cloud and on-prem and edge growth. What's your relationship with AMD now? And how has it changed?

Art Baudo

>> Just so I don't— I think that, when we started, we introduced AMD EPYC processors back in 2018, and we've introduced every generation of EPYC processor to date with Turin. Some of our Turin instances actually just launched in 2026. So very prescient to that discussion. But when we first started, the EPYC instances were using multithreading. And once we moved to the 7th generation, we moved to single-threaded instances, with the performance going from 6th generation to 7th generation, that is some pretty good improvement in performance, raw performance that it's been able to deliver. And customers in the AI space in particular and other spaces in the EDA space are able to take advantage of those performance improvements. We also have high frequency instances as well that are available on the AMD platform. Our X8aedz and our M8azn ones have the high 5 GHz frequency as well. So that provides performance frequency, memory as well. So we have been expanding that overall portfolio since we started that introduction in 2018.

John Furrier

>> What use cases are emerging that you could share? Because a lot of people have reined in cloud spend. Now they've got tokenomics. They're starting to see real value being created by deploying on cloud, some really great use cases. How has that changed what they do and how has that changed the price performance opportunities? what's the pattern recognition?

Art Baudo

>> My take is that as we've added more performance, we've tried to continuously make sure we optimize on price performance for customers. So some of our most recent instances continue to see a reduction in price performance. That is a continuous focus from the AWS side, and that is no different on the AMD instances. As that performance has gone up, so has the price performance that we've delivered to customers. We're seeing— we continue to see a wide variety of workloads that are happening. But one of the things I also mentioned, we're seeing a need to do more simulation for customers, that those simulations that you can do in the cloud— Rivian today was talking about how they're simulating vehicle crashes using the cloud. They need the breadth of the CPUs, the scaling of the CPUs in the cloud to be able to do that. They see the same on the EDA side as the complexity of design is increasing. There is a continuous need for continued improvement in performance that exists in the cloud.

Dave Vellante

>> I want to ask you, you mentioned you started adopting EPYC in 2017. I think that's the same year you launched Nitro, roughly, I think.

Art Baudo

>> Yeah. So we adopted EPYC in 2018 and we launched Nitro in 2017.

Dave Vellante

>> Okay, there you go.

Art Baudo

>> But obviously all of the EPYC instances are off of Nitro.

Dave Vellante

>> Right, so I wanted you to talk about the value, customer value of Nitro. We wrote a piece, our research team wrote a piece years ago on AWS's secret weapon.

John Furrier

>> Dave loves Nitro.

Dave Vellante

>> Well, Nitro was like, yeah, I love Nitro too. It really was, is AWS's secret weapon. But at any rate, the customer value was it gave you the ability to provide optionality and security that you can't go in and see, and enclaves and the like. So I wonder if you could discuss that, both the optionality and the security angle.

Art Baudo

>> Yeah, so first on the optionality, the Nitro System has allowed us, we believe, we credit the Nitro System with being able to deploy so many instances for customers. It has allowed us to faster deploy the instances that are out there. And we believe that has also contributed to performance as well for customers. The Nitro System has allowed us to offload that, some of the functions over into hardware and software. Our hypervisor has been offloaded over there onto the Nitro System. And also with the Nitro System, we have zero operator access, something that over time we've also worked to showcase even more to customers. We did in 2023, we worked with the NCC Group to have our architecture analyzed and the NCC Group Talked about how our architecture is designed so that no one can access that architecture. We've introduced attestation for software-based instances as well. So security as well, that's been in there. So it's been something that evolution has paired with our evolution overall desire to deliver security and performance at scale to customers.

Dave Vellante

>> So how does it fit vis-à-vis AI? What are the, are there trade-offs, or is it just a sort of a natural extension into AI? How are you thinking about that?

Art Baudo

>> So I see it as a natural extension, frankly. So customers are, with the additional performance that the CPUs have been getting, we're seeing customers use the CPUs for AI. The Nitro System is something that's deployed across our entire ecosystem. So it's available for customers. I know we're here at an AMD event, But the Nitro System is deployed across our entire portfolio, from our CPU portfolio to the GPU portfolio and to our own Trainium and Inferentia and Graviton-based products as well.

