Cloud Infrastructure for the AI Age

What are the new building requirements for AI infrastructure to power unmatched scale, stability, and security for the AI era? 

‍

At the Six Five Summit, we're thrilled to present our Cloud Infrastructure Track Opener, Amazon Web Services’ Vice President, Compute & ML Services, Dave Brown!

‍

Host Patrick Moorhead and Dave Brown share a conversation on AWS's role in building the future of artificial intelligence infrastructure. AWS is at the cutting edge, offering an AI stack that is designed to meet the demands of AI workloads, ensuring their customers have the scale, stability, and security needed to foster innovation. 

‍

Key takeaways include:

‍

🔹Meeting Evolving AI Workload Demands: Explore AWS's proactive approach to addressing the ever-increasing and complex demands of modern AI workloads, ensuring customers have the foundational capabilities they need.

‍

🔹AWS's Comprehensive AI Stack: Get an in-depth overview of AWS's robust AI stack, including advanced accelerated compute solutions powered by NVIDIA GPUs and AWS's proprietary AI chips.

‍

🔹Scale, Stability, and Security for AI: Understand the paramount importance of scale, stability, and stringent security in effectively running AI workloads, and how AWS delivers these critical attributes.

‍

🔹Continuous Reinvention in the AI Space: Gain forward-looking insights into future AI trends and how AWS is continuously reinventing its services and technologies to stay at the forefront of AI innovation.

‍

Learn more at Amazon Web Services.

‍

Watch full video on Six Five Media, and be sure to subscribe to our YouTube channel, so you never miss an episode.

‍

Or listen to the audio here:

‍

‍
Patrick Moorhead: The Six Five Summit is back in its sixth year, and we are talking about, unsurprisingly, AI. AI has moved from science projects to POCs to full-scale. Infrastructure and platforms around AI are important, companies focusing on what they do best and leaving the driving to other people. Many are making this choice, and I can't imagine a better guy to talk about this than Dave Brown with AWS. Dave, welcome back to the show.

Dave Brown: Well, Pat, thanks for having me. It's great to be back. I think I've done these a few times before, but six years. That's amazing.

‍

Patrick Moorhead: You have. You're a friend of the Six Five, and we appreciate everything that you bring to the table. I know you don't do a lot of these, so it's been very, very special. AI has just been incredible, and it's interesting. AI didn't start two years ago. In fact, you got the first jump on this with machine learning about seven or eight years ago and ended up being the leading provider for those workloads. Things have changed, though. You're putting significant investments all the way from NVIDIA GPUs to your own silicon. So can you talk to us, maybe do a double-click, maybe even talk about the strategy of how you're not just delivering the raw performance because there's a lot of people that can do that inside of the chip itself, but also the usable compute power with the stability that you provide, the security and also the scale. As I said, enterprises are scaling AI.

‍

Dave Brown: Yeah, absolutely. Well, Pat, it's been an incredible journey. I was... I think it's 14 years ago now that we put the first NVIDIA GPU in the cloud, and 14 years ago was actually a negotiation to convince NVIDIA that it was a good idea to put one of their GPUs in the cloud. And boy, how things have actually changed. It's pretty incredible. And that's been amazing. We've had the Hopper GPUs. We just went live last week with the latest Blackwell, the B200 GPU, so this very close relationship we have with NVIDIA is just continuing to grow, and it's been a lot of... very, very successful. At the same time, though, we know that one of the most important things for our customers is to find ways to innovate on their behalf. And specifically, in this space, we've seen that if we can find ways to reduce the cost of machine learning and to give them more performance for every dollar spent is how we think about it, customers typically do more. And we started to see that in 2017, very early days of machine learning, and folks were starting to think about deep learning. This is long before generative AI, but we saw that the inference workload, we could unlock more innovation for customers if we could find a way to do 

more.

