Amazon CEO Andy Jassy on the $220 Billion AI Buildout

So, these numbers are um you know, they're they're kind of mind-blowing numbers. And um even spending $220 billion in capex this year, we're leaving a meaningful amount of demand on the table. we just the the thing that helps us here is that we we have really good visibility uh into customer demand. And so if you think about the the capex investments in AWS and the way the cash cycle works, there's two macro areas. There's the investment for data centers and then there's the investment for uh for servers and networking. the data center piece, you have to invest, you know, about two years in advance of when you can put servers in the data centers to start to monetize them. The good news is you only make that investment once. It's a useful life asset. That's about 30 years. Um but um but you make that a couple years in advance. The server um and networking gear, it's only a few depending on the component, it's only a few months in advance. So you have very good visibility into customer demand before you pull that trigger.

If it turns out you don't actually need the capacity to fill those data centers, you won't you won't buy the server networking gear that on average they you know they they tend to break even uh in about three years and they have useful lives at this point of about five to six years and most of the deals we're doing are about fiveyear deals. And so when you are building as fast as we're building with so many data centers simultaneously, you're going to spend a lot of capital and your free cash flow is going to have headwinds in those years until you get to that those later stages, you know, a few years down the road where all those data centers are online and they're all being monetized and you're making back your money from the server and the network investment. But it has because of the way the cash cycle works, it has headwinds in the early years when you're growing that fast. So actually on that, you know, data centers going back when you started AWS were probably the last thing in the world that you would believe would be a political issue.

Let's go back to the capex for a second. You the $220 billion, you know, that's $600 million a day. just to you put put in perspective. I'm I'm curious how you think about AI demand right now offline. We were talking about how you see it as a a barbell and and maybe you can sort of situate or contextualize sort of the state of AI demand at the moment. Well, I mean the first uh you know the first context piece is what I said earlier which is just uh I can't believe how how big and fast AI has grown. You know it's you know a few hundred times faster than the beginning of AWS in the first three years. $25 billion run rate for us just in in AI revenue in AWS in the first three years. So it's really large and yet it really is barbelled if you look at how demand it is. And what I would say is on one end of the barbell are um the the AI labs who are consuming gobs and gobs of compute unlike anything I've seen and and then a you know a few runaway apps chat GBT cloud code things like that.

On the other end of the barbell are enterprises who are largely using AI um for cost avoidance and productivity. So automating customer service or fraud or business process automation. In the middle of the barbell are the current production workloads of enterprises. Um uh many of which are using AWS and and using inference today. But I would say the overwhelming majority of which still aren't in any pervasive way like you know and that I think is going to be the largest absolute segment of AI usage and I think you're going to see that. It's hard for me to predict the rate at which that'll happen. It's you know it's happening to a greater extent every year but I think if you zoom out five years five to seven years and and you look at what it'll be like and I don't think it's going to wait for five to seven years. It never happens in one fell swoop. But if you got if you just zoom out five to seven years, I think that um quite a few if not most of the enterprise applications that we use and that exist today are going to be using inference in a pervasive way.

chunk of 28 is sold out like there's so much demand that um you know it is uh and and you can't uh we have if you think about the AI business and what we call core business which think of core as non-AI they're both growing at really rapid rates right now and you know if if you have only a certain amount of memory you have to decide how much the memory you want to you know you want to allocate to to HBM you know to the to the AI accelerators versus um to um DRM you know into you know your core so it's uh these are real trade-offs you have to make but I would say that where we start on this issue is um being customer focused and for us uh we you know we have deep relationships and partnerships with our customers in AWS and in the same way during the pandemic, unlike our competitors who I think tried to figure out how to how to optimize within quarters to drive revenue as much as possible in a in a world where companies were going to cost optimize.

We took a different approach. We, you know, we proactively had our field go out and help customers figure out how they can cost optimize even though we knew it would diminish our revenue um short and medium-term. But you know, we're going to help our customers be successful because our very strong belief is if we help customers over a long period of time, our business will come along with it. And that's the same thing here on the core. You know, I guess you could do calculations and decide that you have customers because of the scarcity and and and their willingness to pay would pay more to squeeze out your core customers, but that they're running their businesses on top of AWS. They're running all their core applications. We're not going to squeeze our customers and we're going to make sure that they can scale. And then our job is to try to figure out how to have as much capacity as we can to allow not just our core customers but all of our customers who want to use us in the AI side and people want to use us in the AI side um to be able to do so and uh and do so in a way where you know it's it's sustainably economically for us as well.

