Even spending $220 billion this year, we're leaving a meaningful amount of demand on the table. We always leave a certain amount for on demand, but '26 is obviously sold out, '27 is largely sold out, a good chunk of '28 is sold out. The financial markets are saying, 'Oh, there's not enough money in the world to finance these things.' Everything we do in the world today — stock market, banking, healthcare — they're all using data centers. Our job is to try to figure out how to have as much capacity as we can. We have an opportunity to help across so many more customer experiences and make their lives better and easier every day.
The big labs being willing to pay almost anything for power, cooling, chips... well, maybe we should buy out the old guys and sell them to the new guys because they'll pay us whatever.
Andy, welcome to the podcast. It's great to have you.
So you're five years into running Amazon, and you've done a lot since being here, and a lot is coming out as CEO right now — $220 billion in capex spend, AWS grown 28%. There's a lot that you've achieved. How do you think about the ambition in 2026 as it relates to Amazon relative to 2021 when you took over formally as CEO?
Oh, 2021 feels so long ago. It feels like a decade ago. But as you know, the world changes fast, and if you compare it to 2021, our overall top line is about 80% higher, our AWS business is almost three times bigger, our ads revenue is about two and a half times bigger. So the business has just grown quite a bit, and I think there are a few things that have really happened.
I think pretty quickly after I started in the role, which was on the tail end of the pandemic, in our retail business we ended up with a real cost-to-serve challenge. We had basically doubled the size of our fulfillment network that we had built over 25 years — we doubled the size of it in about 18 months. And then we built the last-mile transportation network the size of UPS in about 20 months. And whenever you build that quickly, you end up with all sorts of things that you don't have optimized. And our cost to serve in our fulfillment network were not what we wanted. And it created a lot of challenges and angst.
A lot of people at that time — I remember it was just a year or so, or a little less than a year into starting the job — people said, 'Boy, what a terrible time to start and what bad luck.' And I think in retrospect, it was really a great time to start and a real blessing, because it allowed us to consider a bunch of things that we probably wouldn't have considered for many years, but they were really good ideas for our customers and for the business.
The biggest of which, in our fulfillment network, ended up being that we regionalized a flat national fulfillment network in the US. We regionalized it so we got items much closer to customers, which turned out to both lower our cost to serve because of shorter transportation distances, while getting items to customers much more quickly. And that re-evaluation of our most closely held beliefs during that time when we were struggling on cost to serve really unleashed speed of delivery and a lot of demand in our retail business, and a big piece of why we've grown as fast as we have there.
And of course, pretty quickly thereafter, just how AI exploded further on the scene. It's not like AI was new — I mean, we have been all doing machine learning for a long time, and there have been lots of waves of AI that were supposed to be the real tipping points. But you get to 2023 and you really see with generative AI a huge difference. And I would say our AWS business was growing really quickly at that point. We have the broadest functionality by a fair bit, and stronger operational performance and security, and I think a type of partner ecosystem that's different — we were the clear leader, and we felt like that business could be a few hundred billion dollars of revenue over a long period of time, figure over 10 years.
Now we look at what's happening with AI and how every customer experience we know of is going to be reinvented with inference and AI. We actually think the AWS business — not only is it growing really rapidly. I mean, last quarter we grew 36.7% year-over-year on a $169 billion annual revenue run rate business. Just the opportunities today are really striking.
And I think if you layer on top of that the seeds that we have of new businesses — whether it's how the team has progressed our Prime Video business, which is now a large and profitable business, or the low-Earth-orbit satellite we're building with Amazon Kuiper, or autonomous driving ride-hailing service with Zoox, or healthcare business, or what we're doing in robotics — we just have an opportunity, I think, to help so many more customers across so many more customer experiences and make their lives better and easier every day. And that mission of trying to make customers' lives better and easier every day really unifies everything we're doing across the company.
