AWS CEO Adam Selipsky On Amazons $100 Million Investment Into Generative AI

Adam Selipsky, CEO of AWS. It is wonderful to have you with us here in studio. Thanks for coming. Thank you. It's so great to join you. Thanks for having me. So some big announcements today. I'm wondering if before we go into them, you can walk us through Amazon's generative AI strategy. Sure. Well, Amazon has been doing machine learning for, I'd say, 25 years, probably started with

personalization on the Amazon website. And if you fast forward AWS in 2017 launched Sagemaker, which is our machine learning platform and we now have over 100,000 customers doing machine learning on Sagemaker. So it's fair to say that on any cloud, on all clouds, the majority of machine learning happening in the cloud is happening on AWS. And then in terms of large language models or these LMS

some elements of our retail website search. Also Alexa, a lot of her voice responses are driven by by large language models. So we've got a lot of scientists who've been working on LMS in particular and on AI in general for for a long time. And now we're bringing all of that to bear, plus a lot of additional resources in generative AI in particular. When you say

a long time or a good amount of time and how you've been implementing large language models and Alexa and others, is that years, is that months? Were you doing this before the explosion of ChatGPT? Yeah, I'd say years. And the concept of LMS, particularly, particularly if you're a scientist and in that world is not new, there's certainly been really impressive advances over the last couple of

years. And ChatGPT is the application that's sort of taken the consumer world by storm, if you will, so approachable and easy for people to understand. But really, this whole world is going to go far, far beyond chat applications, although that's a cool application. But AWS is really taking a full bottom to top full stack view of generative AI. So we have first you've got the models

which are incredibly large and incredibly expensive to train and just because they need so much data. And so we've actually designed our own chips which go into our own compute capacity that customers can use to, to train all of these models. So and our chips that we've designed the trainium for doing training and inferentia chip for doing running models in production that's called doing inference are

really going to have industry leading price performance. And then of course, a lot of people have heard of GPUs and that's another technology chip type that's used. And we'll we'll have lots of customers using those as well. Which are also trying to develop in-house, correct. The artificial intelligence, higher end chips. Yes. So our own chips are trainium and Inferentia chips. And then we'll also sell people

capacity based on GPUs because there's not going to be one solution in all of AI. It's not not a question of what's the answer. It's a question of what choice can be provided. How do you democratize all of this? How do you provide a whole lot of options? Right now it feels like there is one GPU, one player and that is Nvidia's H100. Are you saying

that there's more out there now that has that capacity? Yeah. Just like anything else to do with cloud computing. There are so many customers, so many people are rushing to the cloud. It is not a single homogeneous world. There's lots of different types of customers, many different use cases. You always have to pick the right solution and the right application. So GPUs are immensely popular and

really good for a lot of different applications, but at the same time, it's not going to be the best answer for everything. So just like for general purpose computing, we've built our designed our own chip, which is very popular, but lots of people are still going to come to AWS to use Intel or AMD chips. Lots of people will come for GPUs, but lots of people

are also separately going to want these chips that we've designed because it's going to be really, really great price performance. To be clear though, those are for a different purpose. They don't have the same kind of power and capacity as Nvidia's. No, they absolutely do. And it'll be used for generative AI and to to train models and to run models in production. I just think the

different customers will make different choices because it's a it's a multi textured, heterogeneous world out there. There's room for a lot of different solutions. Amazon currently has in-house chips that it's designing that has the same capacity and compute power as Nvidia's. A100 have the our own chips that we are in production and shipping today to train and one for training models for doing inference or running

models of production to render our. Generation of chips with Inferentia two that we announced earlier this year. They're in production, They're running, They work amazingly well. And for for many, many applications, they provide absolutely the best price performance. Is that a misperception in the market then, that everyone's trying to get their hands on this particular Nvidia chip that they're selling on the black market in China

that venture capitalists are buying up to entice startups? I mean, is that a misconception then? Why? Why are so many looking to those? Well, I mean, GPUs are immensely popular and they're used very pervasively for all types of machine learning, including generative AI. So absolutely. And we're we're a close partner with them. And many, many customers come to us to use Nvidia based GPUs for sure.

