Amazon CEO Andy Jassy on the $220 Billion AI Buildout

The episode opens on Andy Jassy’s own numbers. Even spending about $220 billion this year, Amazon is leaving a meaningful amount of demand on the table. The company always holds some capacity for on-demand use, but 2026 is sold out, 2027 is largely sold out, and a good chunk of 2028 is already committed. Financial markets, he notes, have started asking whether there is enough money in the world to finance the buildout.

Everything people do digitally—stock markets, banking, healthcare—already runs on data centers, Jassy says, and Amazon’s job is to have as much capacity as it can. The clip also flags the tension Horowitz will press later: the big labs will pay almost anything for power, cooling, and chips, which raises the question of whether cloud providers should buy out older customers to make room.

Erik Torenberg welcomes Jassy to the podcast five years into his tenure as Amazon CEO. He ticks off the present tense of the job: roughly $220 billion of CapEx and AWS growing 28%. He asks how Jassy thinks about Amazon’s ambition in 2026 compared with 2021, when he formally took over.

Jassy says 2021 feels like a decade ago. The company’s overall top line is about 80% higher, AWS is almost three times bigger, and advertising revenue is about two and a half times bigger. A lot happened quickly after he started, on the tail end of the pandemic.

Retail had a cost-to-serve problem because Amazon had doubled, in about 18 months, a fulfillment network that had taken 25 years to build, then stood up a last-mile transportation network the size of UPS in about 20 months. Build that fast, he says, and you do not have things optimized.

Fulfillment cost-to-serve was not where they wanted it. A year or so into the job, people told him it was a terrible time to start. In retrospect he calls it a blessing: the squeeze forced the company to consider ideas it might not have touched for years, ideas that were good for customers and for the business.

The biggest change was regionalizing a flat national fulfillment network in the United States. Putting items closer to customers shortened transportation distances, lowered cost-to-serve, and got orders there faster. Re-examining closely held beliefs, he says, unleashed delivery speed and a large part of why retail has grown as fast as it has.

Generative AI then exploded. Machine learning was not new, and there had been plenty of waves billed as the real tipping point. Around 2023, Jassy says, you could see a huge difference. AWS was already growing quickly, with the broadest functionality, stronger operational performance and security, and a distinctive partner ecosystem.

Amazon was the clear leader and had thought AWS could become a few hundred billion dollars of revenue over something like ten years. With every customer experience going to be reinvented through inference, that view has changed. Last quarter AWS grew 36.7% year over year on a $169 billion annual run-rate business.

On top of that growth are newer bets: Prime Video as a large, profitable business; the low-earth-orbit constellation Amazon Leo; Zoox, an autonomous ride-hailing service; healthcare; robotics. Jassy says the chance to help more customers, across more experiences, is unified by one mission—make people’s lives better and easier every day.

Ben Horowitz follows up on the fulfillment build. Going that fast is sloppy, and there was internal angst about it. Amazon is known for operating excellence and frugality, and here the company was building the equivalent of UPS. How does a CEO communicate that without lowering the cultural standard?

Jassy points to Amazon’s 16 leadership principles, which a lot of people know about. They are supposed to govern how the company operates as a team. Several of them are in tension with one another, which confuses people, but he says that is how the real world works: nothing is perfectly clean, and you are always making trade-offs.

When they decided to expand fulfillment and last-mile so aggressively, the team knew it would be hard, more expensive, and less optimal than going slower. Customers and sellers needed it. Early in the pandemic physical stores were closed, and Amazon’s role in daily life was, in his word, seminal.

The first decision was whether to build fast and imperfectly in order to serve customers and sellers trying to keep businesses alive. As they almost always do, they shaded toward customers. It was then obvious that not everything was going to plan. Fulfillment people were in offices more than the rest of the company, but design, planning, and assessment still involved people working remotely, all over the place.

That put bias for action in tension with relentlessly high standards—the cost structure, delivery, and design you always planned for. They chose customers and speed, and they expected cleanup. They did not foresee every way the network’s topography would fail to be optimized.

What saved them, Jassy says, was metrics plus a truth-telling culture: get the data and the issues on the table, be direct, and do not paper over problems. Then you sort small productivity fixes from architectural ones. Re-architecting key parts of the network, which might have been eighth or ninth on that team’s list, became the way to get both speed and cost.