John Furrier

>> On migration, this comes up a lot. We're hearing a lot of people trying to migrate in the past year or so, maybe 18 months to start here. The— now we're hearing with AI, migration is going faster. You guys have published a lot of examples of migrations for this, that platform, but When you get down to migrating to infrastructure, what's that look like? What are people doing or is that abstracted away completely? If I want to move from Graviton to AMD or how do I mix and match the best elements in the design?

Art Baudo

>> One of the things we try to do is give customers choice. Customers are able to make that decision and migration from an instance type to another, we try to make as seamless as possible. Specifically, if you're going from, 7th generation to 8th generation instances, the architecture is not changing. If you're sticking, let's say, with AMD instances or a Graviton-based instance, so you're able to migrate very quickly, achieve the performance benefits. And if you're someone that needs both that performance and, you're looking for the price performance to decline, that upgrade affords you that ability to do that. And we're seeing customers make that trade-off.

Dave Vellante

>> We were talking earlier, I'd love your opinion on this, Art. About, the old, the Gartner term of the split IT.

John Furrier

>> Bimodal.

Dave Vellante

>> Bimodal IT, remember that? It was old stuff here, new stuff here. You've got a massive install base of applications running general purpose computing, x86, slash, infrastructure. How do you see those two worlds, the AI and the general purpose world, Are they coming together in a new fused architecture? Will they be bimodal? How do you see that playing out?

Art Baudo

>> I see them as complementary. They can exist in the same space. And I think that depending upon the industry also depends upon how they're using some of those AI skill sets. As I mentioned before, with, folks such as Rivian doing and now exploring more AI to help with their simulation and modeling, they're pairing that with some of their knowledge and their existing software and their existing databases. And I see that to be something that customers are doing. They're adding this AI as an additional layer to support some of the fundamentals that they have that exist right now.

John Furrier

>> Has Agentic changed your view on EC2, all the different elements on the infrastructure side? How has that changed, if at all, how you guys operate?

Art Baudo

>> I wouldn't say that it has changed how we operate. From the perspective of, Amazon is one of the most customer-centric companies that's out there. So we listen to what customers are looking for, and that's how these instances come about. And I would say that it's only made me actually more optimistic about EC2. The reasons I think to go to the cloud and use and leverage the performance that exists are even greater now today. And I think the fundamentals of the cloud, the ability to deliver security, the ability to deliver performance on demand for customers, especially in some of these workloads where you have a really, really heavy burst and a need for lots of CPU performance, and then suddenly you may not need it for a period of time. The cloud is phenomenal at doing that.

Dave Vellante

>> I want to ask you about that customer obsession. John, you and I have asked, Andy Jassy this question years ago about why the focus on primitives, and his answer was always pretty consistent, because it allows us to change quickly as the market changes. And my question is, as you focus on the customer within your domain, one of the hard parts about covering Amazon is you do so much. And if you're a generalist analyst I am, I'll go to Amazon, which part of database do you cover? Is it operational or analytic? I'm like, I cover database and data and Infrastructure, the whole gamut. And so as you focus on customers and their needs within your domain, how do you assess feedback from your colleagues that might be further up the stack or adjacent, and how do you bring that together to really nail the holistic customer need? Because you guys have seemingly done a great job of that. Is that because you're so sort of developer-centric, is it because you have the flexibility of those primitives. Does that question, make sense to you? And how do you bring it all together so that it nails the customer need?

Art Baudo

>> I think it's very much about continuing to listen to what customers are doing, talking to and understanding customers. And there are a lot of us that talk to customers and we do try to take into account different customer viewpoints as we deliver different products. But also that sometimes results in the development of different products just to cover different customer needs at the same time. And I think we'll see a continuation of that as we go forward.

John Furrier

>> What's your favorite customer principle? Amazon principle? Bias for Action?

Dave Vellante

>> Favorite LP?

Art Baudo

>> Favorite leadership principle? I say Earn Trust.