So, a couple of follow-ups on that. One, you kind of mentioned when you're building out the fulfillment network so fast that it's going to be a little inefficient — when you go that fast there's some sloppiness to it. And then you mentioned there was kind of angst around that internally. Like, from a CEO perspective, how do you communicate what you're doing without lowering the cultural standard of — you know, Amazon's known for its excellence and operating efficiency, being frugal, all these kinds of things — but here you are going, 'We're going to build UPS,' and so how do you reconcile those?
Well, what's interesting is, you know, we have this set of leadership principles that a lot of people know about — the 16 leadership principles — and a lot of these leadership principles, they're supposed to really, and they do, govern the way we think about operating as a team and culturally. But there are several that are in tension with one another, and sometimes people are confused about that. But the reality is it just reflects the way the real world works. Nothing is really perfectly clean. You're always making trade-offs.
So if you go back to that time when we made the decision to build out the fulfillment network and the last-mile transportation network that aggressively, that quickly, the question the team was facing is: we know if we go this fast it's going to be really hard. It'll probably be more expensive than if we went slower. We will probably not get everything as optimally as we would if we went slower. But customers and sellers really need it. So are we going to actually shade on the side of enabling?
Remember back then, in the early part of the pandemic, no physical stores were open. I mean, if you look at the role Amazon played in everybody's lives during the pandemic, it was very seminal. And so we had to ask ourselves: are we going to build fast and maybe not as optimally as we can, with the trade-off of serving customers and sellers who are trying to also have businesses that sustain themselves, or not? That's the first decision you have to make. And we decided, as we almost always do, to shade on the side of customers.
And then it was fairly obvious that as we were building, it was obvious that not everything that we were planning was going perfectly. We had a lot of — our fulfillment team was more in the office than most of the rest of the company, but still a lot of people that have to participate in a lot of the designing and the planning and the assessing. You had people all over the place and remote. And then now you're balancing the combination or the tension of bias for action or speed and relentlessly high standards, which means you deliver at the cost structure and the delivery and the design that you'd always planned for. And so again, you have to make trade-offs.
We made the trade-off to side with customers and to move quickly to be able to serve sellers and consumers, and we thought there'd be some amount of cleanup. Now, did we know exactly how the topography would work out in the way it did? No. And there were certain things that we didn't have optimized that I'm not sure we foresaw wouldn't be optimized. But if you have good metrics and if you're a culture that's a truth-telling culture — which is another really important part of our culture — we get the data on the table, we get the issues on the table, we're direct with one another, we don't allow people to paper over issues.
If you have that type of culture, it's fairly obvious in the metrics and it's fairly obvious in the discussion. And then you just have to go about figuring out how you want to improve it, and which pieces are just kind of small productivity things and which are architectural pieces. We went back and really rearchitected key parts of the network after we got that capacity installed — that was probably at the time for that team maybe eight or nine on their list of priorities, but became a much higher priority in a way to actually get the combination of speed and cost.
Yeah, fascinating. You know, kind of like at the beginning, when you started AWS, it was kind of 'build it and they will come' — you're building all this infrastructure and then you're going to rent it, and so they needed to come or you'd be in a lot of trouble. But this is that at a much bigger scale, in that you and others — but I would say primarily you, or at least you're certainly one of the biggest players — are building data center and chip capacity so fast that the financial markets are saying things like, 'Oh, there's not enough money in the world to finance these things.' And then we just kind of saw that skittishness in like the core credit debt market and equity markets and so forth.
And so how do you know, like, how much is too much on that build-out given the demand is so steep and you're growing so fast? And then there's a contention for that supply chain that you're kind of fighting for to get your cooling and power and chips and all that.
Yeah, it's a very good question, and I wish there was an equation that would make it simple. It would save us a lot of time as a team. I would say in the first wave of cloud, you know, when we started AWS, there was a little bit of — we didn't know if people would use it. So, build it and they will come. But we had very good signals inside the company, which at the time — this was around the early part of the 2000s — we had teams that were not moving nearly as quickly as we wanted them to, even though we felt like relatively we moved quickly, but for us we just couldn't understand why it took so long.