Horowitz turns to infrastructure. When Jassy started AWS it was build-it-and-they-will-come: put up capacity and rent it, or you were in trouble. This time the scale is much larger. Amazon and a few others are adding data-center and chip capacity so fast that financial markets say there is not enough money in the world to finance it, with skittishness in corporate debt and equities.

Given how steep demand is, and given the fight for cooling, power, and chips, Horowitz asks how Jassy knows how much is too much. Jassy wishes there were an equation. In the first wave of cloud there was some uncertainty about whether people would use AWS, but the company had strong internal signals.

In the early 2000s, Amazon product teams were not moving as fast as leadership wanted. It turned out they were spending 70 to 80% of their time on undifferentiated infrastructure and only about 20% on what made the product different. Every team was reinventing the wheel.

AWS was originally conceived so Amazon itself could move faster. If they built high-scale, high-availability, low-cost, flexible building blocks—compute, storage, database, content delivery, messaging—consumer teams could build quicker. They figured other companies had the same problem. Even before the public launch, so many internal teams wanted in that they knew they were on the right track. Nobody anticipated how fast it would grow outside.

AI had its own false dawns. Six or seven years before the generative wave, people already thought it was AI’s time, and Amazon built AI services that were a pretty good AWS business—nothing like now. Early AWS felt fast and was about a $60 million annual run-rate business after three years.

In the first three years of this generative-AI wave, Amazon’s AI business inside AWS is over $25 billion in run rate: several hundred times faster. They have not seen anything like it. The question becomes how much CapEx to put against the opportunity.

For 2026 they expect to spend about $220 billion. Jassy remembers a leadership-team breakfast 15 years ago when they could not believe a telco spent $50 billion of CapEx. The numbers are mind-blowing. Even at $220 billion this year, they are leaving a meaningful amount of demand on the table.

What helps is visibility into customer demand. CapEx in AWS has two cash-cycle buckets. Data centers have to be funded about two years before you can put servers in and start monetizing them. You make that investment once; it is a useful-life asset of about 30 years.

Servers and networking gear are only a few months ahead, depending on the component, so you see demand before you pull the trigger. If you do not need to fill those halls, you do not buy the gear. On average the equipment breaks even in about three years, with useful lives of about five to six years, and most deals are about five years.

Building this many data centers at once means a lot of capital and free-cash-flow headwinds until later years, when the sites are online, monetized, and earning back the server and network spend. That is the shape of the cash cycle when you grow this fast.

Horowitz notes that when AWS started, data centers were about the last thing anyone expected to become a political issue. In the United States especially, he says, nobody is as anti-data-center. He asks whether the backlash has any merit, and how it changes what Amazon builds at home versus in other regions.

Jassy’s context is that almost everything people do is digitally driven—stock markets, banking, healthcare, entertainment, shopping, education, business, ride-hailing—and all of it uses data centers. Horowitz jokes that everyone is anti-themselves: they do not want to use their phones anymore. Jassy adds that this was true before AI.

AI, he says, is the biggest technology transformation of their time, which is saying something after the internet, the cloud, and mobile. The infrastructure required to support that growth has no recent parallel except perhaps the highways and bridges of the 1950s.

The United States, and the world, cannot afford not to build data centers, or another country—he names China—will set the leadership and the norms, domestically and globally. They have to build. That does not mean the industry cannot do it better; it is happening quickly and everyone is learning together.

He thinks a number of myths and mistruths are being propagated because data centers are an easy villain, and political opportunists like an easy villain. His examples start with water: golf and almond farming consume far more. Golf courses, he says, use 200 times more water than all of Amazon’s data centers, and more in two days than the data centers use in a year.

On pollution, new sites have to register backup generators as if they ran full-time at full capacity. In practice they run about 1% of the year, on the order of 10 hours, for maintenance. Energy rates in geographies with a big Amazon presence have gone up in some places and down in others.

The real issue, Jassy argues, is that the country has not invested in its grids the way it needed to. Something was going to force that investment; AI’s speed and scale happened to be the thing that pushed it over. Once capacity is in, he expects rates to flatline or fall, as they have in Indiana, Georgia, and Mississippi.