In addition, we have many, many customers more and more as we just keeping keep on increasing our capabilities. Many are coming to us to also use the AWS design chip. So it's a big and not an Or. Okay, let's get to the announcement today. Amazon is investing $100 Million into a new generative AI innovation center. So you're leaning into your own technology here. Is that enough?

I mean, when we look across the tech landscape, Google is putting three times that into anthropic. Microsoft put $10 Billion into OpenAI is $100 million into a generative AI innovation center. Is that enough to really be a player in the space? Well, it's a good start. And also that's purely to work directly with our customers. It's one tiny piece in the overall landscape of what we're

doing in in AI overall and then generative AI, we talked about Sagemaker and the whole machine learning platform we have. I mean, Amazon has spent many billions of dollars on AI over the years. And if you look at not only our our chip efforts, our machine learning service, Sagemaker R and then on on top of those, those chips and the compute capacity, we have a new

service called Amazon Bedrock, and it's going to be a very, very important and popular service. It's in private preview right now. And the idea is that what the world needs is not one model. What the world needs is choice. The world needs to have AI democratized, just like I was just like it rather was democratized by AWS. So in Amazon Bedrock, the managed service for doing

generative AI, and we're going to have Amazon's own models that we're building in there. Those are going to be called our Titan models. And in addition, we're going to have a lot of the leading startups and models out there. So Anthropic, whom you mentioned, will be in Amazon, bedrock stability. Ai will be in bedrock. Other startups such as AI 21 will be in bedrock. And choice

is such an important concept here because otherwise you ask yourself the question, Well, where are the different runners? Three steps into a ten K race? Does it really matter? The point is your three steps in and it's a ten K race. And what what people need today is choice and to be able to experiment and to be able to figure out what different types of models,

what different types of use cases are most powerful. And that's why customers are so excited to work with AWS for generative AI. So what you're saying is instead of giving them the large language model, you're giving them the tools to build their own. Is that a better way of we're. Going to give them whatever they need. So there are going to be a relatively small number

of companies who will build their own models. And for them, AWS is the best place to come and build those models. Use whatever chip you want to use, get access to all of AWS, other compute storage machine learning capability, higher up the stack, you're going to have probably the majority of customers who want to use large language models and other models that that other companies have

built. And and that's where Bedrock comes in. So instead of saying, hey, there's one model or one company, I mean, can you imagine, let's say it's 1997 and we're talking about the Internet and we ask ourselves, who's the Internet company going to be? I mean, it kind of seems like a silly question, right? Who knows what you might have said? But I do know that AltaVista

was the number one search company in 97. And I promise you, my kids have have never heard of it. So it's kind of the wrong question to ask. So I think the question to ask now is how do we experiment with generative AI? What are the different models that I could I could try, what are the different use cases inside my enterprise that I might want

to apply generative AI against? And that's why we've got got bedrock with all of these different models. So we will provide secure, private, consistent access to as many of the world's leading models as as we can. Right. So you mentioned, you know, the $100 million that you're announcing today that you're putting into this generative AI innovation center. It's not apples to apples to what the other

companies are spending. Is there a more accurate number that you could give that is more representative of Amazon's spending on generative? Ai? Yeah, I. Don't think I have a single number. It's happening all over the company. It's happened for years and years. We're making use of all that. We're making use of all the accumulated experience we've got, but we're putting rather enormous resources into machine learning

and AI in general, including generative. In particular, but I'm very excited about the the new generative AI innovation Center because it is a specific program to take experts that we have, such as data scientists, engineers, solution architects and strategists, and go work directly with customers. Very much one on one to understand what problems do they have, how can we help them to bring generative AI to

those problems and then to design specific solutions that they can then go and implement on AWS very quickly? So Wall Street is skeptical that Amazon can capitalize on AI. One of the most bullish analysts on the street says Amazon, quote, lags its mega-cap peers in generative AI abilities today. What is he missing? I think the reality of it is that Amazon has been working on artificial

intelligence, as I mentioned earlier, for longer than almost any of these companies, and we have long accumulated experience. We have many, many experts, thousands of practitioners of different kinds inside the company. Many of those experts are now working specifically on generative AI. If you look at the full, nobody else has the full stack of capabilities that we're that we're putting together. So if you take the