And it just turned out that our product teams were spending 70 to 80% of their time on the infrastructure pieces, which didn't differentiate them, and only about 20% of their time on what differentiated the product pieces. And we realized every team was doing it and reinventing the wheel. And so we originally conceived of AWS to allow ourselves to move more quickly internally. And we figured if we built a set of high-scale, high-availability, low-cost, highly flexible building-block services like compute and storage and database and content distribution, messaging and things like that, we would be able to build much quicker in our consumer business.
And then we just said we're pretty good relatively speaking technically — probably other companies have the same issues. But even before we launched externally, we had so many teams internally interested that we knew we were on the right track. I don't think any of us anticipated it would grow as fast as it did externally.
On the AI side, there were, as you know, these different waves of people saying, 'This is the time. AI is finally here.' And even maybe six, seven years before the generative AI wave, people felt like it was AI's time. We built a whole bunch of AI services that actually was a pretty good business for us on the AWS side, but nothing like what we're seeing right now. Like in the early days of AWS, we felt like we were growing really fast and we felt good about it, and we were about a $60 million annual run rate business after three years. In the first three years of this generative AI wave, our AI business in AWS is over $25 billion in run rate. So it's several hundred times faster. We just haven't seen anything like it.
And then the question becomes how much capex do you want to invest to go after the opportunity? And what I would tell you is, we've said this year in 2026 we expect to spend about $220 billion of capex. By the way, I'll just tell you that I remember being at a leadership team breakfast, you know, 15 years ago when we couldn't believe the telco company spent $50 billion in capex. We're like, how are they spending that much on capex? So these numbers are, you know, they're kind of mind-blowing numbers. And even spending $220 billion in capex this year, we're leaving a meaningful amount of demand on the table.
The thing that helps us here is that we have really good visibility into customer demand. And so if you think about 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 servers and networking. The data center piece, you have to invest 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. But you make that a couple years in advance.
The server and networking gear, 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 buy the server networking gear. On average, they tend to break even 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 five-year 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 those later stages, 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 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 — data centers, going back to when you started AWS, were probably the last thing in the world that you would believe would be a political issue. But here we are. And particularly in the United States, it's funny, and I'm sure you've experienced this, but nobody anywhere in the world is as anti-data center as the kind of US politics, for whatever reason. So how do you — how does that end up affecting this? I guess, is there any merit to the anti-data center argument, and then how does that kind of inform your strategy of what you build here, and what you build in the other regions?
Yeah. Well, I mean, I'll just give you some of the context that I have, or at least I believe, which is: if you think about everything we do in the world today — stock market, banking, healthcare, entertainment, shopping, education, business, ride-hailing — like everything we do is digitally driven. They're all using data centers in a very real way.
No, everybody is anti-me. I don't want to use my phone anymore.
And they all use data centers, and that's before AI. And AI is the biggest technology transformation of our time, which I think is saying something since we've lived through the internet and the cloud and mobile. And so the reality is the type of infrastructure build-out we need to support the rapid growth we were just talking about on the AI side — there hasn't been anything like it in the world, and in this country, since maybe when we built out the highways and bridges in the 1950s. I mean, that's the type of build-out we're going through.
And the country — I think the world in general, but I say the US specifically — cannot afford not to build data centers, or we will find that another country like China will establish all the leadership and all the norms there, and I don't think any of us are going to like that, either domestically or globally. So we have to build data centers.
Now, do I think there are ways that we can be doing it better, all of us as an industry, than we have? Of course. I think it's happening quickly and I think everybody's learning together. But there are a number of what I would consider somewhere between myths and mistruths right now being propagated, because it's one of those issues where everybody — you have an easy villain, and politically and sometimes opportunists like to kind of find that villain to point at.