John Furrier

>> Earn Trust. Okay. So on the agent piece, I want to get back to this because one of the things I've been watching very carefully is how fast you guys move to meet the need with Bedrock. And SageMaker and getting closer to the primitives or the machine code. I asked Matt Garman last year, all the Anthropic engineers at that time, Perplexity, this is when that was in the heyday, early kind of breakout, they were getting closer to the silicon. So the code was getting closer to that. How has that changed? And SageMaker became that layer and Bedrock became more for the developer, pick your model. But you guys have been very accurate on that. Choice is key. We hear that all the time. But on the developer side, they want to get closer to the processors. They want to get closer to the instances. What are you guys doing for the ecosystem? Because one of the themes here is open source. That's huge. You guys are part— we see you guys at the Linux Foundation all the time, so we know you're active. What tools are you building for these new AI-native developers, if at all, to interact with? You guys think about that software? Piece of it, because the software with the hardware seems to be the big advantage. What's the update? What kind of stuff are you building for the ecosystem from a tooling standpoint for optimizing and tuning?

Art Baudo

>> I have to turn over part of this answer to my other colleagues who are working on that space a little bit more than I am, but I can say that we're working on that tooling and trying to make sure we are working hand in hand with developers to make sure that we meet those needs. As for the tools and other updates that are happening, I'll have to leave it to some of them to offer.

John Furrier

>> My next question, that would be a little more internal to Amazon. So last year we did something new.

Art Baudo

>> Okay.

John Furrier

>> We had what we call a re:Invent halftime report.

Art Baudo

>> Okay.

John Furrier

>> So in June, I went up and sat down with all the leaders because it seemed the first half of 2025 was so fast. We got an update. We didn't have time to go up this year. It was so busy. You guys seem to be super busy. Everyone's cranking. What's been the velocity change of product releases, cadence of the culture? I thought last year was fast. We've got re:Invent coming up in December. Yeah, it seems this year is probably 2 to 3x faster. How would you describe the pace of play in your world and what things do you see at re:Invent coming down the pike?

Art Baudo

>> Well, I would say that this has been an incredible time to be in compute and also in the cloud space. I've seen, this year so far we've introduced a series of the AMD instances early in the first quarter as well. We're seeing even more customers use those instances that I was privileged enough to talk about today. And then we've also added our next generation of Graviton instances have deployed, and I'm excited. re:Invent, as you mentioned, is going to be here before we all know it. And there's going to be even more acceleration of those deliverables. But that's also in direct response to customers accelerating the need for more and more compute as well. And I see that happening as we deliver.

John Furrier

>> We saw— we were talking about the news on our last interview when the cloud leader came on from AMD, Peter, and Google had earnings out. You guys have talked about the CapEx. The forecasts for CapEx keep getting bigger. We're seeing model routing as a big theme when you talk about CPU in relation to the GPU, but workload migration, workload routing, we're starting to see kind of tier 1 workloads might want the best of the best infrastructure. Do you guys see the CapEx growth changing the quality of service in terms of being smarter around, okay, these are great instances for that, let's move that workload there.

Art Baudo

>> So I think we have, as we have added more instances, we've been even more targeted for the workloads that those instances target. So, and that has been as customers have had some specific needs for specific workflows, workloads, excuse me. And one of the things that we had was, we designed some instances around electronic design automation, particularly on the X8aedz side of the house. And that's an example of a workload that has been delivered over there. Not that those instances aren't used for other things, but I see that more workloads have contributed to even more instances, more performance that's needed, and more memory as well.

Dave Vellante

>> Maybe another way of asking that question is, 2020, 2021, remote work, everybody went to the cloud, huge explosion in demand. And then as we entered 2022 and got through the year, cost optimization became a big thing. So AWS came up with a number of tools and was very aggressive actually about cost optimization. I remember there was a lot of talk in the market, oh, they're going to hurt revenue. And, ultimately it showed you did the right thing for your customers and they bought more. And we come into, this era. Do you see a similar sort of discussion going on with customers on sort of token cost optimization? And if so, are there things that you're thinking about, how are you helping them with that?

