Amazon CEO: AI business annual revenue exceeds 25 billion USD, growth rate is hundreds of times that of AWS in its early days, spending 220 billion USD still cannot meet demand

Recently, Amazon CEO Andy Jassy appeared on the well-known venture capital a16z's podcast, engaging in an in-depth conversation with a16z co-founder Ben Horowitz. In this interview, Jassy comprehensively revealed Amazon's strategic landscape in the wave of generative AI.

In response to Wall Street's concerns about the seemingly bottomless capital expenditures of tech giants, Jassy provided the most direct market feedback: demand has not only not slowed down but is rapidly overextending future capacity. At the same time, he detailed the current "barbell" demand structure in the AI market and clearly pointed out that to truly unlock enterprise-level AI applications, the cost of computational inference must drop by another order of magnitude.

"Even with $220 billion spent, we still left a significant amount of demand on the table"

The core issue currently most concerning the market is whether the giants' capital expenditures, often in the hundreds of billions, will turn into excess sunk costs. Jassy directly responded to market doubts with a set of astonishing order data.

"Even if we spend $220 billion this year, we still leave a significant amount of meaningful demand on the table." Jassy bluntly pointed out the current extreme supply-demand imbalance in computational power: "We always reserve a certain proportion of computational power for on-demand services, but capacity for 2026 is clearly sold out, 2027 is basically sold out, and a large portion of 2028 has also been sold out."

From the perspective of business growth, the growth curve brought by AI is extremely steep. Jassy compared the early development speed of AWS (Amazon Web Services) with the current pace: "In the first three years of AWS, we felt we were growing fast, with an annualized revenue run rate of about $60 million. But in the first three years of this wave of generative AI, our annualized revenue run rate for AWS's AI business has already exceeded $25 billion. Its speed is hundreds of times that of back then; we have never seen anything like it."

In addressing concerns from the capital markets, Jassy explained the cash flow cycle logic of data center investments. He stated that investments in data center construction need to be made two years in advance, but have a lifespan of up to 30 years; whereas the procurement of servers and networking equipment only requires a few months in advance. This means the company has very clear visibility into customer demand. "If you find that you actually don't need to fill those data centers, you won't buy servers and networking equipment. Most server contracts are five years, and typically break-even can be achieved in three years. Due to this cash cycle, when you grow at such a fast pace, early free cash flow will indeed face headwinds, but once everything is online, the returns will be very substantial."

Refuting data center controversies: "The water consumption of a golf course in two days equals our annual usage"

In the United States, the energy consumption and environmental issues of data centers are evolving into a political controversy. Jassy strongly rebutted this and elevated it to a macro level of national competitiveness.

"The infrastructure construction we are experiencing is comparable in scale to the construction of highways and bridges in the United States in the 1950s. The country and even the entire world cannot afford the cost of not building data centers; otherwise, we will find countries like China establishing all leadership positions and norms, which is unacceptable for us both domestically and globally."

Regarding environmental accusations, Jassy presented striking data facts: "There are many claims circulating that are somewhere between myth and lies because politicians and speculators like to find a simple 'villain.' In fact, the water consumption of golf courses or almond farming far exceeds that of data centers. As a data point: the water consumption of a golf course is 200 times ours; they use more water in two days than our data centers do in a year."

He also added that the pollution from backup generators is severely exaggerated: "You have to register pollution levels as if they are running at full capacity all the time, but in reality, we only run backup generators 0.1% of the time each year (about 10 hours) for maintenance."

The "barbell effect" of AI demand and the core logic of reducing inference costs

When dissecting specific AI demand, Jassy proposed a vivid "barbell" theory, which also explains why Amazon insists on a self-developed chip strategy.

"One end of the barbell is AI labs, which are consuming massive amounts of computational power; the other end is enterprises, mainly using AI to reduce costs and improve productivity, such as automated customer service or business processes. And in the middle of the barbell is the current production workload of enterprises. The vast majority of enterprises have not yet adopted (AI inference), but I believe that in 5 to 7 years, almost all existing enterprise applications will deeply utilize inference technology."

However, to activate the vast market in the "middle of the barbell," price is the biggest stumbling block.

"In the past few years, inference costs have decreased significantly, but in my view, it is still an order of magnitude too expensive." Jassy emphasized, "If you ask enterprises to rewrite their existing CPU applications using AI accelerators, if the cost difference is huge, the threshold for them to rewrite will be very high. But if you can bring down inference costs, the threshold will be greatly lowered. That's why we are focusing on our self-developed chip business."

Jassy highly praised the acquisition of chip design company Annapurna Labs in 2015, calling it "the most successful acquisition in Amazon's history, not even a close second." He revealed that through the self-developed Graviton processors (which are 40% more cost-effective than other x86 processors) and Trainium AI accelerators (which are 30-40% more cost-effective), Amazon is aggressively driving down the cost of computational power.

Organizational restructuring: 6 engineers did the work of 40 people in a year, all products must be rebuilt from scratch

In addition to outputting computational power externally, AI is internally reconstructing the operation of Amazon's massive machine, with personnel structure and product development logic being the first to change.

Jassy shared a shocking internal case: the Bedrock (Amazon's AI inference service platform) team restructured the inference engine with just 6 engineers in 76 days, whereas this previously required 40 people an entire year.

"This opened everyone's eyes." Jassy candidly stated, "In the future, there will be fewer humans doing the same work than today. I'm not saying we won't need people, but as the scale of the business expands, we won't invest as much manpower as we did in the past. Teams will become smaller, and people will need to work cross-functionally because models can help you complete 80% of the work."

At the product level, Jassy believes that incremental improvements are no longer viable; there must be the courage to "rebuild from scratch."

