2026-10-24 20:06

Introduction: Jassy bets on in-house chips, driving inference costs down by another order of magnitude. Recently, Andy Jassy, CEO of Amazon, appeared on a16z’s popular podcast, engaging in an in-depth dialogue with a16z co-founder Ben Horowitz.

Recently, Andy Jassy, CEO of Amazon, appeared on a16z’s popular podcast, engaging in an in-depth dialogue with a16z co-founder Ben Horowitz. In this interview, Jassy provided a comprehensive overview of Amazon’s strategic vision amid the generative AI wave.

Facing Wall Street’s concerns about tech giants’ seemingly bottomless capital expenditures, Jassy offered the most direct market feedback: demand is not slowing down—it’s rapidly outpacing future capacity. He also meticulously dissected the current “barbell-shaped” structure of AI demand and clearly stated that to truly unlock enterprise-grade AI applications, compute inference costs must be reduced by another order of magnitude.

“Even with $22B in spending, we’re still leaving massive demand on the table”

The central question dominating today’s market is whether multi-billion-dollar capital outlays by tech giants will become excess sunk costs. Jassy directly addressed these concerns with striking order data.

“Even if we spend $22 billion this year, we’re still leaving a substantial amount of meaningful demand on the table,” Jassy candidly stated, highlighting the current extreme shortage of compute capacity: “We always reserve a portion of capacity for on-demand use, but 2026 capacity is clearly sold out, 2027 is nearly sold out, and even a significant portion of 2028 has already been pre-sold.”

From a business growth perspective, the AI-driven growth curve is extraordinarily steep. Jassy compared AWS (Amazon Web Services)’s early development pace with its current trajectory: “In the first three years after AWS launched, we thought we were growing fast—our annualized revenue run rate was around $60 million at the time. But in the first three years of this generative AI wave, our AI business on AWS has already surpassed a $2.5 billion annualized run rate. Its speed is hundreds of times faster than back then—we’ve never seen anything like it.”

To address financial market concerns, Jassy explained the cash flow cycle behind data center investments. He noted that data center construction requires two-year lead time, but has a 30-year lifespan; server and network equipment procurement only needs a few months’ lead time. This gives companies high visibility into customer demand. “If you realize you don’t actually need to fill those data centers, you won’t buy servers or network gear. Most server contracts are five-year agreements, typically achieving breakeven within three years. Due to this cash cycle, when growing at such a rapid pace, early free cash flow does face headwinds—but once fully deployed, the returns will be substantial.”

Countering Data Center Controversy: “Two days of golf course water use equals our annual consumption”

In the U.S., data center energy use and environmental impact have become political flashpoints. Jassy responded forcefully, elevating the debate to the level of national competitiveness.

“What we’re experiencing in infrastructure development is on par with the scale of America’s highway and bridge building in the 1950s. The nation—and indeed the entire world—cannot afford the cost of not building data centers. Otherwise, countries like China will establish all leadership positions and standards, which would be unacceptable both domestically and globally.”

On environmental claims, Jassy presented impactful data: “There are many myths and half-truths being circulated—politicians and speculators love finding a simple ‘villain.’ In reality, golf courses or almond farming consume far more water than data centers. As a single data point: golf courses consume 200 times more water than our data centers; their two-day usage exceeds our annual consumption.”

He also added that generator pollution is grossly exaggerated: “You must register emissions assuming full 24/7 operation, but in practice, we run backup generators only 0.1% of the time annually (about 10 hours) for maintenance.”

The “Barbell Effect” of AI Demand and the Core Logic Behind Reducing Inference Costs

When breaking down specific AI demand, Jassy introduced a vivid “barbell” theory—explaining why Amazon remains committed to its in-house chip strategy.

