The CEO of Amazon, former CEO of AWS, Andy. Great to see you. Thanks for having us in your little den here. Used to be doing media hits all week. It's great to be with you guys. This is a tradition. We've done this a lot of years together. I love the fact that you're doing videos and you came on stage. The energy in the room completely changed when Matt introduces the OG, the Godfather of Cloud, and then the crowd reaction was pretty strong. You could really feel the energy, and then you dropped the mega announcement of having your own model. But just in general, since then it's just been a great event. My first question is: do you miss it? You're back.
Oh, it's well, first of all, I thought Matt did an awesome job in the keynote. And it's a thrill to be back. What I miss most about it is the people. Just being around our customers. We work so hard all the time and we see all the warts and all the things that we think we can be doing better for customers that we want to deliver for them. You're kind of heads down, and then to be around all your community partners, your customers, and your partners, and have them excited. We hear stories. One of my favorite parts about re:Invent every year, and I've heard this at least three times in the last 24 hours, is that people come here to learn from each other, to learn about what they can be doing differently, and to be inspired to go back to their own businesses and change their customer experiences. You get teams that come together, they see a bunch of services released or they hear what others are doing, and they go to the bar and they say, 'We're doing this, let's go,' and they go back and they make a change. That's what this is about.
The energy was high, and we're kind of in this AI generation. You can see some of the cards you guys are playing, the infrastructure advancements, but now with the AI, I have to ask you: you've operationalized the cloud and did all the greatness, turned the world on fire, created disruption, kept taught people working backwards, documents, operationalized everything. Now with AI, is that the same plan that you see? Because these builders are going to go faster, they're going to have the AI at their back. How do you look at this next wave of cloud?
I think they're very much connected. Just perspective-wise, we felt like we grew the AWS business really quickly, and it took us about eight or nine years to build. To use AI, you have to have your data organized and architected in such a way that you can access it. Trying to do AI from a mainframe is nearly impossible. So you really need to have your infrastructure modernized in the cloud and your data accessible to run AI. Then the reality is that, as you've heard us talk for years, every application is compute, virtually every application is storage, almost every application is database and analytics. Another one of those core building blocks is going to be inference. Every application is going to have some of AI and inference in it. So it's very much a building block. I also think that you don't show up and just have one big service that everybody uses. We have a bunch of different chips and services that we think are going to help people be much more choice-performant in their training and their inference, and you'll continue to see that from us.
When you think about the impact that Amazon has had on the industry broadly, it's pretty remarkable. John mentioned working backwards, customer obsession, two-pizza teams, the flywheel concept, what I call Jesse's law: 'There's no compression algorithm for experience.' Two questions: has that changed? Is there a compression algorithm for experience in this AI era? And are there new sort of operational frameworks that are emerging as a result of this?
I think that the models... it's not like people started working on these big transformer models two years ago. People were working on them for a long time, and the first several versions... the first real application that just exploded was ChatGPT. But if you think about ChatGPT, it's kind of a pretty thin user interface on top of a model. It's remarkable. One of the things that people don't realize, I was talking about this in the keynote, is that to build great generative AI applications, it's not fast actually. If you're doing a software development project, we can get on a whiteboard and map it out, and of course there are differences, but it largely functions as you design it. Whereas with generative AI apps, it's very iterative. You think you're going to make a big advancement with what you've done, and then it turns out you don't. You think you're making a tiny change, and it gets much better. So it very much applies that there's no compression algorithm for experience. What we tell a lot of companies: before the pandemic, companies were on this march to modernize their infrastructure to the cloud. The pandemic hit, we got into survival mode, then there was an uncertain economy and everyone was cost optimizing. Now as people start to spend again, people are asking us, 'Should we modernize our infrastructure or should we do generative AI?' Of course the answer is yes. But if you don't take the low-hanging fruit of modernizing your infrastructure, you're actually not in a position to do AI well.
AI clusters, people try to do it on prem. Cloud has advantages. Also in the cube, the word 'blast radius' has come up a lot, which I know is that Amazon term that James Hamilton and the team uses. It's hard to build large-scale infrastructure. Dave Brown and I talked about that, talked about the dynamics of the infrastructure at scale because you guys see stuff at scale that others don't. Things do break. Building a system to power the kind of apps that are coming, certainly the demand is going to go crazy with Gen AI, certainly Bedrock and some of the models, but to run it all you need the horsepower.
We've been building very large-scale infrastructure and running the largest workloads in the world for a really long time. We're going to run a cluster of a hundred thousand chips for Anthropic for their future models of training. I also think that one of the biggest inhibitors for generative AI applications to be as broad as they're going to be is skill set and experience, but a good piece of it is the costs need to keep getting lower. The cost of the compute and chips so you can do training much more cost effectively, optimizing inference, which again is the cost of the compute but also some efficiencies. We've done a lot of inventing there. As we are building very large-scale Gen AI applications inside Amazon, we're building about a thousand of those apps already, as well as running for customers. If you take HyperPod and combine that with what we're going to be able to do with Trainium, with Ultra Servers and Ultra Clusters, it's just a different level of skill to be able to train. To replicate that on prem, it's really hard.