‍

And as you know, there's nothing that stops AWS from innovating. We've been at the hardware and custom silicon level with our processor on Graviton, and we thought we could do the same thing in the AI space with Inferentia and now Trainium. And so, as you say, we are building our own custom silicon. We are trying to find ways to give customers that better performance for every dollar spent or price performance while supporting customers on NVIDIA. And that'll always be the path long-term. Customers need choice when they come to the cloud, and we're very happy to give it to them. On the specifics, though, if you think about you can go anywhere, and a lot of cloud providers provide you with NVIDIA GPUs. So, what are the things that we at AWS think really differentiate ourselves? And one of them is stability. And so when you're running a very large cluster, any sort of node failure, any sort of networking issue causes you to lose time. And these clusters aren't cheap, so any lost time leads to lower utilization, which leads to increased costs and just not using those resources. So we have teams of people at AWS working on stability, and we design our servers differently from anybody else with our Nitro system, which we've spoken about before, where it's custom silicon that's actually running all of the management layers and networking layers and StorageIO to really give you the best performing and most stable server available on the market. And we hear from our customers all the time that AWS is the place to be if you want that stability. We always say security is our priority zero, and we live it every single week at AWS.

‍

You know, I attended a number of meetings in our proactive, I'd say paranoid security, and we want to bring the same level of security we've given to customers with our GP... with our CPUs and our normal compute to the GPU space as well. So that starts with Nitro. We have zero operator access. We've actually gone as far as putting in our terms and conditions that we have no access to customer data. We've had folks like the NCC Group in the UK validate that AWS has no access to customer data, and we just think about it very, very differently.

We also working on a product called Secure AI... Secure AI, which we've announced, which actually brings that level of security all the way into the GPU or into the Trainium process to ensure that as a cloud provider, we have no access to model weights and we also have no access to customer data. And that really resonates with our customers from the smallest startups to the largest enterprises where they say we want to be on AWS because of the level security they provide.

‍

And then, finally, scale. One thing about this space is the scale is, I would say, larger than we've ever seen before, whether it comes to power data centers, network processes, GPUs, whatever it might be. And we have a long history of scaling in this way and doing many, many deployments at really large scale. One of the places we've rarely invested in is our network with high elastic fabric adapter, where we are able to provide the level of latency, low latency, and high throughput with very, very consistent in the cloud. And so, for very large training clusters, we call them ultra-clusters, we're able to give customers that. We're actually building a cluster at the moment with Trainium. Actually, on both NVIDIA, we're building in Project Saber for NVIDIA, which is 20,000 Blackwell GPUs. And then we're doing many hundreds of thousands of Trainium accelerators for Anthropic for their next training cluster, which we call Project Rainier. So being able to scale once customers start small and go large is something we really specialize in.

‍

Patrick Moorhead: Yeah, you specialize in it. You also have a long track record for it. I have talked to some end users who just weren't finding the stability that they needed with the Neo Cloud. And the entire system matters when it comes to this. Cooling networking, the supply chain, the suppliers you pick, the levels of quality that you put into that. And some people have to try the stuff before they realize that, but it's pretty known fact down here on your reliability, so it's pretty good to see. You talked a little bit about optionality, and we've talked on the show about your optionality with silicon, right, your own silicon to merchant players like NVIDIA, and even folks like AMD, but this also goes across the way that you consume AI. If I look at your entire stack, you can go all the way from, I'll call it, piece parts to fully managed solutions. Can you talk a little bit about the thinking that went into? Is it as simple as, "Hey, we're AWS. We provide optionality, and these are the different ways that people want to consume from us, so we deliver that?"