"For products like Alexa, its 'brain' was very innovative during development, but with the emergence of generative AI, the boundaries of capability have completely changed. If you want to have a winning customer experience, you must rewire Alexa's brain."

Not only Alexa, Jassy pointed out that all core consumer experiences at Amazon, including retail, Prime Video, and music, must be thoroughly reimagined in conjunction with AI in the coming years. "Teams often resist due to pride and historical baggage, saying 'we tried that, it didn't work.' But this (reconstruction) is a painful process that must be undergone; if you only think about incremental improvements, you will be eliminated in three to five years."

The following is the full translation of the podcast:
Andy Jassy 00:00

Even if we invest $220 billion this year, we still leave a considerable portion of demand on the table. We always reserve some capacity for on-demand use, but 2026 is clearly sold out, 2027 is basically sold out, and 2028 has a significant portion already sold out.

Andy Jassy 00:14

The financial markets are saying, oh, there isn't that much money in the world to finance these things. But everything we do in this world—stock markets, banks, healthcare—they all rely on data centers. Our job is to find ways to have as much capacity as possible; we have the opportunity to cover many more customer experiences, making their lives better and easier every day. Those large labs are willing to pay almost any price for power, cooling, and chips. Well, maybe we should buy out old customers and sell to new ones because they are bidding at our price.

a16z 00:48

Andy, welcome to the show. It's great to have you here. Thank you for the invitation. So, you've been at the helm of Amazon for five years, doing a lot of things. It's not easy to be here, and as CEO, there are many things progressing one after another—you currently have $220 billion in capital expenditures, AWS has grown by 20%, 28%. Well, you have indeed accomplished a lot.

Andy Jassy 01:09

You've achieved so much, so how do you view Amazon's ambitions for 2026—compared to when you officially took over as CEO in 2021? Oh, 2021 feels like a long time ago; it feels like, you know, like ten years ago. But, you know, the world changes so fast. And, you know, it's just...

Andy Jassy 01:30

If we compare to 2021, our overall revenue is about 80% higher, AWS business is nearly three times, and advertising revenue is about two and a half times. So this business, you know, has grown quite a bit.

Andy Jassy 01:47

I think a few things have indeed happened. Shortly after I took office—just as the pandemic was winding down—our retail business faced a real fulfillment cost challenge: we essentially doubled the scale of our fulfillment network. The network we built over 25 years was doubled in about 18 months; then, in about 20 months, we built a last-mile delivery network comparable to UPS. And when you build that fast, there will inevitably be all sorts of things that are not optimized. The fulfillment costs in our fulfillment network were not what we wanted. Well, you know, this brought many challenges and a lot of anxiety at the time.

I remember that I had been in office for about a year, not even a year, and everyone was saying: oh, taking over at this time is the worst timing, such bad luck. But looking back now, I think that was actually a very good starting time; it was truly a blessing—because it forced us to consider a lot of things that we might not have thought about for many years, and those things were very good ideas for our customers and our business. One of the biggest was our fulfillment network: we regionalized what was originally a flat national fulfillment network in the U.S. By regionalizing, we placed products much closer to customers, resulting in lower fulfillment costs due to shorter transportation distances, while also speeding up delivery to customers.

Well, during that phase when we were struggling with fulfillment costs, re-examining our most deeply rooted beliefs truly unleashed a lot of demand for delivery speed and retail business, which is also a key reason we were able to grow so quickly there. Of course, you know, right after that, AI further ignited the whole scene. It’s not that AI is something new—we’ve all been doing machine learning for a long time, and there have been several waves of AI that were considered "real turning points" before. But when you get to 2023, you really see the huge difference brought by generative AI. What I want to say is that by then, our AWS business was already growing very fast.

We have, you know, obviously the broadest functional coverage, stronger operational performance and security, and I think there is also a different kind of partner ecosystem. We are clear leaders, and at that time we felt that this business could achieve hundreds of billions of dollars in revenue over a long period of time—let's say over ten years. Now, looking at everything happening with AI, and seeing that every customer experience we know will be reinvented because of reasoning and AI, we actually believe that the AWS business is not only growing very fast—last quarter we grew 36.7% year-over-year, and that’s on an annualized revenue base of $169 billion—but the opportunities in front of us today are truly amazing.

And you know, I feel that if you layer on top of that, we have those seeds of new businesses: whether it’s how the team has developed Prime Video—it’s now a large and profitable business; or the low Earth orbit satellites we are building with Amazon Leo; or the autonomous ride-hailing service we are doing with Zoox; as well as our healthcare business and what we are doing in robotics. You know, we really have the opportunity to help many more customers, covering many more customer experiences, making their lives better and easier every day.

And this mission of "making customers' lives better and easier every day" truly unifies what we do across the company. Now, I have a few follow-up questions about this. First, you just mentioned that when you rapidly expand the fulfillment network, it inevitably becomes a bit inefficient. You know, when you push so fast, there will definitely be some rough edges. So how do you—and you also mentioned some anxiety internally around this—how do you communicate what you are doing from a CEO's perspective while not lowering cultural standards? I mean, Amazon is known for excellence, operational efficiency, and frugality, all of these things. But now you are saying: we want to build something like UPS, and so on. How do you reconcile these two?

Well, the interesting thing is that we have this set of leadership principles that many people know, a total of 16. Many of these principles should and do govern how we think about operations and culture as a team.

But several of them have tensions with each other, and sometimes people get confused about that. The reality is that this reflects how the real world operates—nothing is ever truly clean-cut; you are always making trade-offs. So going back to that moment when we decided to expand the fulfillment network and last-mile delivery network at such an aggressive and rapid pace, the challenge the team faced was: we knew that pushing this fast would be very difficult; costs would likely be higher than if we did it more slowly; and we might not be able to optimize everything as well as we would if we took our time. But customers and sellers really needed it. So, should we lean towards "just get it done first"?