People also forget the tax. In St. Joseph County, Indiana, Amazon has paid about $3 billion, money communities spend on what matters locally. The company has programs such as Built Together and just committed another billion dollars over the next five years in communities where it has data centers.

The work with communities has to keep going, he says, as a public-private partnership, because the buildout is too important to the country’s future to lose perspective. Imagine not having built the highways and bridges, and what that would have meant for prosperity and capability. Be partners in the places you build.

Torenberg puts $220 billion in daily terms: about $600 million a day. Off-line, Jassy has described AI demand as a barbell, and Torenberg asks him to situate the market that way. Jassy starts with the same shock: a few hundred times faster than the beginning of AWS, and a $25 billion AI run rate inside AWS in three years. It is large, and it is still barbelled.

On one end are the AI labs, consuming gobs of compute unlike anything he has seen, plus a few runaway apps—ChatGPT, coding tools, things like that. On the other end are enterprises using AI mostly for cost avoidance and productivity: automating customer service, fraud, business-process automation.

In the middle sit enterprises’ current production workloads. Many already run on AWS and some already use inference, but the overwhelming majority still do not, not in any pervasive way. That middle, Jassy thinks, will be the largest absolute segment of AI usage.

He cannot predict the exact rate. It increases every year, and it will not happen in one fell swoop. Zoom out five to seven years—and he does not think the world will wait that long—and quite a few, if not most, of the enterprise applications people use today will be using inference pervasively. There is a lot of growth in front of them.

Horowitz turns to the impending shortage of data centers, chips, and memory, especially into 2027: everyone is building now, it will not all be ready by year-end, and demand is not slowing. Labs will pay almost anything for power, cooling, and chips. Longtime customers pay regular rates to run their businesses.

Is there pressure, Horowitz asks, to buy out the old customers and sell that capacity to the new ones, because the new ones will pay whatever it takes? You cannot fairly double the old prices. It is a strange split: people who have been on AWS for 20 years, whose workloads are probably heading toward inference, versus insatiable new buyers.

Jassy says they do not think about it quite that way, though he understands the arithmetic. There is only so much data center, power, and memory, and he does not think the shortage ends in 2027. That is one reason more capital would not be a bad use of money: 2026 is sold out.

2027 is largely sold out. They always leave some for on-demand—Bedrock, EC2—but the big chunks are taken. A good chunk of 2028 is sold out. Both the AI business and what Amazon calls core, meaning non-AI, are growing at rapid rates, so memory itself is a trade-off: how much to allocate to HBM for accelerators versus DRAM for core.

They start from being customer-focused. During the pandemic, Jassy says, competitors tried to optimize within quarters to drive revenue while companies were cost-cutting. AWS did the opposite: it sent people out to help customers cost-optimize, knowing that would diminish revenue in the short and medium term.

The belief is that if you help customers over a long period, the business comes with them. You could run the numbers and decide that scarce capacity should go to whoever will pay more, squeezing core customers. Those customers run their businesses and their core applications on AWS. Amazon is not going to squeeze them. It will make sure they can scale.

The job is to have as much capacity as possible so core customers and AI customers can both use AWS, in a way that is sustainably economic. He wishes supply constraints were only a 2027 issue. He thinks they will show up in different components for several years.

Amazon is probably putting in more capacity than anybody else, he says, because it is passionate about remaining the significant leader in infrastructure and AI. They are spending $220 billion this year and could spend a lot more—medium to long term it would not be a bad use of capital—and they still probably would not satisfy all the demand. There is just an incredible amount of it.

Horowitz goes back to 2015, when Amazon bought Annapurna Labs. There was not much AI demand then, so the purchase was probably not about giant training clusters. What drove it, and how has it paid off in this new world?

Jassy calls the Annapurna team one of the most talented groups he has worked with, and the deal one of the best acquisitions in Amazon’s history, if not the best. They foresaw some pieces, but they did not buy the company for AI chips.

At a certain scale you see customer patterns large enough that vertical integration can create more value for customers and better economics for the business. Annapurna started on the core EC2 server. Too much of each machine was going to virtualization, networking, security, and storage. They wanted nearly 100% of the server for customer compute.

Amazon had already been working with Annapurna on a NIC that offloaded those functions. They liked the team and the product. That part was going to end up in virtually every EC2 server; it has ended up in every server they have. They bought the company for that core compute capability, and because the chip talent might build other chips too.