‍

Dave Brown: Yeah, it is that we do talk about choice a lot, right. We like the fact that when customers come to AWS, whether it comes to custom silicon or whether they're going to run their own service, be it a database or training cluster or whether they're going to use one of our managed services at some layer, and we have different layers of those as well, they get to choose what's right for their business. And I think that is so important for customers. And you see customers making a choice when they start. They may change their choice over time, but they have the freedom to move around. And that is so important from a cloud provider. It drives competition as well, which we think is excellent in the markets. But one of the things... when customers are looking to run their own training clusters, one of the hardest things about running a training cluster is actually the distributed systems side of it. And it's not something you would normally think about, but just keeping it stable, making sure you can recover quickly from any sort of GPU accelerator failure or is hardware failure is so, so important. So one of the services, we provide a number of them. SageMaker is a great example of a service that has job-level understanding of what you're trying to do.

‍

And we have a new feature there called HyperPod, which we launched a few years ago, that really makes running a training cluster very, very easy. A layer below that, we have our Elastic Kubernetes Service, EKS. Kubernetes has really sort of come to the foreground as the underlying architecture that most of the model providers are using for training, and we have a service there that we're also spending a time optimizing that to ensure that any sort of node failure and also cluster size, right. Kubernetes is not going to be able to naturally scale to the size of clusters that the provider's going to need. So we've invested a lot in scaling Kubernetes as well and making sure we can support hundreds of thousands of nodes in the future. And then things like ParallelCluster as well that provides just better management of things like Slurm and whatever you might want to use. So there's this at various different layers customers can choose.

‍

Now we've seen customers like Perplexity that have actually chosen SageMaker HyperPod. That was the right choice for them, and they liked what it gave them and liked how they managed the job, and they were able to make... most important thing is obviously make progress in training their models, make progress in fine-tuning whatever they would do. The customers like Adobe, they've chosen EKS, where they've said, "Hey, we want to be at that layer below." For us as an engineering team, that's the layer we want to be at, and we really... For us, it's a customer choice, and we'll support them at whatever layer they want to be at. So, really, across the portfolio, I'm sure we'll talk more about Bedrock and inference for customers that say, "Hey, I don't want to manage anything. I just want to get the value of AI in my application." Bedrock is great at giving them the inference capabilities that they need.

‍

Patrick Moorhead: Are there core variables that customers you've seen should consider with knowing how to make these decisions? Sophistication might be one, the amount of resources might be another. Is there a cheat sheet for this, Dave?

‍

Dave Brown: Yeah, I think it's so true in the space. One of the things Andy Jassy's often said is, "There's no compression algorithm for experience," which is such a great thing because it's so geeky as well. But I think a lot of it is just customers learning. I think customers and it's what is your core value? And I think a lot of engineering teams, I see, they do want to start layer... lower down the stack, but very quickly, they start to realize that a lot of their time is actually spent managing underlying GPUs or managing the container ecosystem or dealing with failures.

And normally, what we see is they'll start to move up the stack if they aren't already there. So SageMaker HyperPods is being just so incredibly successful. We use it internally as well with our Nova team that's doing training, so they're a big part of that feedback loop with a very, very large frontier-size cluster as well. So a customer that's just looking to get that you want to move quickly, you want to get the value out of AI, you don't want to spend all your time worrying about a node failure, and how you recover, SageMaker HyperPods is really where you want to be.

‍

Patrick Moorhead: Yeah. No, I love that. Different strokes for different folks. Definitely impressed. I mean, I'm impressed with a lot of the stack. Bedrock, to me, is the farthest that she went, essentially saying, "Okay, we're not really great at the tech, but we know what we want to do with it. Make it easy for us. We don't have to pick all the different factors for everything below it." So that, to me, was a step-up for you when you introduce it that I thought was unique and valuable.

‍

Dave Brown: Yeah.

‍

Patrick Moorhead: So kind of joke that, okay, AI didn't start two years ago. Okay. It's been going on for a long time. But if I do look at my top 10 enterprise challenges with adopting AI, some of them have gone down in the list, some of them have gone up in the list and some of them have just stayed constant. I mean, cost has been a stated challenge, and fear of enterprises diving into that. I mean, they just have to look at the price for a GPU as example, and they're doing the multiplication on their own, and it pretty much scares them. What are some things that you are doing to maximize the value of these enterprise AI investments? Things to maybe allay their fears that, hey, once they get in, it's not going to go crazy.