You have to remember, that was early in the pandemic when all the physical stores were closed.

I mean, if you look at the role Amazon played in everyone's lives during the pandemic, it was quite groundbreaking. So we had to ask ourselves: should we build quickly, perhaps not as optimized, in order to serve those customers and sellers—those sellers who were also struggling to keep their businesses afloat? That was the first decision we had to make. And we decided, as we almost always do, to lean towards the customer side. You know, then, during the building process, it quickly became clear that not everything we planned was progressing perfectly.

We have, you know, our fulfillment team is in the office more often than most other departments in the company. But even so, there are a lot of people who need to be involved in a lot of design, planning, and evaluation. People are scattered everywhere, working remotely, and then you have to balance the combination or tension between "bias for action" or speed and "maintain extremely high standards"—the latter means you have to deliver according to your planned cost structure, delivery methods, and designs. So you have to make trade-offs again.

Our trade-off was to stand on the customer side, move quickly, in order to serve sellers and consumers well, and we knew there would be some cleanup work later on. So, did we know at that time that it would evolve into what it became later? No. There were also some things we did not optimize properly, and I am not sure we foresaw that they would not be optimized. But if you have good metrics, if your culture is one that speaks the truth—this is also a very important part of our culture—you know, we put the data on the table, we put the problems on the table, we communicate directly with each other, and we do not allow others to gloss over problems. If you have such a culture, then it becomes quite clear in the metrics and in discussions.

Then you just have to figure out how to improve it, distinguishing between what are just small efficiency issues and what are structural issues. Once we had that capacity in place, we went back and really restructured key parts of the network—that was probably around the eighth or ninth priority on that team’s list at the time, but later became a much more important priority because it could really achieve a combination of speed and cost. Yes, it’s interesting. You know, it’s a bit like when you first started AWS—you started AWS, and the approach back then was a bit like "build it and they will come," right? You built this whole infrastructure and then you operated it. So they had to come, or you would be in big trouble.

But now this is playing out on a much larger scale: you and some other companies—though I have to say, mainly you, at least you are definitely one of the biggest players—are building data centers and chip capacity at such a rapid pace that the financial markets are starting to say: wow, there isn’t that much money in the world to finance these things. And then we recently saw that kind of tension in the debt and equity markets of CoreWeave and so on. So how do you judge—considering the demand is so steep and you are growing so fast—how much is too much for this round of construction? And there is competition in the supply chain; you have to compete for your cooling, power, chips, all of these things. Yes, that’s a very good question. I really wish there were a formula to simplify it, that would save our team a lot of time.

Well, I think I would say this: in the first wave of the cloud, when we started AWS, there was indeed a bit of "we don’t know if people will use it, so let’s build it and they will come." But we had good signals internally. At that time, you know, around the early 2000s, we had quite a few teams whose progress was far below our expectations—even though we felt we were fast enough relatively speaking. But we just couldn’t figure out why it was taking so long. The reason was that our product teams were spending 70% to 80% of their time on the infrastructure part of the work, which didn’t allow them to differentiate; only about 20% of their time was spent on the parts that truly made the product different.

So we, you know, realized that every team was repeating this and reinventing the wheel. Therefore, we initially conceived AWS to allow ourselves to run faster internally. Our judgment at the time was: if we built a set of large-scale, highly available, low-cost, highly flexible foundational module services, such as computing, storage, databases, content delivery, messaging, etc., we could build much faster in consumer business. Then we just said, relatively speaking, our technology was decent, and other companies probably had the same issues. But even before we launched it externally, there were so many teams interested internally that we knew we were on the right track. I don’t think anyone among us anticipated it would grow that fast externally.

On the AI side, you know, as you know, there have been waves of people saying: now is the time, AI has finally arrived. Even about six or seven years before this wave of generative AI, people felt that it was the era of AI. We built a lot of AI services, and it was actually a pretty good business on AWS. But it cannot be compared to what we see now. In the early days of AWS, we felt we were growing very fast, and we felt good about it—three years later, we achieved about $60 million in annualized revenue. And in the first three years of this wave of generative AI, just our AI business in AWS has annualized revenue exceeding $25 billion. So it has grown hundreds of times faster. This, we have really never seen anything like it. So the question becomes: how much capital expenditure do you want to invest to chase this opportunity? I can tell you that we have said we expect to spend about $220 billion in capital expenditures by 2026.

By the way, I remember about 15 years ago at a leadership team breakfast, we couldn’t believe that a telecom company would spend $50 billion on capital expenditures; we said: how can they spend so much money on capital expenditures? So these numbers, you know, are truly mind-boggling numbers. And even if we are going to spend $220 billion in capital expenditures this year, we still left quite a considerable portion of demand on the table. What really helps us here is that we have good visibility into customer demand. So if you look at the capital expenditure investments of AWS and how the cash cycle operates, it has two major components: one is investment in data centers, and the other is investment in servers and networks. For the data center part, you have to, you know, invest about two years in advance before you can put servers in the data center and start monetizing. The good news is that this investment only needs to be made once; it is a long-lived asset, about 30 years.

But you have to make this investment a few years in advance. The servers and network equipment only need to be invested a few months in advance, depending on the specific components. So before you really pull the trigger, you have good visibility into customer demand. If you find out that you actually don’t need that capacity to fill those data centers, you won’t buy servers and network equipment. On average, they, you know, can break even in about three years, and currently their lifespan is about five to six years.

