AWS CEO Matt Garman Argues Amazon's AI Strategy Can Win in 2025

Yeah, thank you for coming. I appreciate you being here.

Well, I'll tell you, in the AWS business, we're incredibly excited about where we are and the potential we have going forward. As you mentioned, our results, we added $1.6 billion of revenue quarter over quarter. The growth is really great. And when we talk to customers, the potential out there is just massive. We have maybe 10 to 15% of workloads that have moved to the cloud. So there's enormous potential out there and customers are super excited about AI. When we think about our strategy, we're laying the groundwork for where we think the real value is for enterprises from an AI perspective, from a modernization perspective, from a cloud perspective for the next two, three, five years. If you take a look at some of the things that we're launching, a couple weeks ago we launched what we call Bedrock Agent Core. Agent Core gives you this secure platform for customers to really go build agentic workflows that they can then drive a lot of the value in their business. When we talk to customers about AI, which everyone's excited about, they should be. It's such a transformative technology that is really going to change everything that customers do and every business and every workflow and every job out there. When you think about that, a lot of the benefits that customers are going to get in the future when they really start to measure the ROI and when they really start to see how this is materially changing the business is going to be from these agentic workflows. So what we're doing is laying the groundwork for how do we have a scalable, secure set of capabilities that customers are able to do. That means how do they have every model available that they want to be able to use? How do they have a framework that isolates them? Make sure they have secure agents that can go and do work on their behalf in a secure way, authenticated way, a measured way that they can actually audit and look at. And then how do they have all of the data capabilities that they can be sure to actually go and drive value from that? So all with a really low-cost infrastructure that has purpose-built infrastructure built for it.

Yeah, it's a really good question and frankly it's a place that we don't love being. So that's something that we work really hard to make sure that we don't get supply constrained on. Sometimes there are a couple of workloads that we're unable to support from very large customers, particularly building very large training clusters or other things like that where there's just massive amounts of compute. We have longer term relationships with some of those customers where we plan for future years out. But a lot of what we do is try to make sure that we don't impact mission-critical enterprise or startup workloads that are core to what they're doing. We think about some of these opportunistic workloads where they're really large swaths of capacity. Our view is we want to make sure that customers can run their businesses on us and that they can take advantage of the innovation and the technology that we have and that they can trust us for the long term. So those are the workloads that we really prioritize for our customers as we think about what we can support and what we can't support. We're working really hard to be able to say yes to everybody. I feel very good about where we are in the medium to long term. In the short term, the whole industry is growing incredibly fast and no one can say yes to everyone. Part of that is how do you continue to knock down the various constraints there are in the world. Sometimes it's chips, sometimes it's power, sometimes it's other components, other times it's demand. There are always constraints in business when you're trying to figure out which thing is perfect and growing. Right now we continue to knock those down for this week, for next month, for next year, and we feel pretty good about it. Part of the other thing we do is we're very focused on how we also grow our business responsibly from a capital allocation perspective and a spend perspective and a risk perspective, but also from a strategic perspective.

I don't think so. No, no, no. I don't think so. On DGX Cloud, look, we're great partners with Nvidia. DGX Cloud runs in AWS. In fact, we're one of their biggest DGX Cloud partners. Jensen and I talk all of the time. We're very close partners with Nvidia. We do add value to that though. There's the DGX Cloud where Nvidia actually adds some really great value to their cloud. They have some software on top to help customers train and get efficiencies out of their hardware, which is great, and they sell that on AWS today. We have many customers taking advantage of that in AWS today. We also sell Nvidia servers, GPUs, as EC2 instances, and in those we do add capabilities there. So you're right, it's not just the same thing that you go get from everybody else. It actually has what we call Nitro. That gives customers much better enterprise security, better isolation, much better security, much better encryption. From a security perspective, we feel pretty convicted that that's an important part of running in a customer environment. But that's unrelated to being supply constrained. We've been building the Nitro system for a decade now. We know how to add that in. We work very closely with the Nvidia engineering team to make sure that seamlessly integrates.

We love the AWS business and we really believe in the growth going forward. I think it's a massive opportunity and so we're investing pretty significantly in the AWS business and in Amazon. As you saw from our numbers this last quarter from a capital spend perspective, we're out there spending and building data centers and getting power and getting capacity for customers as fast as we possibly can. So from that perspective, we're spending a lot. I think you still want to think about this as a very dynamic business though. You could always add a zero on what you're going to spend and at some point you make a mistake and it could be a problem for the business. So our job is also to be responsible there and think about how do we grow fast? How do we make sure we have enough capacity for our customers and don't put the company at risk by committing hundreds of billions of dollars many years in the future when technologies may change? Those are a number of things I think we think about. We balance those. Unfortunately, we've had multiple decades to get good at this. Amazon in particular is quite good at this both from our retail side where we think about capacity for our fulfillment centers as well as on the data center side where we think about how do we make sure that we have enough power, how do we make sure that we have enough low-cost infrastructure so that our customers aren't stuck with higher prices further down the road when something may change. So I think we try to balance all of those risk things and responsibly grow as fast as possible.

At Amazon, we've been incredibly fortunate to have many of the most talented, smartest technology minds on the planet for the last two decades. We continue to have those folks. By the way, we pay them well. But we also want people that are missionary. We want people that are here because we're working on exciting problems, because we're building interesting technology. We want people who are customer focused and really want to work for customers and solve customer problems. We don't want people that are just here for a paycheck. There are other places you could go to potentially make a little bit more money. But we think that Amazon and in particular AWS gives you the most exciting interesting problems to work on working together with many of the most talented people. Those are the folks that we want on the team. We think that produces the best results as a team for our customers.