‍

Dave Brown: Yeah. I mean, I think the biggest impact that we can have for generative AI today is to lower the cost for enterprise customers and startups alike. I think there are many, many proof of concepts that don't find their way to production today because they just can't justify the cost, right. The cost of doing the inference or training the model exceeds the savings. And I strongly believe that in the months and years ahead, many of those POCs will be deployed as costs come down. Now, those costs are going to come down in several ways. So one of the ways is model innovation. We saw that with some of the things we... recent model innovations with the DeepSeek thing that surprised the world. Honestly, it shouldn't surprise the world. I hope we see many more of those. I think the investment that's happening in models, including what we're doing with our own Nova, are going to bring cost improvements for customers in significant ways. And so that's one of them. The other way is going to be through hardware innovation. And so it is incredibly important that we have more competition in the market, more solutions in the market, more options available, more innovation in the hardware that's going to allow us to... We always talk about price performance, and for every performance... for every dollar you spend, how much performance do you get?

‍

And it's a little bit like the re-imagining of Moore's Law. Nobody's made any statement about how it's going to improve, but I think that same thing's going to play out is the cost of inference from a hardware point of view or training has to come down. It just has to come down. It has to come down significantly. Our big investment there is obviously what we've been doing with Trainium. We're on our second generation of that now, which you know today offers up to 30 to 40% better price performance. And for us, that sort of 40% price performance number has been so important for the growth of things like Graviton. We just announced at re: Invent that with Graviton 2, we now land more Graviton 2s every year than all of our other processes combined in our data center, which is an incredible statistic and really speaks to the ability for us to go and innovate a custom silicon and give customers better cost and cost just means they do more and they do more innovation.

‍

And so we're hoping that the same thing happens with Trainium over time as we're able to bring that to market. One of the other things is just being more efficient with the GPUs that you do have. And so we're talking about SageMaker HyperPod we've seen customers save up to 40%. Again, I said magic 40% number using their Hyperpods just because they're driving better utilization, the cost of a large cluster, even a single percentage point in utilization can save you millions of dollars. And so those are the things that you really want to think through. So how do we make sure we have the latest models available through Bedrock so you get those benefits? How do we make sure SageMaker makes you more efficient? And then how are we building custom silicon that's going to drive that cost down? But we know when we lower costs, customers innovate more, and that's what we really want to be able to see in this space.

‍

Patrick Moorhead: Yeah. Listen, I mean, going all the way back to S3 and the price reductions on that, right. I mean, people just really gravitated. I know that that was the start of AWS and then EC2, and the rest was history. I do like the way that you expanded almost the definition of cost. It's not just this lower price, but it's being more efficient at what you're already paying for.

‍

Dave Brown: Yeah.

‍

Patrick Moorhead: There's a rule of thumb that in a standard configuration, the GPUs are sitting around 30% of the time, right. That's not good. The other part of it is on-prem versus AWS. You're sharing these services across as opposed to the sunk cost that you're paying for, whether they're doing something for you or they're not doing something for you.

And my final comment is I've been tracking Graviton forever. I feel like I was part of at least the group of analysts who are trying to educate enterprises and other companies about them, and that really is the poster child for how to do this and also do it again with the option of, "Okay, if you want Intel and AMD? Hey, we got that too. You want Graviton for this? It's good." And one thing-

‍

Dave Brown: Good.

‍

Patrick Moorhead: ... I did on that too is you didn't oversell it, right.

‍

Dave Brown: Yeah.

‍

Patrick Moorhead: You said exactly what the different generations of Graviton did well, and then you added to those workloads as you rolled it out there. And I can't tell you how much trust that that got you in the industry and with your customers, and it's pretty phenomenal.