Most of the deals we do are about five years. So when you are building so many data centers at such a fast pace like we are, you will spend a lot of capital, and your free cash flow will face headwinds in those years—until a few years later when you reach that later stage: all these data centers are online, all monetizing, and you are making money back from your investments in servers and networks. But due to the way the cash cycle operates, in the first few years of growing so fast, it does create headwinds. So following this, you know, going back to when you started AWS, data centers were probably the last thing in the world you would believe would become a political issue. But now we are in such a situation, especially in the United States.

You know, it's quite interesting, and I believe you have experienced it too—there's no place in the world where people are more opposed to data centers than in the American political sphere, regardless of the reasons. So, you know, how will this situation ultimately affect this matter? What I want to ask is, does the argument against data centers have any merit? And how will this in turn affect your strategy—what you build here, what you build in other regions? Well, okay, I can give you some background that I have, at least what I believe. Think about it, everything we do in this world today—stock markets, banking, healthcare, entertainment, shopping, education, business, ride-hailing—everything we do is data-driven, and they all rely on data centers, plain and simple. No, everyone does; I oppose myself, I no longer want to use my phone.

But they all use data centers. And this is just before AI. AI is the biggest technological transformation of our time—considering we've gone through the internet, cloud, and mobile internet, this is a significant statement. So the reality is, to support the kind of rapid growth we just talked about on the AI side, the infrastructure we need has not been seen in this world, in this country, since we built highways and bridges back in the 1950s. I mean, what we are experiencing now is that level of construction. And this country—I think the whole world is generally the same, but I want to specifically mention the United States—cannot afford the cost of not building data centers; otherwise, we will find that another country like China will establish all leadership and all rules there. I don't think any of us would like that, whether domestically or globally.

So now we must build data centers. Now, do I think that as an entire industry, we have ways to do better than we are now? Of course. I feel things are developing rapidly, and everyone is learning together. But right now, there is a batch of—I would call it something between myth and misinformation—that is being spread because it's the kind of issue where everyone can find a ready-made "villain." Politically, sometimes opportunists like to point out that villain. I can give you some examples. The water consumed by the golf industry or almond farming is much more than that of data centers.

Just to give a data point: golf courses consume 200 times the water of all our data centers; their water consumption over two days is more than our data centers' annual water consumption. So there are many such examples, and this is a good one. Or you can talk about pollution. The reality is, when you register a new site and equip backup generators, you must register and declare: if you run these generators at full load, 24/7, what would the emissions be? But of course, that never happens. Backup generators only run a fraction of the time, like 10 hours a year for maintenance.

So, you know, or you talk about electricity prices—if you look at the different regions where we have large-scale data center layouts, some places have seen electricity prices rise, while others have seen them fall. But the real issue is: as a country, we have not invested in our grid infrastructure as we should. There will inevitably be significant growth, so something will push it to the critical point where investment is needed. The result is AI—because it is growing so rapidly and to such a scale—has become the thing that pushes it over. But once you really build the capacity, you will find that in most regions, electricity prices will eventually level off or even decrease, just like what has happened in Indiana, Georgia, or Mississippi.

So there are some areas where we can do better. Another point that people easily forget is: in these regions, large data center providers pay a significant amount of taxes. For example, in St. Joseph County, Indiana, we have paid about $3 billion in taxes. The community has received this money and can spend it on things that truly matter to these communities. So the community also benefits from this. We have built some projects, like Built Together, and we just committed to investing another $1 billion in the communities where we have data center layouts over the next five years. So I believe we must continue to work with the community to advance this matter. It has to be a public-private partnership. But this round of construction is too important for the future of this country—imagine what would have happened if we hadn't built the highways.

a16z 20:44

And bridges—that would determine the economic prosperity and capacity boundaries of this country. So we need to view it in a reasonable perspective, and we must become partners with the communities we are building in. Let's go back to the issue of capital expenditure. That $220 billion, which is $600 million a day. I just want to, you know, put it in a reference frame. I'm curious how you view AI demand now—you described it as a...

Andy Jassy 21:10

Barbell structure. Maybe you can outline or position the current state of AI demand. Well, I want to say the first background is what I mentioned earlier—I can hardly believe how big and fast AI has grown. You know, it's hundreds of times faster than the early days of AWS; in the first three years, we achieved an annualized scale of $25 billion in AI revenue in AWS. So it is really very large. But at the same time, if you look at the structure of demand, it really is a barbell. One end of the barbell is those AI labs, consuming massive, massive amounts of computing power, at a scale I have never seen before. Then there are a few blockbuster applications, like ChatGPT, Claude Code, and so on.

Andy Jassy 21:59

The other end of the barbell is enterprises, which mainly use AI for cost avoidance and efficiency improvement, such as automating customer service, fraud detection, or business process automation. The middle of the barbell is the current production workloads of enterprises, many of which run on AWS and are already using inference today. But I want to say that the vast majority of them have not been used in any penetrative way—you know, and I think that will become the absolute largest piece of AI usage. I think you will see that.

Andy Jassy 22:38

I find it hard to predict how fast this will happen. It, you know, happens to a greater extent every year. But I feel if you pull back the perspective—five years, five to seven years—and look at what it will be like; and I don't think it will wait until five to seven years to happen, it never happens all at once. But if you can pull back the perspective to five to seven years, I think a considerable amount, even the majority of the enterprise applications we use and exist today, will be used in a penetrative way for inference. So there is still a lot of growth space in front of us.

So how do you view this? Because we are facing an impending shortage of data centers, chips, memory—all things—really in 2027, because everyone is building now, but they may not be ready by the end of this year, and demand is not slowing down at all. So how do you think about this situation—when those large labs are willing to pay perhaps any price, almost any price, for electricity, cooling, chips; and you have these long-standing old customers paying at regular rates to run their businesses. Is there a commercial pressure that makes you think: well, maybe we should buy out the old customers and sell to new customers because they are bidding along with us? And the old customers, I can't double their prices, that wouldn't be fair. Or, how do you view this? Because it is really strange—you have a group of people who have been with you for 20 years, doing something that, as you said, is probably going towards inference; then you have these new customers who are greedy and willing to buy anything.