Our view around AWS and particularly in Bedrock has always been to make every single possible AI model available for our customers. We want the best models available. But it takes two partnerships to make that work. Our goal has always been to have this set of models so that customers can build anything that they want with the very best technologies capable. When you're building an AI application, the model is a core part of it. It's not the only part. People want lots of other things. They need guardrails, great management of their data, the right ability to control agents, a bunch of these other pieces. But the model is a key piece of that engine. Many customers love the OpenAI models. They've been one of the leading model providers for the longest time. They're the first.

They're definitely the first ones out there. They had early adoption and first mover advantage. I think folks love the OpenAI models just because they started getting going with them first. They probably had a year head start on many of the other models and customers are familiar with them. They wrote a lot of prompts for them. They're familiar with them and they're quite good models. For different use cases, we don't think there's going to be one model that's the best model for every single use case. Sometimes people want to use OpenAI. Sometimes people want to use Anthropic Claude models. Sometimes people want to use open source like Meta's Llama models. Sometimes people will custom make their models for their very own use case. We see that many customers actually use a variety of those models and combine them all together to build interesting applications. OpenAI launched an open source version of their own mini model and we're super excited about it. It has fantastic performance particularly for text-based workloads and an incredibly low price. It's a brand new competitor in the open source area. Many customers also love using open models. One of the benefits of an open model is you have a lot more flexibility to go play with it. You can play with the weights, fine-tune it, customize it to your own use case. So we're super excited to see how customers inside of AWS use that. We've seen a lot of excitement already.

We've been a huge fan of open source for a really long time. We've built much of our business on technologies like Linux. We build businesses around things like OpenSearch. We build different open platforms that either we've created or that we partner with the ecosystem to invest in. I think it's a really great model where open source can be secure if you really know what you're looking at. In fact, most times it's more secure because you have more eyes on it, more people thinking about it, more people finding ways to patch bugs and other things. Everybody can see the code as opposed to it being a black box that only one particular company can see. From that perspective, lots of our customers and Amazon love open source. We're big fans of that. Part of what we do is when we vend open source, we make sure that it's the very latest offering, that we apply patches when they become available in the open source community, and we have teams jumping on to make sure that we're also looking for patches and secure environments. We also give customers a great way of running those. Many times open source technologies are actually hard to operate. They're great pieces of software, but they're hard to operate in a scalable way. Part of what AWS is really great at is offering hosted versions of open source where it's open source technologies, but you don't have to manage servers, you don't have to manage patching, you don't have to manage scaling. We do all of that for you, but you have all the flexibility of using open source. That gives our customers a lot of times the best of both worlds.

I believe that's right. I think those are still proprietary on Microsoft.

They're a great partner of ours and the team has built some incredible technology. Kudos to that Anthropic team. They have a really great team there and we share a lot of the ethos around really wanting to support customers and being missionary about what we're going after. One of the things that I love about that team is they really leaned into AWS and Trainium. They recognized early on the path that we were on and how that technology was going to allow them to scale and grow at a lower price and a lower cost and better performance for their customers. They basically saw early on, we worked with that team before Trainium 2 was even launched on the market. They looked at it, realized it's going to be a great technology for us to build on, and made a bet together with us before it was even built into a server and launched. Now they're quite happy with it and we're scaling. It worked out as a good bet for both of us, but they made that bet before Trainium 2 was made as a chip.

It's a similar thing and we've seen this already. We've been making our own custom Graviton chips, which is a general purpose processor, ARM-based core, for 7 to 10 years now. We've seen a massive shift where many of our enterprises have moved big chunks of their workloads over. With Graviton 4, the latest version, customers see absolute benefit. They see 20% better performance than the absolute best performance they can get on an x86 chip at 20% lower cost. That is a massive benefit to customers. When we do the same thing to AI, we did the general purpose processor first. It was a huge success. We moved on to doing an AI chip. I think that we'll see similar opportunities there. We're leaning in as you mentioned with some of the early adopters to really get some of that benefit, but the cost benefits are massive when we really get there. As you look at one of the limiters of adoption in enterprises today, one of the things that could slow down AI growth is cost. These things are still expensive. Our long-term view is how do we continue to work with model providers to make more efficient models, better models, better capabilities, better ways to manage these things, and a lower cost infrastructure to make sure that we keep driving these down so that as these things really scale into enterprises where you really get the value and enterprises see this ROI, the math is just easier because the capabilities are cheaper.

I think it's probably more around capabilities. Honestly, when enterprises and customers of all sizes, whether you're an early startup looking to build your business somewhere or you're an enterprise saying, 'Who do I bet my business on for the next 10 years?' I really think the key is who are they going to pick and they're going to pick based on capabilities, based on value that they get, and who they can trust from a reliability and security perspective. That applies to AI, it applies to core workloads, it applies to all of those things. I don't think that enterprises are going to be willing to give up availability. They're not going to be willing to give up security. They're not going to be willing to give up cost. For a long time, we said we view the cloud relative to traditional IT companies that were 80, 90% gross margin businesses in software. The cloud was going to be a lower margin business. For Amazon, it's a higher budget business because we come from the retail world, but it was going to be that and I think you've seen that play out where it's a very nice business. It's a good margin business for us. It's very capital intensive that you have to invest in. I think that will sustain, but I do think that the choice of who people are going to go with is going to be much more based on who they can trust, where the innovation is, and because businesses really are betting their future. When they're betting on that, they're not betting on financials. They're betting on who's the customer they can count on or the partner. It really is a partnership on who they can partner with for the next 10 years.