‍

Dave Brown: Yeah, it's been an incredible success story, and we are very careful with benchmarking as well to ensure that we really... we use real-world workloads to try and benchmark, and we want to make sure if we say something, customers are likely to see that.

And that's been a lot of the success behind Graviton is it's relatively easy to get to the numbers we've quoted, and then it sort of snowballs as one team in an organization sees this big one, another team wants to do the same thing. And so that's really what's driven the adoption. There is one other thing as well, which I didn't mention in the AI space, that's been interesting, and that's just, you reminded me of it when you were talking about on-prem versus the cloud, right.

So the sunk cost to spend upfront to get the GPUs versus being able to get them in an on-demand-like fashion. Now, we haven't been able to maintain on-demand for GPUs. Is been... That's been one of the ways we've sold in the cloud in the history of AWS, and it just didn't work when the Hoppers came out. And really, because if I had a GPU on-demand, it was there for maybe one second, and it was gone. I never saw it again. And so I realized I actually couldn't maintain on-demand.

‍

And so we came up with a new construct called Capacity Blocks, and it's something that we're seeing a lot of our customers, including very large enterprises, use. You think of it as sort of hotel room bookings. You want to book a hotel room. You want a cluster of a certain size, maybe for a few hours or a few weeks. I think we can do it now up to six months. We can give you the capacity immediately if it's available, or you could book it a few days out if you knew you were going to be ready with it. And so the great thing about that is you really only need the cluster for the time that you want it, and then you give it back to us. And so it's a sort of new form of on-demand that we've actually had to innovate around to ensure that we've allowed customers to get access to that. And what we've seen is customers then share these GPUs between different businesses. And so you get that next level of cost saving, the next level of utilization improvement by using the cloud as well. So we're excited about where that's been going.

‍

Patrick Moorhead: Yeah, I appreciate you bringing that up. So, hey, I want to talk about inference. It's kind of funny, Dave. People are all surprised that inference is growing, and I mean, training's still going, but inference is just rocking. I mean, you and I had a conversation, I'm pretty sure five years ago when we went from 80/20 training and inference to 80/20 inference to training. And I'm curious at how are you thinking about this today. I mean, things like reasoning models that make a difference on inference as well?

‍

Dave Brown: Yeah, it is so true. Actually, I remember that conversation with you. In 2017, when we decided to build Inferentia, which is our inference chip, the reason we chose inference is inference was actually 90% of the spend. And so, 10% was training, 90% was inference. And that was like, remember when your phone could show you pictures of dogs and we were all surprised at how did ML do that? That's incredible. And things like BERT and ResNet and those other models and very little training and a lot of inference. And then generative AI changed the whole thing because, the big frontier model providers, there was just so much being spent on training, and the inference workload hadn't caught up yet. And what we're seeing right now is that inference workload is actually catching up. And I wouldn't be surprised if we got to a 80/20, 90/10 again in terms of inference versus training because it's just for it to give its value, really, inference is where the value is. And so Bedrock, as you mentioned earlier, the space that we've been... the service we built specifically for inference. And so it is a serverless service. It means you don't have to run any infrastructure yourself. 

‍

Bedrock runs all the infrastructure and manages the GPUs or the Trainium accelerators behind the scenes manages the complexities of running the model, manages the performance, whether it's output tokens per second, whether it's latency, all geographical locations and regions around the world, all that is managed for you by Bedrock. And you, as a developer or customer, just as to call the API and you get access to the latest models. Obviously, Code 4 is available on Bedrock along with Llama 4 and all of the other models, Nova our own model as well. And so you get to play around and pick these models. That's the other thing, Pat, is we, in most applications, they don't use a single model. They normally use either different variants from a single model or, typically, they'll use many different models for different use cases as well. So that choice on that selection is so important, and we're just seeing great growth on inference. And I think the vast majority of that is in Bedrock today, and we're expecting that to continue to grow and get back to sort of numbers we've seen previously.