Andy Jassy 24:35

Yes, that's a good question. You know, the way we look at this is somewhat different from what you mentioned, but I completely understand. If you think about it, today there is only so much data center and power, and only so much memory, and I don't think this shortage will end in 2027.

I believe—going back to the topic of capital investment—one of the reasons I feel we can invest much more, and that it won't be bad capital usage, is that 2026 is clearly sold out at this point, and 2027 is basically sold out. We will always leave some for on-demand use, whether in our inference service Bedrock or in EC2—but basically, everything is sold out, all the big chunks are sold out. A significant portion of 2028 is also already sold out. Demand is that large; you can't... If you look at AI business and what we call core business—the core can be understood as the non-AI part—they are both growing at a very fast pace now.

And you know, if you only have that much memory on hand, you have to decide how much memory to allocate to HBM, that is, for AI accelerators, and how much to allocate to DRAM, that is, for your core business. So these are real trade-offs you have to make. But I want to say that our starting point in this matter is customer-centric.

For us, we have deep relationships and collaborations with customers in AWS. Just like during the pandemic—at that time, our competitors, I think they were thinking about how to optimize within a quarter to maximize revenue in a world where everyone is trying to optimize costs; while we took a different approach. We proactively sent our teams out to help customers find ways to optimize costs, even though we knew this would weaken our short-term and mid-term revenue. But you know, we want to help customers succeed because we firmly believe: if we help customers in the long term, our business will follow. The same goes for the core business. I guess you can indeed do some calculations and determine: due to scarcity and customers' willingness to pay, you can squeeze a bit more money from core customers. But they run their businesses on AWS, and all their core applications run on it.

We will not squeeze our customers; we will ensure they can scale. Then our job is to find ways to have as much capacity as possible, so that not only our core customers but also all customers who want to use AI with us, and those who want to do AI with us, can use it, and do so in a way that is economically sustainable for us. So, I really wish I could tell you that the supply tightness will only be an issue in 2027; I think we will see it across different components for several years. We have always been aggressive, and I believe we are putting out more capacity than anyone else because we are truly committed and very focused on continuing to be a significant leader in infrastructure and AI, just as we have always been.

But as I said, we are going to spend $220 billion this year, and we could have spent much more. That wouldn't be bad capital usage. Moreover, looking from a mid-term to long-term perspective, I think we are still likely to not meet all the demand. The current demand is simply astonishing. Continuing on this topic, you did something in 2015—you acquired Annapurna Labs. At that time, you probably weren't thinking about the huge AI demand that would arise in 2015, as the AI demand back then was quite limited. So what drove the decision to acquire? And how has it generated returns in this new world? Yes, I feel that the Annapurna team is one of the most talented teams and groups of people I have had the privilege to work with; it is one of the best acquisitions we have made in Amazon's history, if not the best.

Yes, we foresaw some parts of it, but I don't think we acquired Annapurna specifically for AI chips. Here's the thing: when you reach a certain scale, you can access and observe customer behavior patterns at a massive scale, and you realize that if you can do vertical integration and build certain capabilities yourself, you can provide more value to customers and also bring better economics to the business itself. So the starting point for Annapurna with us was just around the core EC2 servers: we felt we were spending too many server resources on virtualization, networking, security, and storage, and we wanted to use almost 100% of the server resources to provide computing for customers.

So before we acquired them, we were already working with Annapurna on this network interface card (NIC)—offloading these components onto the NIC so that we could leave as many servers as possible for computing. We really liked this team and the product we created together. We realized that what we were going to buy would ultimately be installed in almost every EC2 server—and indeed, it was installed in all our servers.

So we initially acquired it because it would become a core capability of our computing power; also because this team was so talented at chip-making, we felt they could probably make other chips as well. At that time, we had a very deep partnership with Intel—we still maintain a deep collaboration with that company today. But the problem was that Intel held almost all CPU market share at that time; and when you have a partner that holds almost all the market share, they often don't think about how to improve cost-effectiveness; they think about other features that can give them more market share and more revenue.

By the way, this is completely understandable; it's part of the dynamics of how markets operate and part of the reason for the "innovator's dilemma." So we wanted to push for cost-effectiveness, but it was hard to get their attention. However, we had customers wanting to move more and more local workloads to the cloud, and they wanted better cost-effectiveness. So the next thing we had the Annapurna team do was to create a CPU chip. They produced Graviton—today, 99% of our top 1,000 EC2 customers are using it extensively. Its cost-effectiveness is about 40% higher than other x86 processors, which has had a huge impact on our customers and our business. Then a few years ago, we saw the same "movie" replay in the AI field: there was indeed a very strong dominant supplier. We had a deep partnership with them—we still have a deep collaboration, and I expect we will continue to do so for the foreseeable future.

But they had a huge market and many different products and priorities they were working on. And we knew our customers needed better cost-effectiveness, which is why we set up another team within Annapurna to work on AI accelerators, which we called Trainium, and its cost-effectiveness is also about 30% to 40% higher than comparable processors. And you know, when we talk about the middle part of the barbell—what we just discussed—if you ask: what things will change the middle of the barbell to enable more current enterprise production applications to use inference? There are a few things. Part of it is talent: some people are already accustomed to post-training and fine-tuning, and there is a need for organizational capability that knows how to prioritize, what to tackle, and how to do it.