They still had a deep partnership with Intel, which had virtually all of the CPU share. A partner in that position is often not thinking about how to drive price-performance; it is thinking about features that grow share and revenue. Jassy says that is understandable—it is part of how markets work, and part of the innovator’s dilemma.

Amazon had tried to push on price-performance and could not get attention. Customers moving on-premises workloads to the cloud wanted better economics. The next assignment for Annapurna was a CPU: Graviton. Today 99% of Amazon’s top 1,000 EC2 customers use it pervasively. It is about 40% better price-performance than other x86 processors, which has been hugely impactful for customers and for the business.

A few years ago they saw the same movie in AI: one dominant provider, very good, a deep partnership they expect to keep, with giant markets and many other products and priorities. Customers still needed better price-performance. They spun out a separate Annapurna team to build an AI accelerator, Trainium, which Jassy puts at about 30 to 40% better price-performance than comparable processors.

Coming back to the middle of the barbell, he asks what will get more of today’s enterprise production apps onto inference. Some of it is talent that is comfortable with post-training and fine-tuning, and organizational muscle that knows what to prioritize. A big chunk is the cost of inference.

Inference costs have come down, but in his opinion they are still an order of magnitude too expensive for pervasive use. If you ask enterprises to rewrite CPU applications onto accelerators, and the cost difference is large, the incremental benefit has to clear a high bar. Take the cost of inference down and that bar falls. Every application gets better, smarter, and more capable. Driving that cost down is a really important part of making AI pervasive, and the chips business is a big piece of it.

Horowitz offers a data point from his own firm: one employee spent $2 million on inference, doing good work, which is how you notice it is not cheap. He then asks about Amazon’s large investments, on the order of $18 billion in Anthropic and $50 billion in OpenAI. Similar checks from Nvidia got Wall Street in a tizzy, recalling the dot-com boom. Why did those deals make sense, especially since the labs are also customers?

Jassy calls both companies impressive, with top-line growth he is not sure anyone has seen before. The investments are a combination of the teams, the models and applications they have built, Amazon’s belief in how transformative AI is, and the view that these will be two of the most successful companies in the space over a long period. Not everyone will win, but there will be many successful companies, and he believes these will be two of them.

The partnerships are close. Anthropic is the longer relationship: many customers run Anthropic on Bedrock, Amazon’s inference service, and Jassy thinks customers are most successful running it there. The companies have a deep chip partnership as well. Anthropic has been building on Trainium, which has been good for Anthropic and for the Trainium team; they have learned from each other. Belief in AI, belief in those companies, and the partnership itself made the investments make sense.

On the OpenAI side, many AWS customers wanted to use OpenAI and could not, because of OpenAI’s prior business arrangements. For a long time the companies had mutual respect and a desire to work together. Over months of conversations they got OpenAI onto Bedrock. They also built a unique product together: a stateful runtime environment, which is how Jassy thinks all agents will be built going forward. You are going to need state.

Like Anthropic, OpenAI made a multi-year, multi-gigawatt commitment to Trainium. Amazon believes OpenAI will be one of the most successful companies in the space over a long period. All of that, Jassy says, made deeper partnerships and investments make sense.

Horowitz switches to how Amazon uses AI to build the company. Jassy has said the Bedrock team rebuilt the inference engine with six engineers in 76 days—work that would have taken 40 people a year. Amazon is famous for two-pizza teams. Meanwhile the industry specialized into designers, product managers, front-end and back-end engineers, and suddenly everybody is a renaissance person. What is the current thinking on org design?

Jassy agrees the Bedrock story was eye-opening. They spent time sharing it internally to inspire people and remind them what is possible. He started at Amazon 29 years ago as an assistant product manager owning five products. As companies get successful, a product manager owns one product, or one slice of it—pricing, say. Everyone becomes more specialized and accepts that as the cost of building something big. He says it is not, especially with AI.

Even in the week of the taping, a consumer-business idea around agents was agreed on Thursday. Four or five people had a working prototype that was 80% of the way to shippable by Sunday. Those are different worlds. For organizations, he thinks, number one is that fewer humans will be doing the work they do today.