chips that we talked about, we're designing our own really priced performant machine learning AI specific chips. And then on top of that, we have a managed service for accessing the most important models in the world, both Amazon's, which will come out later this year, the Titan models, as well as a whole slate of really important other models that startups and others have developed. And then on

top of that, we're going to build some really exciting applications and solutions. So this year we launched Codewhisperer, which is a coding companion. So you type in English, we'll get to other languages eventually, but for now you type in English. It returns code and in internal tests. Codewhisperer reduced the time to complete a task by up to 57% for for Amazon developers. And we released a

couple of months ago. It's already immensely popular. So if you look at the other players out there, you really don't have anybody who's designing the chips, has a really a choice oriented, democratizing layer to access the models and is building solutions on top of those models. So customers really need us, all those customers. We are by far the the biggest, most widely adopted cloud and customers

need generative AI from AWS. And you don't think. That's something that Microsoft and Google are working on? Also that sort of full stack of infrastructure and applications? Well, honestly, we focus way more on customers at Amazon than we do on anyone else, but to my knowledge, nobody, including those companies is working on, certainly have not announced that they're working on those full that that full suite

of capabilities bottom to top of the stack. No. Then I wonder what is Wall Street missing? Is it that Microsoft and Google have sort of this product that has captured the mainstream? I think you said earlier taking the consumer by storm. Does Amazon need something like that? Does it matter that you're not being recognized on Wall Street? You there's an expression in Amazon, which is a

pioneers need to be willing to be misunderstood for long periods of time. And I think Amazon has had many examples in its history where it said we're going to focus on customers and have a steadfast belief in that. We're going to work with customers, we're going to build what they want. And if people want to perceive us in a certain way for misunderstood, that's okay as

long as customers understand where we're going. And so what we're doing is trying to stay laser focused on building what our customers need in this space. And and as I mentioned, there's not one customer need the world is not I think I think chat applications are amazing, but it's only one of many, many things that the world needs in generative AI. So we need we need

an AI in order to discover drugs better. We needed to monitor industrial equipment better. We needed to create autonomous driving systems in vehicles. We needed for to make media and entertainment better. We needed really for for every sector of the economy, every, every application. And I think that probably what's happened is because it's super easy to understand, oh, I can I can ask for a haiku

about medical equipment and you get a haiku about medical equipment. But there's more. But there's a lot. I mean, that's fun, but it's not nearly as profound as the many, many innumerable generative AI based applications that the world is going to need that the world will benefit from. And we're focused on on that mission, not on being understood by a small group of folks. And the

cloud is a good example. Amazon had such a big head start before even your rivals caught on. So point taken. Well, I helped start up AWS and in the early days, if I had a nickel for every time I was asked, But what does this have to do with selling books? It would be great. So we were misunderstood for a long period of time. Now, of

course, that's turned into the cloud and all of these other companies eventually followed us in After a number of years in. Being far and away the leader, far and away. The leader significantly larger and more deep, more broadly deployed than any other cloud. Okay. Well, let's talk about that traditional cloud, I guess let's call it, and move away from generative AI. We're at a stage now

where the cloud is now become traditional. Isn't that funny? I guess we finally arrived. We've arrived. I would have said that a few years ago. Well, we've also seen growth slowing, though, so maybe it's arrived and maybe we've gone even a little bit further at AWS as well, some of the other hyperscalers. I wonder last run on AI. Has AI perhaps been a convenient distraction? Can

AI business and the monetization of it make up for slowing enterprise spend in the cloud, more infrastructure and applications? Well, I think that I is going to be this this next wave of innovation in the cloud. It's going to be, you know, the next big thing that that pushes even more customers to want to be in the cloud. Really, you need the cloud for generative AI,

you know, the amount of the amount of computing that needs to happen and the amount of storage, the amount of IT capacity overall. But are people. Paying for it. Now? Is that being monetized? It absolutely is today. And I think and we're only scratching the surface. So I think if you're going to if you're going to build a model, you're going to run an application for

your customers. The cloud is going to be by far the best place to build and run those generative AI capabilities. And at that point, it's like, okay, well, which cloud am I going to use? And actually the calculation is not going to be any different, I don't believe, than than the existing calculation for existing applications. Which cloud do I want to use? It's about security first