“One end of the barbell is AI labs consuming massive, unprecedented compute power; the other end is enterprises using AI primarily to reduce costs and boost productivity, such as automating customer service or business workflows. The middle of the barbell is the current production workloads of enterprises. Most enterprises haven’t yet adopted AI inference widely—but I believe that within 5 to 7 years, virtually all existing enterprise applications will deeply integrate inference technology.”

However, unlocking the vast middle market hinges on price.

“Over the past few years, inference costs have dropped significantly, but in my view, they’re still too expensive by an order of magnitude (an order of magnitude too expensive).” Jassy emphasized, “If you ask enterprises to rewrite their existing CPU-based applications using AI accelerators, and the cost difference is huge, the barrier to rewriting becomes extremely high. But if you can lower inference costs, that barrier drops dramatically. That’s precisely why we’re pushing hard in our in-house chip business.”

Jassy highly praised the 2015 acquisition of chip design firm Annapurna Labs, calling it “the most successful acquisition in Amazon’s history—no contest.” He revealed that through self-designed Graviton processors (40% better cost-performance than other x86 chips) and Trainium AI accelerators (30–40% better cost-performance), Amazon is actively driving down compute costs.

Organizational Restructuring: Six Engineers Did What 40 Took a Year to Do—All Products Must Be Rebuilt from Scratch

Beyond external compute delivery, AI is internally reshaping how Amazon operates, starting with organizational structure and product development logic.

Jassy shared a striking internal case: the Bedrock (Amazon’s AI inference platform) team rebuilt the inference engine in just 76 days with only six engineers—something that previously required 40 people working a full year.

“This shocked everyone,” Jassy admitted. “In the future, fewer humans will be needed to do the same work. I’m not saying people aren’t needed, but as business scales, we won’t pour in as many people as before. Teams will be smaller, and individuals will need to work across functions because models can handle 80% of the work.”

At the product level, Jassy believes incremental improvements are no longer viable—radical reinvention is essential.

“For products like Alexa, their ‘brain’ was innovative at launch, but with generative AI, capabilities have changed fundamentally. If you want a winning customer experience, you must rewire Alexa’s brain.”

It’s not just Alexa—Jassy pointed out that Amazon’s retail, Prime Video, music, and all core consumer experiences must undergo complete reimagining within the next few years. “Teams often resist due to pride and legacy baggage, saying ‘we tried that before—it didn’t work.’ But this painful reinvention is unavoidable. If you only aim for gradual improvement, you’ll be obsolete in three to five years.”

Below is the full transcript of the podcast:

Andy Jassy 00:00

Even if we spend $22 billion this year, we’re still leaving a considerable amount of demand on the table. We always keep some capacity reserved for on-demand use, but 2026 is clearly sold out, 2027 is nearly sold out, and a significant portion of 2028 has already been pre-sold.

Andy Jassy 00:14

Financial markets say, 'Oh no, there isn't enough money in the world to fund all this.' But everything we do in the world today—stock markets, banks, healthcare—they all rely on data centers. Our job is to secure as much capacity as possible; we have the opportunity to serve far more customers and improve their lives every day. Large labs are willing to pay almost any price for power, cooling, and chips. Maybe we should buy out old customers and resell to new ones—because they’ll pay whatever we ask.

a16z 00:48

Andy, welcome to the show. Great to have you here. Thanks for having us. You’ve been CEO for five years now, and you’ve accomplished so much. Just being here is impressive—so many things happening under your leadership—you’re managing $22 billion in capex, AWS grew 20%, 28%. Yeah, you’ve done a lot.

Andy Jassy 01:09

You’ve achieved so much—how do you see Amazon’s ambitions in 2026 compared to when you officially took over as CEO in 2021? Oh, 2021 feels like forever ago, like… you know, like ten years ago. But the world changes so fast. And you know, well…

Andy Jassy 01:30

Compared to 2021, our overall revenue is up about 80%, AWS business is nearly tripled, advertising revenue is about 2.5 times higher. So this business, you know, has grown substantially.