‍

Patrick Moorhead: Dave, has reasoning and agents changed core infrastructure at all? I remember, back in the day, a tiny card that didn't require any fans. You could do object recognition, you could do speech-to-text, you could do some basic types of things. And now we have reasoning that seems to take multi-turn agents to take a tremendous amount of compute.

‍

Dave Brown: Yeah, absolutely. We are absolutely seeing reasoning and agents and specifically in the coding space. Code 4 is so good at coding a number of these startups out there that are using code today to just provide brand new coding experiences that are changing the way that people code and significantly improving our ability to code and speed at which we're able to code and innovate. And I think those agents are going to make their way into other use cases as well. We're seeing that. And it has. The demand for inference is obviously increased significantly. The size of the models and the reasoning models typically a lot larger. And so we have to think about can I fit the entire model into a single server, right? A normal server back in the day, eight GPUs. And that's probably not enough anymore for some of the larger models. 

And that's where we're deploying the GB200s, literally in the next couple of weeks, they'll be available on AWS. 

‍

And that'll actually have 72 accelerators in a single, what we call ultra server. We've done the same thing already with training, where we have 64 accelerators now within a single ultra server, and they're all memory-coherent. They have access to each other's memory. And the thing you're really avoiding there is the network trip time because if you did it on the GPUs with age, you would've to do a network... you'd have to use multiple machines. And now, these network time, and no matter how much we optimize the network, you're always going to have lower latency and lower tokens per second. So the scale-up domain, as they call them on these ultra servers, is really the next generation of inference workloads, but also training workloads. And so we're excited to be taking that next step. And there's a lot of complexity from water cooling to different types of communications in the server through PCIe, a lot of complexities for us to work through, but we've made a lot of good progress together with NVIDIA and then on our own Trainium accelerators as well.

‍

Patrick Moorhead: So Dave, we've talked about the good old days of machine learning seven or eight years ago, and we're talking about the workloads today, but you have to plan many, many years in advance to intercept the needs of what you think is going to happen in the future. Building out a new data center or filling up space that you already have is not a fast endeavor. So, what are some of the key infrastructure challenges that you see on the horizon as AI continues its evolution?

‍

Dave Brown: Yeah. I mean, the first one is obviously just the scale and performance demands of infrastructure. And I think if you look at sort of 2018, 2019, it was the... the world was probably the most stable. We learned to run the cloud really well. We had a great forecast capabilities in knowing what our customers demand. I always call it the illusion of infinite capacity. How are we actually delivering this illusion of infinite capacity to customers without actually having infinite capacity, which is incredibly expensive to do. And so AI has introduced just a whole lot of new challenges in that area. So making sure that we have the right power available. We're building the right data centers years out. That's something, as cloud providers, we've had to do for our entire history, but something we've had to get a whole lot more focused on with generative AI. You're helping customers to be able to onboard. One of the things we see with customers is data is so important in AI. You can use a model, and it's just a model that somebody else trained, but when you bring that model together with your data, suddenly, it becomes something that can change your business, literally puts you on a new trajectory.

‍

So how are we giving customers that data foundation, and how are we helping them process that? Now, a lot of customers today have already put all of their data on AWS with data lakes or S3 or whatever it might be, and we've got a suite of analytical tools that help customers make sense of that data and now allow them to use it with our AI tools. So whether it's things like SageMaker for training or Bedrock for inference, but even higher level tools like Q, you know, Q for business and Q for developer, deep integration with customers data that allows them to really get the value of AI. Security and governance is always going to be a priority. And we're going to have to maintain and always be paranoid around security and say, what is that next thing, the next type of attack we might be seeing? How do we want to build our architecture to ensure that customer data is secure? And all the way through their entire life cycle. We've got guardrails on Bedrock as well, which allows customers to ensure that a model isn't responding in a way that maybe they don't want their business to be represented as.