and foremost, and then about operational excellence and uptime, then about choosing the cloud with the broadest and the deepest set of capabilities, and then having platforms that all work together saying, here's where my my data resides and I need the data to weave in gracefully with my generative AI. And AWS is the leader in all of those areas. So can you say then with increased spending

on AI applications and programs, has growth bottomed at AWS? I don't believe that growth has bottomed at all. At it's slowed down permanently slowed down at all. Has it reached the bottom, though? Do you see it coming back? You know, there's a separate phenomenon, which is obviously we're in a time of great economic uncertainty. You're seeing that in a lot of sectors in the entire tech

sector. Certainly, you know, a lot of software companies, certainly most cloud based companies, growth has decelerated over a number of quarters. And really what's driving that is we have a lot of customers who are doing their own belt tightening and their cost optimizing. And because we're so long term focused at Amazon and because we're so customer obsessed, we're actually leaning in and. Helping to save the

money. And I know you've talked about that. They come in and say you've got to help us cost optimize and we say, Great, we'll dig in. My question is, is that period over? Are you starting to see customers spend more again because of this generative AI ship, because they feel better about the economy, because tech is taking off again? Yeah, I think there's a couple of

you sort of have to unpack it d average it, if you will. So I think a lot of customers are largely through their cost optimization, but there are a lot of other customers who are still right in the middle of it. So it's hard to predict exactly when when that particular trend will be over. But we're still in the middle of it. And so a lot

of customers are either slowing their cloud growth or in some cases absolutely reducing the size of their spend for a specific workload. But at the same time, many of them, even as they cost optimize, are signing large long term commitments with us and talking with us about the next hundreds of applications. They want to move to the cloud. So I think that the we're still very

early in the move to the cloud. We might be hard to really know, but call it 10% of it is probably moved to the cloud. And I think most of it will move to the cloud over time. So I think there's lots and lots of robust growth ahead of us. We've talked about how Amazon had the head start. Aws is the clear leader. However, some on

Wall Street thinks that that could change over the next decade or so. Morgan Stanley predicts that Microsoft's cloud unit Azure will actually surpass that of Amazon's in terms of market share by 2032. What are you doing to prevent that shift from happening? Well, I can't comment, nor do I particularly spend a lot of time on any one analyst projection, to be honest. It's it's just not

I don't think it's. Do they have. An advantage, though, in terms of their enterprise customers and their reach? And, you know, today Google filed an FTC complaint saying that Microsoft was anti-competitive. No, I mean, look, every we are in a market segment with robust competition, as we should be. And I always say, if you look to your left and you look to your right and you

don't see anybody next to you, you may have dramatically overestimated the attractiveness of your segment. So we have a lot of strong competitors, and that's good for customers. And I think ultimately, by the way, it's good for us, too. It makes us a better competitor. So everybody's going to have their advantages. But that being said, I think AWS has more significant advantages and provides more customer

value than any other cloud provider. And if we do our jobs, it's not our it's not our birthright, by the way. But if we do our jobs and if we wake up paranoid every day about disappointing our customers, not paranoid about competitors, but paranoid about our customers, then I think we should have every ability to stay. The industry leaders that we are today by a significant

margin, by the way. But you probably agree that anti-competitive practices would hurt your customers. So do you care if Microsoft is employing sort of anti-competitive procedures for their cloud? Do you think they are? Well, I certainly. Think that there should be fair competition. Everybody should engage in fair competition. And there have been a couple of different software providers who in the cloud space have actually tried

to apply restrictive licensing provisions and basically tried to slow down their existing customers from being able to run the software that they've purchased on the cloud that they want to run it on. And I don't think that that's right. And I think, you know, a lot of folks have recognized that. And hopefully we'll continue to recognize that. I would rather have us all compete on the

merits and let customers choose the solutions that are best for them. I know you don't like to think about your competitors. You focus on your customers, but your competitors are thinking about you. Oracle's Larry Ellison said he's winning customers. Is that true? Uh, what? A number of years ago, Larry said he didn't know what the cloud was. So I guess it depends on on what day