And we just a few years ago we saw that same movie in the AI space where you really had one dominant provider um who's very good and we have a we had a deep partnership with we still have a deep partnership with we will for as long as I can foresee but they have giant markets and and and lots of um different products they're building and priorities and we knew our customers needed better price performance and that's why we we spun out a separate team in and aerna to build an AI accelerator which we call tranium which is also about 30 to 40% more price performant than the other comparable um um processors and you know I think when you talk about that middle of the barbell we were just talking about if you ask what are the what are the things that will change that middle of the barbell to have more of those current enterprise production apps using inference a really big there are a couple things you know some of it is you know some of the talent is comfort able doing post training and fine-tuning and some of the talent and organizational muscle that knows how to prioritize what to go after and how to do it.

But a big big chunk of it is the cost of inference. The cost of inference has come down quite a bit in the last few years, but it's still an order of magnitude too expensive in my opinion for it to be used as pervasively as possible. Because if you're asking enterprises to rewrite their applications that use CPU today with AI accelerators, if the difference in those costs are very significant, the amount of incremental benefit they have to get from having the AI accelerators on there is a much higher bar to want to rewrite those apps. But if you can take the cost of inference down, that bar is much lower and all applications are going to be better and smarter and more capable with AI accelerators and inference. So that is a really important part for customers to being able to use AI pervasively is driving the cost of inference down and our chips business is a big piece of that. Yeah. You know, it has I I can tell you from uh our own firm that it is it can get very expensive.

There are going to be different jobs and new jobs, some of which we know about, some of which we don't know about that will um arise as AI and robotics and space develop. Like, you know, they're going to be all sorts of of new jobs, but we're going to have fewer humans doing what they've done today. I think the teams will be um much smaller than they are right now. You know, I I think you're going to have smaller teams that have um they they have people who can work crossf functionally. They don't only one you they don't only own one sliver because the models are going to be able to get them 80% of the way there on a lot of these um different functions. I think that every team will start at a much more advanced stage than we have in the past. And so I think you're going to be able to invent much more expansively and much more quickly. I do think there's a bunch of things that have to change.

I think that uh you know the the startups starting today most of them are starting aentically and agent native you know um uh you know agent native or AI native right from the start. All the larger companies who've been in business for a while are going to have to figure out how to actually build at that same pace. And I think that I I I think that some of them will figure it out and um and you always like when you have a a a big business with a lot of capabilities and a loyal customer base. Some of them will move very slowly to figure it out and they will be more vulnerable in my opinion to startups. But you're going to I think you're going to have to I predict that companies will have to rewrite um almost all of their code to be agentically capable. um they're going to have to allow agents to be updating different pieces of code. You're gonna have to build constraints there. I think if you just look at security, uh you know, security is I it's kind of interesting if you look at all of the nervousness of the last month or two about models.

They are focused on the more fantastical scenarios. Um and by the way, they're possible. Like is it possible an evil botnet will take over you know everything? I mean it's I guess it's possible and it's possible you can infuse robots with um you know with malicious intent. They're all possible. I personally think we're going to figure that out together you know as an industry and private public segment. But you got to start by paying attention to some of the things. Yeah. But I mean the thing that's interesting is that I don't think that there is enough discussion about the more immediate term risks which are really around cyber security. These models are so they find so many more vulnerabilities than you did before and that humans you know weren't finding and and maybe you could argue if they could or couldn't but they just weren't finding. And so if you fast forward then, I thi

nk that the way that applications are going to have to be able to respond to imminent security threats is is going to be in minutes, you know, and today if you think about it, you know, you find vulnerabilities, people classify the vulnerabilities, people develop patches, and then even if you do that part quick, most companies take a long time to deploy because there are all these dependencies. You're going to have to deploy in minutes. And so if you don't actually have your applications rewritten to be able to be updated by agents, you're going to have no chance in in that new world. And so I just think, you know, all organizations are they're going to be reinvented. And I think the companies that prioritize are going to be much better shape than those that don't. Yeah. And it seems like, you know, we as an industry did not do a great job of architecting everything from the big systems to, you know, embedded devices with like good security architecture.