Look at how much coding is already being done agentically, and at customer service, business intelligence, and research, then fast-forward to how spreadsheets will be built and how many lawyers you need. You will not have no people, but you will not throw as many humans at scale as businesses did as they got bigger. There will be different jobs, some known and some not, as AI, robotics, and space develop. There will still be fewer humans doing what they do today.

Teams will be much smaller, with people who can work cross-functionally instead of owning one sliver, because models will get them 80% of the way on a lot of functions. Every team will start at a more advanced stage, and you will be able to invent more expansively and more quickly. A bunch of things still have to change.

Startups founding today are mostly agentic and AI-native from the start. Larger companies that have been around will have to figure out how to build at that same pace. Some will, especially those with capabilities and a loyal customer base. Some will move slowly and, in his opinion, be more vulnerable to startups. He predicts companies will have to rewrite almost all of their code to be agentically capable, and to let agents update pieces of systems inside constraints.

Security is the example. The last month or two of public nervousness about models has focused on fantastical scenarios. They are possible: an evil botnet taking over everything, robots infused with malicious intent. He thinks the industry, public and private, will figure those out together. Horowitz interjects that you have to start by paying attention to some of the things.

Jassy’s point is that there is not enough discussion of the more immediate risks, which are around cybersecurity. These models find so many more vulnerabilities than humans were finding. Fast-forward, and the way applications have to respond to imminent security threats is going to be in minutes.

Today you find a vulnerability, classify it, develop a patch, and then—even if that part is quick—most companies take a long time to deploy because of dependencies. You are going to have to deploy in minutes. If your applications have not been rewritten so agents can update them, you will have no chance in that world. Organizations will be reinvented, and the companies that prioritize this will be in much better shape than those that do not.

Horowitz says the industry did not do a great job architecting security, from big systems down to embedded devices. The day of reckoning is here. That is a big opportunity for new companies, but the threat is not a thought experiment. The theoretical risks may or may not happen. This one is clear and present.

Jassy says almost every AWS partner and customer of any size is worried about the security piece. Amazon has a professional-services practice helping many of them, and a service called Continuum that identifies issues and builds patches. It is easier to say than to do. For a lot of companies and engineers, security is not the exciting part, so it is easy to wait until there are problems. If you protect data that matters, that is a bad idea.

Horowitz ties it back to specialization: esoteric security engineers who do not like working with other engineers, a scarce resource, and as a result most software companies’ practices for writing software securely are not that good. That has to change, and the good news is there are tools for it. Jassy agrees you can only hire so many pen testers and they can only get to so many applications. Models plus automation will let you pen-test the entire surface and find things you might otherwise have found only by fluke.

Horowitz lays out three things you can do with AI. One is cut costs—fewer accountants, fewer lawyers—which has attracted venture firms that buy old companies to shrink them. That is not interesting to a16z, he says, because it means you have very few ideas. The second is build something the world has never seen, which is exciting. The third is what Amazon did with Alexa: take a successful product and rethink it from the beginning because there is a new platform. How do you get to that, and where does it make sense?

Jassy likes the three-way split. Organizations will get more streamlined, and if you do it right that lets your best builders on small teams go after the second bucket—things people have not seen. When you have experiences like the four-day prototype that is 80% of the way there, it is so inspiring that people spend more free time and thought cycles on new ideas. That, to him, is one of the most exciting parts of AI.

For existing companies with successful customer experiences at scale, he warns against treating AI as primarily an incremental improvement. If you do, you will very likely not have the winning experiences three to five years from now. Every experience we know, even the ones that seem hardest to change, will be meaningfully different with AI in a few years.

Thinking more than incrementally means getting teams to reimagine from scratch what those applications look like if you were building them today, with AI available. In Alexa’s case he felt it was fairly obvious. The brains were inventive when they were built, but generative AI made the possible brains so different that you had to rewire them to stay relevant. Obvious did not mean easy.

Teams are connected to what they have built, proud of it, with hard-earned scars about what works. Getting them to start over takes time, because people say that is not how we have done it, or we tried that and it did not work. Harder still are experiences where the technology’s advantage is less obvious: retail, Prime Video, Fire TV, Prime Music. He believes every one of those businesses needs to be reinvented, and there are teams working on totally reimagined versions.