But a large part lies in: inference costs have come down quite a bit over the past few years, but in my view, it is still an order of magnitude too expensive to be used as widely as possible. Because if you want enterprises to rewrite applications that currently run on CPUs to use AI accelerators, if the cost difference between the two is very significant, then they must achieve very high incremental benefits from the AI accelerators to be willing to rewrite those applications—that's a much higher threshold. But if you can bring down inference costs, that threshold becomes much lower, and all applications will become better, smarter, and more powerful because of AI accelerators and inference. So, a very critical link for customers to widely use AI is to bring down inference costs, and our chip business is a significant part of that. Yes, indeed. I can tell you from our own company that it can really become very expensive. We had an employee who spent $2 million over the past few years—doing valuable work.

But yes, the feeling at that time was: well, inference is really not cheap. So, one thing you did was make some investments, particularly in Anthropic and OpenAI, with very large investments; I recall $18 billion for Anthropic and about $50 billion for OpenAI. So what made you want to do these? And these types of investments, along with similar investments in Nvidia and others, have made Wall Street a bit nervous because it reminds them of the bad old days of the internet bubble. How do you view this—especially considering they are also your customers? Why is this reasonable? Both of these companies are outstanding teams and very impressive companies; the rate at which they are growing their businesses—at least on the revenue side—I don't know if we've ever seen anything like it. Really, it's just incredible.

So for us, it's a combination of several factors: what kind of teams these are, the technology and models they create, and some of the applications they produce; we believe in how transformative AI can be; we believe they will be two of the most successful companies in this field for a long time—of course, not everyone in this field will be a winner, but many very successful companies will emerge, and we believe they will be two of them. Additionally, we collaborate so closely. Let me start with Anthropic—we have been collaborating with them longer. We have many customers running Anthropic in our inference service Bedrock. And I believe that customers running Anthropic in Bedrock are the most successful. We have a deep collaboration with them on chips—they have been building on Trainium for some time. I think this is very good for Anthropic and very good for our Trainium team; I believe both sides have learned a lot from each other. So we have a deep collaboration—whether it's our belief in the future of AI, our belief in these companies, or the collaboration itself, it makes sense for us to invest as part of that partnership.

Then on the OpenAI side, we just have so many customers wanting to use OpenAI on AWS but couldn't do so due to previous business arrangements. And for a long time, I think both companies have had a lot of respect and willingness to collaborate with each other. So this is also the situation: after a series of conversations lasting several months—being able to run OpenAI in Bedrock; we created a unique product together, which is the stateful runtime environment, and I believe all intelligent agents will be built this way in the future; you will need state; and they, like Anthropic, made multi-year, multi-gigawatt commitments to Trainium. And we believe it will be one of the most successful companies in this field for a long time. All of this makes us feel that deepening collaboration and investment is reasonable. Changing the topic, let's talk about how you use AI to build a company. I remember you said that the Bedrock team rebuilt your inference engine with 6 engineers in 76 days, which would have originally required 40 people working for a year. That's quite a difference.

And you know, Amazon is known for its "two pizza teams." So now, what is your organizational design thinking? Considering this technology—it really seems to change everything. Over the years, we have become extremely specialized: we have designers, product managers, front-end engineers, back-end engineers, and so on. And now it seems everyone has become a generalist again. So from an organizational design perspective, how do you view this? Yes, I agree with what you said. I think the Bedrock experience you mentioned was very eye-opening for everyone here.

We also spent a lot of time sharing that story internally to inspire people and remind them of what is possible. Because over time, such things will happen—I've been with the company for 29 years. I started as an assistant product manager, and at that time, I was responsible for, I don't know, five products by myself. And today, as the company succeeds, scales up, and products grow, you reach a state where a product manager only works on one product and only owns one slice—like they are only responsible for pricing, etc. Everyone has become more specialized, and everyone accepts: if you want to grow and succeed, this is the way to do things. But this is not the case, and when you look at everything AI is doing, it is absolutely not.

Take just the past week, for example; we are doing some things around intelligent agents in consumer business. We have a lot of investment there, but we had a very exciting idea, and we decided on Thursday to try to see if we could create a truly functional prototype. The result? Four or five of us got to work over the weekend, and by Sunday, we had it running at 80% release-ready status. This is true—it's completely two different worlds. So I think what this means for the future of organizations: first, I believe that in the future, there will be fewer humans doing the work we do today. Just look at how much coding is being done by intelligent agents, and see how much customer service, business intelligence, and research work is being done.

Yes, think about it as you fast forward: how will the tables be created, how many lawyers will you need------I just believe that you won't become devoid of people, but you won't need as many humans; in the past, as businesses grew, we kept piling people in, but the future won't be like that. Different jobs and new jobs will emerge, some of which we know, and some we don't yet know. With the development of AI, robots, and space, they will come forth. There will be all kinds of new jobs, but there will be fewer humans doing today's tasks. I believe teams will be much smaller than they are now. You will have smaller teams where people can work cross-functionally; you won't just let one person own a small slice of it because the model can help you get to 80%. In many of these different functions, I think every team will start from a much higher baseline than in the past, so I believe you will be able to invent more broadly and quickly.

I do believe there are a lot of things that must change. Most of the startups formed today are starting out in an agent-like or AI-native way from the beginning. And all the mature companies that have been operating for a while must figure out how to build at the same speed. I think some of them will figure it out. And you always have this situation: when you have a large business with many capabilities and a loyal customer base, some companies will be very slow to figure it out, and in my view, they will be more vulnerable when facing startups.

And I predict that companies will have to rewrite almost all their code to give them agent capabilities; they must allow agents to update different code snippets, and you have to establish constraints there. I think just looking at security is very interesting. If you look at the various tensions around models in the past month or two, they have all focused on those more fantastical scenarios.