Art Baudo

>> So I think we spent a lot of time on cost optimization, as you mentioned, and I think we've continued to try to do that for customers. And I mentioned before that price performance number, we've tried to make sure that we've continued to provide the best price performance that we can to customers. That helps as, your workloads are increasing and you're needing more performance and you're sending in more tokens to try to optimize that cost. We've also initiated additional tools for customers. on the Graviton side, we issued a Graviton Savings Dashboard for customers to see. We've issued— we've been trying to help customers get more information about Spot Instances as well so they can take advantage of these cost-saving vehicles as well within their ecosystem to do the best that they possibly can to save money within the infrastructure that we have and also offer them and show them where instances are available and give that information to them so they can use the spot side as well even more.

Dave Vellante

>> All right.

John Furrier

>> Great to see you. We want to catch up with you at re:Invent for sure. I guess my final question for you, I'm just curious because you're at AWS, In 2022, Amazon held a conference called re:MARS.

Art Baudo

>> Yes,

John Furrier

>> re:MARS was one of my favorite events. They only happened for one year. It was really machine learning, automation, robotics, and space. It was really a— even I saw James Hamilton running around and he was just there for fun. They had everything there. AMD showing a lot of robotics, seeing physical AIs here. Is there a perspective on AWS with physical AI being really kind of an edge kind of thing.

Dave Vellante

>> How do—

John Furrier

>> and I know Amazon has a ton of experience with robotics, so it's not new to you guys. How do you see the robotics cloud thing playing out? How do you guys view that? I'm just curious because this is right after agents, physical AI is going to be right around the corner. And we talk to these developers, they want to develop in the cloud, they want to have edge capability. Any thoughts around how you guys think about it? Because a lot of those learnings could be really moved into

Art Baudo

>> Yeah, I'll say

John Furrier

>> that with re:Invent coming up, we will talk about physical AI there.

Art Baudo

>> We will even deliver sessions talking about physical AI. So I want to be clear to people out there that we do see this as incredibly important. And as you mentioned, we do have experience in robotics at Amazon, as well. And we see this being complementary to the overall usage in the cloud here. And I look forward to giving more information about how we do some of that as well.

John Furrier

>> all the engineering you guys have done in AWS, the

Art Baudo

>> learnings.

John Furrier

>> Yeah.All right. You're not going to say anything, but I tried my best.

Dave Vellante

>> Well, but the point is, you have a huge, huge amount of experience with

Art Baudo

>> this.

Dave Vellante

>> Yes. even things that go back to, remember, Saildrone? that's still Saildrone.

John Furrier

>> Yeah. Really? We did an interview with them last year, as an example. And re:MARS was a visionary event. It didn't happen. That was 4 years ago. It didn't happen. But It was such a great collection of research and scientists and commercial. And now we're seeing next year probably going to be a breakout year with robotics, certainly in the US. China's a little bit ahead. But all those hardware, software engineering problems that you guys work on every day to power cloud and AI, they're in the— basically a robot is a supercomputer with all the latency challenges, all the right compute, all the right things have to happen. Safety.

Art Baudo

>> Yeah, it is a sweet spot in combining lots of different elements of compute. Somebody took a picture of one of the robots here today and there's the most— there's a head, we all see it and it looks a cute human, but behind it is a $1,000 camera type thing and memory that's basically behind it. It's a big walking giant computer that's videoing you and processing that information. Computer

John Furrier

>> vision. It's great to have you onWish we could go longer and we'll see you at re:Invent. Thanks for coming on.

Art Baudo

>> Absolutely.

John Furrier

>> All right.

Art Baudo

>> Thank you.Thanks for having me in the house.

John Furrier

>> Again, the Amazon Web Services perspective, continue to see the massive growth. It's a rising tide. Again, tons of CapEx build out the demand for AI infrastructure, the power in this next generation of computing and AI is all being paved by the leaders here. We're doing our part to bring you the content. I'm John Furrier with Dave Vellante. Stay tuned for our next segment after this guest leaves. Thanks.