But I'll give you examples. The golf industry or the almond farming industry consume way more water than data centers do. Just as a data point, golf courses consume 200 times more water than all our data centers, and they consume more in two days than we do in a year in our data centers. So there's just a lot of — that's a good example.
Or you can talk about pollution. What happens is when you register new sites and you have backup generators, you have to register and show the amount if you ran all of those generators full-time at full capacity. But of course that never happens. You only run backup generators 0.1 of 1% of the year, like 10 hours a year for maintenance.
Or you talk about the energy rates. If you look in the different geographies in which we have a big data center presence, some of the energy rates have gone up, some of the energy rates have gone down. But the real issue is that as a country we have not invested in the infrastructure of our grids the way we've needed to. There's going to be meaningful growth, and so something was going to push it over to investment, and it just turns out that AI, because it's growing so rapidly and so significantly, is the thing that pushed it over.
But building it out, of course, what you're going to find is once you actually get that capacity in most of the geographies, they're going to end up — the rates will flatline or go down, as they have in Indiana or in Georgia or Mississippi. And so there are things that we can do that are better.
I mean, the other thing people forget about is in these geographies, the big data center providers are paying a lot of tax. So if you look in St. Joseph County in Indiana, we've paid about $3 billion of tax. The communities get that money and they get to spend it on things that really matter for those communities. So the communities get something out of it, too. We've built programs like Built Together that we just committed to another billion dollars in the communities in which we have data center presence over the next five years.
And so I think we have to keep working on it with the communities. It has to be a public and private partnership, but the build-out is so important to the future of the country. Like, imagine if we hadn't built out highways and bridges — what the economic prosperity and the capabilities of the country would be. And we have to kind of keep it in perspective, and we have to be partners in the communities in which we build.
Let's go back to the capex for a second. The $220 billion — that's $600 million a day, just to put it in perspective. I'm curious how you think about AI demand right now. Offline, we were talking about how you see it as a barbell, and maybe you can sort of situate or contextualize the state of AI demand at the moment.
Well, the first context piece is what I said earlier, which is just I can't believe how big and fast AI has grown. 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 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 is.
And what I would say is, on one end of the barbell are the AI labs who are consuming gobs and gobs of compute, unlike anything I've seen, and then a few runaway apps — ChatGPT, Claude Code, things like that. On the other end of the barbell are enterprises who are largely using AI 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, many of which are using AWS and using inference today, but I would say the overwhelming majority of which still aren't in any pervasive way. 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 happening to a greater extent every year, but I think if you zoom out five years, five to seven years, 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 just zoom out five to seven years, I think that 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. And so there's just a lot of growth in front of us.
And how do you think about — because there's kind of this impending data center, chip, memory, everything shortage that's upon us in really 2027, just because everybody's building now but it's not going to be ready necessarily by the end of the year, and demand's not slowing down at all. And so when faced with — how do you think about it when faced with the big labs being willing to pay maybe anything or almost anything for power, cooling, chips, you know, kind of the environment?
And then you've got these kind of old, been-around-a-long-time customers who are paying their regular rates to run their businesses. Is there like any business pressure to go, 'Well, maybe we should buy out the old guys and sell them to the new guys because they'll pay us whatever, and the old guys — I can't double their prices, that won't be fair'? Or like, how do you think about that? Because it's such a weird thing, that you've got people you've had in there for 20 years doing a thing that's kind of, as you say, probably going to inference, and then you've got these new guys who are like insatiable — they'll buy anything.
Yeah, it's a good question. We don't think of it quite the way you're talking about, but I can completely understand. If you think about it, there's only so much data center and power today. There's only so much memory, and I don't think that shortage ends in 2027. I believe one of the reasons, coming back to the capital investment, why I think we could invest a lot more and it wouldn't be a bad use of capital, is that '26 is obviously at this point sold out, '27 is largely sold out. We always leave a certain amount for on demand in our inference service, Bedrock, or EC2, but largely sold out. All the big chunks — a good chunk of '28 is sold out. There's so much demand.