They will not get it right from the get-go. They will iterate. It is already yielding new ideas and customer experiences they find compelling. Then comes a race: do you want a completely different experience, almost a new app, or do those advances land fast enough that you fold them into the existing product and change the overall experience bit by bit? Either way, he thinks every one of those experiences has to be different in the next few years if you want to remain the successful leader.

Horowitz calls it a tricky leadership problem. Early in his engineering career, after an argument, his boss told him everything is up for discussion until the first line of code is written, and then it is religious. That is the Alexa problem: you are giving up a lot of the knowledge that went into the original thing. Customers, too—some large set loves the current experience.

Jassy says almost every big transformation is huge change management. Moving from on-premises to the cloud was the same: infrastructure leaders proud of what they had built, who knew it inside out, who were in the middle of the action, and who did not want to learn something new.

The same will be true of reinventing customer experiences. People are proud of the last ten years, have a hundred improvements they are excited about, and spend their waking hours on them. Completely starting over is intellectually and emotionally hard. Horowitz does not doubt it.

Torenberg notes that a16z recently incubated Andreessen Horowitz Academy to help young people prepare for this transformation. Jassy has said AI makes building easier but young people still need real expertise. If he were 18 in 2026, what advice would he give himself?

Jassy has a 25-year-old and a 22-year-old. The younger one was 19 when this really took off, and he learned quickly that you cannot tell them not to use AI in education. That train has left the station. Universities would be wise not to fight it. It is a tool people are going to use.

At the same time, Amazon has a writing culture, and people are using AI to write. You can tell when something has been written by AI: the pieces are not put together quite right. The advice he gave his kids is that you cannot run away from it or wish it away. It will be incredibly impactful in work and will let you invent in ways that were not possible before.

You need to learn about it, educate yourself, use it a lot, and try things outside your comfort zone. Neither of his children was a developer; both have used AI to build applications. Use it as a focal point for how you operate your day and your tools at work. He does not think it replaces critical reasoning and thinking, and he thinks that is true for universities too.

Business school, as an example: you used to get three cases a day, do a lot of reading, take notes, figure out the key issues, work the exhibits, and put together a plan. Plenty of people just did not do the work. A lot of that goes away with AI. The models will get everybody to an advanced starting point much more quickly.

Putting the pieces together—to build a strategy, an approach, a plan, a new product, or a new customer experience—is not something he thinks AI will do for most people most of the time. It is not that it will never happen. You will still need human judgment, ingenuity, invention, experimentation, and a sense of how the pieces come together. Use the tools; they will be huge aids. Critical reasoning remains very important in everything we do.

Closing, Torenberg points at the range: satellite connectivity, robotics, healthcare, from a company that started selling books on the internet. What threads the businesses together? How does Jassy decide which lines are worth years of commitment, and how does he make sure Amazon does not spread itself too thin?

Jassy says we are living through three of the biggest transformations of a lifetime, maybe ever: AI, robotics, and space. They will completely change a number of customer experiences, and Amazon is in the middle of all three. Look at retail and how they use robotics; at making it easier for internal and external builders to create AI applications and agents; at reinventing consumer experiences.

Amazon Leo, the low-earth-orbit constellation, is aimed at the 500 million households around the world that do not have broadband. There is an incredible set of transformations customers can leverage, in ways that make their lives better and that can become long-term sustainable businesses. He thinks Amazon has a chance to be much more relevant across many more areas—and that is before grocery, autonomous ride-hailing, and healthcare.

Internally, and from outsiders, people have long asked whether the company is spreading itself too thin. Jassy thinks that depends on how you organize and what you are willing to invest. Amazon does not try to have the same group of people manage multiple businesses. It organizes in a decentralized, single-threaded way.

If you have the right leaders and builders, an experience big enough to move the needle, a differentiated idea, and a space that is not being well served, you can pursue a lot of these things simultaneously. Amazon has, over a long period. Then it comes down to how much you are willing to invest.

Some companies, understandably, stick to their knitting: grow the core business you have. That can be a very successful strategy. Amazon is intent on growing its existing businesses, and Jassy thinks it has done so at a good clip. There are so many other customer experiences people want Amazon to help solve. They will help customers and change the business long term, and the company will keep pursuing them.

Torenberg calls it an inspiring note to end on—endless ambition, from a bookstore to all of this. He thanks Jassy for coming on the podcast. Horowitz adds his thanks. Jassy thanks them for having him, and says he appreciates it.