or what year you ask the question. Look, as I said before. We're making strides, though, aren't they? Well, there's robust competition in the cloud space. That's not a competitor that we see in a lot of deals, to be honest with you. But I mean, the cloud is a very, very attractive market segment. We've always said literally from the day we launched, we said it would not

be winner take all. We said that there would be probably a small handful of disproportionately significant providers. And I think you're seeing that play out. So I think over the years you've seen some rise, you've seen some fall. I think over the next few years you'll continue to see some rise and you'll continue to see some fall. And but, you know, we we like our existing

position and we like our prospects if we continue to innovate faster than anybody else and provide the best security, which we do, provide the best operational excellence, which we do, and provide the broadest and deepest set of capabilities which we do, including generative AI. I want to go back to sort of the chip conversation that we were having. We were talking about the importance of GPUs

and does Amazon have enough of that Nvidia H100 that are in such high demand to give your customers the best performance? Yeah. Well, I think that the entire world would like more chips for doing generative AI and whether that's GPUs or whether that's Amazon's own chips that we're designing and, and so short term everybody's kind of looking under every rock and and so it'd be hard

to. Venture capitalists are buying them up in some. Cases it would be you know entrepreneurship. There's no stopping it. So I think it's fair to say that everybody would like more today. I think if you look forward a number of months, it feels like that shortage should ease over the coming months and into next year. And I think we have a significant advantage and can provide

a significant additional value that's hard for others because not only are we a leading host or based on on GPU capacity, but as we've discussed, we also have these other whole families of chips that that designs and that we control the supply chains for them and we can cause more supply to be brought on board without having to be reliant on anybody else. And so if

you look at all of the generative AI compute capacity that's going to be needed, I think that we're in a better position than anybody else on Earth to supply the capacity that our customers collectively are going to want. How does that affect or not affect Amazon's CapEx plans? Do you foresee a big bump up in spending to be able to have the infrastructure and the servers

to accommodate this shift? Well, we're in a capital intensive business for sure. I mean, if you just look at the infrastructure we build, we have 31 full infrastructure regions around the world and those are all significant CapEx efforts. So for sure, there will be a lot of CapEx spent on generative AI, just as there has been in every other major technology initiative at AWS. So it

will continue to be capital intensive, but also I think good business and steady. Or do you think that this shift requires a big outlay. At some point? Well, you know, there's always we're a big enough business now where AWS is an $85 Billion a year run rate business. So there are always things which are growing more rapidly and things which are growing less fast and in

some cases shrinking. So I think you'll see CapEx shrink in some areas. Generative AI is certainly an area where CapEx will will increase significantly and then we'll have to see where it all shakes out. Bottom line, you. Mentioned a few times that Amazon is developing its own chip in house and you're doing that with the help of ARM as a partner, the chip company. Can you

explain how you work with ARM and how that relationship has evolved since you rejoined AWS? Sure. Well, we've had a the relationship has not changed since, I'm happy to say the relationship hasn't changed, particularly since it came to AWS. It's just been expanding constantly for for years and years. We've been working with ARM for I guess a decade, but I have to have to check and

they've been a very important partner for us as we've kind of brainstormed and led and directed our own chip design efforts and started with these general purpose chips. Graviton We're Now in Graviton three, which is our third generation of general purpose. Chip And one of the coolest things about Graviton three is that it's 60% more energy efficient than the equivalent x86 based based chips, and customers

care a lot about that sustainability these days. So it's a really important effort for us, as is our entire chip effort. And we work very closely with ARM. We meet with them frequently, other important design partner. It's all part of a broader community or ecosystem of companies in this chip area. And their relationships with other companies in that community is really important to us as. Well,

we often note how critical or how unique or strategic arm is in the chip space. Is there anyone else who could do what they're doing in terms of helping you design your chips in-house? I think that we're very happy with the ARM relationship. I think in anything we do, we always want to make sure that we have choices. The customers will ultimately have choices because part

of our job is to ensure consistent supply so customers don't have to think about a supply chain. We're just the supply chain, so you're always going to want to have optionality. But we love our ARM relationship and only see it increasing for years to come. Great. Well, Adam, thank you so much for taking the time to chat with us. Appreciate you being here. Thank you. I

appreciate it.