If you think about the AI business and what we call core business — think of core as non-AI — they're both growing at really rapid rates right now. And if you have only a certain amount of memory, you have to decide how much of the memory you want to allocate to HBM, to the AI accelerators, versus to DRAM, into your core. So these are real trade-offs you have to make.
But I would say that where we start on this issue is being customer-focused. And for us, 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 optimize within quarters to drive revenue as much as possible in a world where companies were going to cost-optimize, we took a different approach. 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 short and medium-term.
But 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. I guess you could do calculations and decide that you have customers because of the scarcity and their willingness to pay would pay more to squeeze out your core customers, but 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 on the AI side — and people want to use us on the AI side — to be able to do so, and do so in a way where it's sustainably economic for us as well.
And so I wish I could tell you that I thought that the supply constraints would be just a 2027 issue. I think we'll see them in different components for several years. I think that we're probably putting more capacity in than anybody else because we really are very passionate about continuing to be the significant leader in the infrastructure and AI space like we've been.
But as I said, we're spending $220 billion this year, and we could spend a lot more, and it wouldn't be a bad use of capital, both medium to long-term, and I think we still probably wouldn't satisfy all the demand. There is just an incredible demand right now.
So kind of speaking of that, you did something back in 2015 when you bought Annapurna Labs, which was probably before you were really thinking — I mean, maybe you were — about giant AI demand in 2015, because there just wasn't much AI demand in 2015. So kind of what drove that decision, and then how has it kind of paid off in this new world?
Yeah, I think the Annapurna team is one of the most talented teams and set of people I've ever had a chance to work with, and it's one of the very best acquisitions we've done at Amazon in the history of the company, if not the best. And we foresaw some pieces, but I don't think that we made the Annapurna purchase for AI chips per se.
When you get to a certain scale, you start to have access and visibility to patterns of customers at very large scale that you realize if you can vertically integrate and you could build certain capabilities yourself, you'll be able to provide more value for customers and better economics for the business itself.
So what Annapurna started as for us was just on the kind of core EC2 server. We felt like we were using too much of the server for virtualization and for networking and for security and for storage, and we wanted to be able to use as nearly 100% of the server for compute for customers. And so we had been working with Annapurna before we acquired them on this NIC that offloaded those components to the NIC to allow us to have as much of the server for compute as possible.
And we really liked the team, and we really liked the product that we collaborated with them on together. And we realized that was going to end up being in virtually every EC2 server, and it's turned out to be in every server we have. So we originally bought it because that was going to be a core capability of our compute capabilities, and because the team was so talented at building...
chips, we felt like they could probably build other chips too. And at that time we had this very deep partnership — which we still have a deep partnership with this company — but with Intel, Intel had virtually all of the CPU share. And when you have a partner who has virtually all the share, they're not often thinking about how to drive price performance. They're trying to think about other features that allow them to get more share and drive more revenue. By the way, it's totally understandable. These are some of the dynamics of how markets work. It's part of how the innovator's dilemma comes up. And so we had tried to push on price performance and it was hard for us to get attention there. But we had customers who wanted to move more and more of their on-premises workloads to the cloud and they wanted better price performance. So the next thing we turned that Annapurna team loose on was to build a CPU chip, and they built Graviton, which 99% of our top thousand EC2 customers use very pervasively today. It's about 40% more price performant than other x86 processors. And so that has been hugely impactful for our customers and for our business.
And just a few years ago we saw that same movie in the AI space, where you really had one dominant provider who's very good and 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 lots of different products they're building and priorities, and we knew our customers needed better price performance. And that's why we spun out a separate team in Annapurna to build an AI accelerator, which we call Trainium, which is also about 30 to 40% more price performant than the other comparable 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 things that will change that middle of the barbell to have more of those current enterprise production apps using inference, there are a couple things. Some of it is that some of the talent is comfortable 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 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 being able to use AI pervasively — is driving the cost of inference down, and our chips business is a big piece of that.