‍

And so that allows customers to really fine-tune those options. Just infrastructure flexibility as well, right. So we've spoken about a lot of different services, multiple different layers of the stack. That is so important where we can really meet customers where they're at, and then as they go on their AI journey, some of them might decide right now to fine-tune a model, but in the future, they might just say, "Hey, Bedrock's going to be the way to go. I'll do fine-tuning today. I'm not going to use SageMaker." So that flexibility to move, that flexibility to use different types of accelerators or processors or chips is so important. Really put themselves in a place where as the... as competition continues in the market and innovation continues to happen, you're very well positioned to get onto the next thing without having to do another large CapEx investment like you were talking about earlier. And then, finally, just cost and efficiency. We've spoken about that. We've so focused right now on just how do we get the cost down, how do we improve the models? How do we improve the infrastructure because we know that's the largest thing we can do to unlock the next wave of innovation for our customers by allowing them to use all of the services and models that are available. So yeah.

‍

Patrick Moorhead: Yeah. I appreciate you explaining it. And I like the way you answered it. It was more than, "Hey, what new hardware am I going to be putting in?" It's even getting more out of your current investment that you have in onboarding. Sometimes, I forget just how hard some of this stuff is. But with AI, I mean complexity is just, it's high, high in the radar, but well worth the investment. I mean, the 10xing of what can be done, I think, is a conservative view of what companies are going to be able to do with AI, and it's going to be a workload after workload. You just don't turn on the switch and my Fortune 100 company is AI-ified, and everything is AI, right. This is going to be a multi-year, multi-year build-out. We're going to see companies who didn't get on it quickly enough, shake out. We had a shake-out during dotcom. I was part of dotcom and part of dotbomb, by the way, watching that Echo up and down, and there's just a string of companies that just don't exist because they didn't get into the technology, they didn't get into the web, they didn't get into e-commerce quickly enough. And I think that we will see the same thing. So any enterprises out there, if you're not starting your AI journey, you are too late. You need to pile the resources and get serious about that.

‍

Dave Brown: And Pat, I think that's a great point, and I'd add to it in saying one of the things is when you try some of these things early, you may not get the results you want. And what's important for these companies is to find a way to stay at it, to find a way to stay in the experimentation, to find a way to try the new model, to try the new silicon that comes out to work with a cloud provider like AWS. As you know, we work very closely with customers. We're always looking for feedback. That's how we get better and that's how we innovate. But stay in it. Don't write it off because it is going to change things significantly. You talk about, I can't remember a time where we said 10x or something, and feel like that's probably the low end of the estimate. And we've been saying that for a while, and I think folks are struggling to believe it, but we've seen so much improvement just in the last six months, and the pace of innovation.

I've never seen a time where there's something new every single week that just surprises us in what customers are building and how the world's moving, and the pace of innovation is quite honestly astounding. So you do need to be in it, stay in it, give feedback. If you're using AWS, you'd love to see something. We always love to hear that feedback from our customers and see how we can innovate on their behalf. So very, very exciting times.

‍

Patrick Moorhead: Yeah. Dave, thanks for the time. I appreciate. You've been very generous as well. I think I probably talk too much, but I love this stuff. I get excited about this stuff, and I know you do too. So thank you for coming on the show, Dave.

‍

Dave Brown: It's a fantastic conference. So thank you very much, Pat. It's great to be you. Thanks for having me.

‍

Patrick Moorhead: Thank you. And thank you for joining us here for the cloud infrastructure track opening keynote here in the Six Five Summit in its sixth year. Stay connected with us on social, which I'm on way too much, I'm sorry. But explore more conversations at sixfivemedia.com. More infrastructure conversations to come. Thank you.

‍

Disclaimer: The Six Five Summit is for information and entertainment purposes only. Over the course of this webcast, we may talk about companies that are publicly traded and we may even reference that fact and their equity share price, but please do not take anything that we say as a recommendation about what you should do with your investment dollars. We are not investment advisors, and we ask that you do not treat us as such.

‍

‍