By the way, those scenarios are possible: Is it possible for a malicious botnet to take over everything? I mean, I think it's possible; is it possible to inject malicious intent into robots? It's all possible. Personally, I believe that as an industry, along with the public and private sectors, we will work together to solve this issue. But you have to start by focusing on certain things. Yes, but the interesting part is that I think the discussion around those more recent risks is not enough------those risks are mainly around cybersecurity. These models are so powerful that they can find many more vulnerabilities than humans have ever found in the past. You might argue whether humans could find them, but the fact is they didn't. So fast forward, I believe the way applications respond to urgent security threats must be measured in minutes.

Today's situation is: you find a vulnerability, someone categorizes it, someone develops a patch; and even if this part is done quickly, most companies take a long time to deploy because of various dependencies. In the future, you must complete deployments in minutes. So if you haven't truly rewritten applications to be updatable by agents, you will have no chance in that new world. So I believe all organizations will be reinvented, and those companies that prioritize doing this will be in a much better position than those that do not.

Yes, and it seems that as an industry, we haven't done a great job of architecting everything------from large systems to embedded devices------with security architecture in mind. The reckoning day has come. But I think the good news is that this presents a huge opportunity for new companies, new attempts, and people to do things. But yes, as we said, this threat------those theoretical, thought-experiment threats, well, they might happen, or they might not. But this is clear and real.

Yes, almost all the partners and customers we talk about on the AWS side, regardless of size, every company is concerned about security. You know, we have a professional services business helping customers with this, and we built a service called Continuum that can identify a large number of these vulnerabilities and generate patches. But this thing, you know, is easier said than done. In fact, because many companies and many engineers sometimes feel that focusing on security isn't that exciting, it's easy to wait until something goes wrong. But if you are protecting important data, that's a bad idea. Moreover, I think in most companies------going back to our earlier conversation------specialization in security is probably the worst scenario: you have these very esoteric security engineers who don't like to work with other engineers, they are scarce resources, and so on.

The result is that in most companies, most software companies, their security practices when writing software are not very good. I think this clearly must change. The good news is we have tools that can help solve this problem. Following what you said, the good news is: you have to hire penetration testers, but the number of security personnel you can hire is limited, and they can only cover so many applications. Models combined with automation will allow you to better conduct penetration testing across the entire attack surface, and you will find some things that you might not have discovered without good luck. So when we observe this market, one interesting thing is: with AI, you can do a lot. One way is to use it to cut costs, reducing the number of accountants or lawyers you need.

Now, there are indeed quite a few venture capital firms saying: well, we are going to buy those old companies and cut their costs. This has never really interested us because we think: well, if that's your idea, then your idea is just too limited. Another way is: well, we want to create something the world has never seen before. That is very exciting. But there's a third way, which is doing something like Alexa: you have a product------a successful product------and then you say, well, let's completely rethink this thing from scratch because we have this new platform. How do you get to that point? Under what circumstances does it make sense? Yes, I think you draw an interesting tripartite distinction. And I believe, as we just discussed, organizations will truly change and become more streamlined.

If you do it right, it will allow your truly great builders and these small teams to create what you refer to as the second type of things------those things that people have never seen before, you can achieve that. And when you have the kind of experience I mentioned earlier------getting a new idea to 80% in four days------this is incredibly motivating for people, and they will really spend more of their spare time and more of their idle thinking cycles coming up with new ideas. For me, this is one of the most exciting parts of what is happening with AI.

But I also believe that for those existing companies that already have successful customer experiences at scale, if they primarily understand AI as just a way to incrementally improve experiences, then three to five years from now, they are likely not to have those winning experiences. Because I believe that every experience we know------even those that seem the least likely to be changed------will be significantly different in a few years because of AI.

So this means that if what you want is not just incremental improvement, you have to find a way to have the team reimagine what these applications should look like from scratch------assuming you build it today with all the new technologies available. In some cases, like with Alexa, I think it is quite obvious that we must do this. The "brain" of Alexa was indeed very creative at the time it was created. But due to generative AI, the capabilities that this "brain" can possess are completely different now, so if you want to provide an excellent customer experience and want to remain relevant in the long term, you must rewire Alexa's brain. So that is quite obvious. By the way, even though it is quite obvious, it is actually very difficult------because teams have an emotional connection to what they create, they take pride in what they build, and they have hard-earned scars of understanding what works and what doesn't.

Getting them to completely start over and reimagine requires time because teams will say: well, we don't do it that way; we tried that, and it didn't work. Or something like that. But I think what's even harder is: taking our other consumer experiences------like our retail business, or Prime Video business, Fire TV business, Prime Music------in those experiences, the technology has much better solutions today, which is not so obvious. But I actually believe that in each of these businesses, we need to reinvent them; and we already have teams working on thoroughly reimagined versions of these businesses.

So, do I think we can get it right from the start? Probably not. We will find ways to iterate our way in, but it is already generating all kinds of new ideas and customer experiences that we find very attractive. Then the interesting race will be: do you want a completely different experience------almost like a new application------or do you want those new things to advance quickly enough that you can integrate them into the existing experience, gradually and substantively changing the overall customer experience over the next two to three years? As you find the elements that work. But I do believe that if you want to be a successful leader, every one of these experiences must become different in the coming years.

Yes, this is a very tricky leadership challenge. It reminds me of when I first became an engineer. I was arguing with someone about something. My boss said: listen, Ben, you have to understand one thing------before you write the first line of code, everything is up for discussion; after that, it becomes a belief. I guess you experienced this with Alexa as well. But what's interesting is: all that knowledge, all that experience, all those things used to build the original thing------now you are saying, well, we are willing to give up a large part of it to move into this new world.

And by the way, it's the same for customers: you will have a large group of customers who love the current experience, right? Yes, every major transformation------at least in my experience, maybe you can cite examples that are not like this------but almost every major transformation requires a huge change: technology, product, customer experience, and business elements must be compelling enough to make people want to change. But even so, it won't happen automatically. This is a massive change management------just like moving from on-premises deployment to the cloud, it is the same kind of transformation: many infrastructure leaders take great pride in what they have built, they know it inside and out, they are at the center of action, and they don't want to learn new things.

It will be the same thing, just like the experience I mentioned earlier. The same will apply to reinventing all your customer experiences. We all take pride in what we have built over the past decade in these experiences, we all have 100 things that could make it better, we are all excited about it, and we spend all our waking hours on it. To think "how do I completely start over," you know, is difficult both rationally and emotionally. Yes, no doubt.

a16z 53:06

Also, Andy, we just incubated Harwood's and Jason Academy a few weeks ago to help young people prepare for this transformation we are talking about. You mentioned that AI makes building easier, but young people still need to learn, you know, real expertise. I'm curious: if you were 18 years old in 2026, what advice would you give yourself to prepare for this transformation we are entering?

Andy Jassy 53:29

Well, I have a 25-year-old and a 22-year-old child. So I think that 22-year-old child really started to take off when they were 19. I quickly realized: you can't tell them not to use AI in education. That ship has already sailed, or that train has already left the station. And I think the wise thing for universities to do is not to fight it. I think it's just a tool that people will use.

Andy Jassy 53:54

But at the same time, I think all young people— and I want to say the same for us, we have a lot of people, we have a writing culture, and everyone is using AI to write things— you can tell what is written by AI: it just doesn't quite fit, those parts are not stitched together. And you know, so my advice to young people— which is also my advice to my children— is.

Andy Jassy 54:18

You can't escape it, you know, and you can't expect it to go away. It will have a huge impact on your life, on every experience you try to use, in your work; and it will allow you to invent in ways we couldn't before. So you really need to understand it, you need to self-learn it, you need to use it a lot. You need to try. Try to do things that are outside your comfort zone.

Andy Jassy 54:48

Neither of my two children are developers, but they, you know, have used AI to create applications. Treat it as the focus of how you organize your day and use tools at work. But I don't think it will replace critical reasoning and thinking. And I think the same goes for universities. I often think about this in many ways— I’ll take business school as an example. You would get three cases to analyze, you have to do a lot of reading, take notes, figure out what the key issues are. Then you would get a bunch of attachment materials, and you also have to take notes on those. Then you have to integrate what you think the company should do. And you know, often there are a bunch of people who don't do this homework at all— they don't read, they don't research, they don't take notes. And all of this, AI can help you with, it can get everyone to that higher starting point faster.

Andy Jassy 55:52

But I also want to say, really connecting those points, those fragments— whether it's to form a strategy, a thought, a plan, a new product, or a new customer experience— I don't think AI will do that for most people most of the time. I'm not saying that will never happen.

Andy Jassy 56:10

I still believe you need human judgment, creativity, inventiveness, experimental spirit, and the ability to think about how all the pieces fit together to ultimately create amazing experiences and businesses. So I think you should use these tools, they will be a huge help, but I still believe critical reasoning and thinking will be very important in everything we do. We're nearing the end. The business you are doing is like...

a16z 56:43

From satellite connections to robots, to healthcare, so diverse. I mean, this ambition is incredible. Think about how this business started, you know, selling books on the internet, and now it has become this. How do you view, well, what connects these different...

Andy Jassy 57:00

business lines together? Or how do you judge which business lines are worth investing years in? Or how do you ensure you don't spread yourself too thin? Or how do you understand this extremely broad landscape?

Andy Jassy 57:11

Well, I would say a few things. First, I think we are in the middle of the three biggest transformations of our lifetime— perhaps in history—: AI, robotics, and space development. They are all enormous and will fundamentally change many customer experiences we all encounter. And we are right in the middle of these three major transformations.

Andy Jassy 57:40

If you look at what we are doing in retail, and how we are using robots, what we are doing on the AI side— which includes making it easier for both internal and external builders to create successful AI applications and agents, as well as reinventing all of our consumer experiences.

Andy Jassy 57:58

If you think about Amazon Leo— the low Earth orbit satellite constellation we are building for the 500 million households around the world that do not have broadband connectivity— there is a series of incredible transformations happening, and we have the opportunity to help customers leverage them to make their lives better. And the way we are doing this helps customers and also brings us long-term sustainable business. And you know, I think we have the opportunity to become more relevant to customers in more areas. This doesn't even account for what we are doing in the grocery sector, what we are doing in autonomous ride-hailing and healthcare, and, you know, those.

Andy Jassy 58:47

You know, regarding how we are organized, we have often had this internal debate over the years, and many outsiders have asked the same question: is this spreading too thin? I think it partly depends on how you organize yourself and how much you are willing to invest. Because we don't let the same group of people manage multiple businesses; we organize in this decentralized "single-threaded" way. As long as you have the right leaders and builders on these teams, as long as the experience is big enough to drive growth metrics, as long as you have differentiated ideas in an area that is not well served, you have the opportunity to pursue many of these things at the same time— and we have been doing that for a long time. Then it comes down to how much you are willing to invest.

Some companies— completely understandable— will say: we just want to do our own thing, we will stick to our core business and grow it. By the way, that can also be a very successful strategy. And we are also very focused on continuing to grow and strengthen our existing business, and I think we are doing quite well at that, at a good pace. But there are so many other customer experiences that we know customers really want us to help solve, and we can solve— they will help our customers and also change our business in the long term. So we will continue to pursue these. This is an inspiring conclusion— endless ambition. Andy, thank you very much for coming on the show. Yes, thank you for having me. Thank you very much, everyone.