Andy Jassy 01:47

I think several things happened. Shortly after I took over—right at the tail end of the pandemic—our retail business faced a real fulfillment cost challenge: we essentially doubled the scale of our fulfillment network. The network we built in 25 years was doubled in about 18 months; then, in roughly 20 months, we built a last-mile delivery network equivalent to UPS. And whenever you build that fast, you inevitably leave behind inefficiencies. Our fulfillment costs weren’t what we wanted. You know, that brought a lot of challenges and anxiety at the time.

I remember, I’d only been CEO about a year—maybe less—everyone said, 'Wow, this is the worst timing to take over, terrible luck.' But looking back now, I think it was actually a great start—truly fortunate—because it forced us to consider things we might never have thought about for years, and those ideas were excellent for our customers and our business. The biggest one was our fulfillment network: we transformed our originally flat, nationwide network into a regionalized model. By regionalizing, we placed goods much closer to customers, which not only reduced fulfillment costs due to shorter transport distances but also drastically improved delivery speed.

During that period of struggling with fulfillment costs, re-examining our most entrenched beliefs truly unlocked enormous demand in delivery speed and retail operations—the key reason we grew so quickly. Of course, you know, right after that came AI, which further accelerated everything. It wasn’t that AI was something new—we’ve been doing machine learning for a long time, and there were several previous waves considered 'true turning points' in AI. But when you get to 2023, you really see the massive difference generative AI brings. At that time, our AWS business was already growing very fast.

We had, you know, the broadest feature coverage, stronger operational performance and security, and I believe a different partner ecosystem. We were clearly leaders, and we felt this business could sustain thousands of billions in revenue for over a decade. Now, seeing what’s happening with AI and knowing every customer experience will be re-invented through inference and AI, we actually believe AWS not only grows very fast—last quarter, we grew 36.7% year-over-year—with a $16.9 billion annualized revenue base—the opportunities ahead are truly astonishing.

And you know, if you layer on top, we have seeds of new businesses: whether it’s how the team turned Prime Video into a large, profitable business; or our low Earth orbit satellite project, Amazon Leo; or the autonomous ride-hailing service with Zoox; or healthcare, and our robotics initiatives. You know, we genuinely have the chance to help far more customers, cover far more customer experiences, and make their lives better and easier every day.

This mission—to constantly make customers’ lives better and easier—is what truly unifies everything we do across the company. So I have a few follow-ups. First, you mentioned earlier that when expanding the fulfillment network so quickly, it necessarily involved inefficiencies. You know, when you push forward that fast, you’ll inevitably have rough edges. How did you—especially given you mentioned internal anxiety—communicate what you were doing as a CEO while maintaining cultural standards? I mean, Amazon is known for excellence, operational efficiency, frugality—all those things. Yet you’re now saying: we’re building something like UPS. How do you reconcile these?

Well, interestingly, we have these well-known leadership principles—16 in total. Many of them guide how our teams think about operations and culture.

But several of them inherently create tension—sometimes people get confused by this. Yet reality reflects the true nature of the world: nothing is ever clean-cut—you’re always making trade-offs. So returning to that moment, when we decided to expand our fulfillment and last-mile delivery networks at such an aggressive, rapid pace, the team faced the dilemma: we knew that moving so fast would be difficult; costs would likely be higher than if we moved slower; we probably wouldn’t optimize every detail as thoroughly as we would if we moved slowly. But customers and sellers really needed it. So should we lean toward ‘just get it built’? Remember, this was early in the pandemic—every physical store was closed.

I mean, if you look at Amazon’s role during the pandemic in people’s lives, it was truly groundbreaking. So we had to ask ourselves: should we move fast—even if not perfectly optimized—to better serve those customers and sellers—those sellers who were also trying to keep their businesses alive? That was the first decision. And we chose, as we almost always do, to side with the customer. You know, then it quickly became clear during the build process that not everything went perfectly.

We had, you know, our fulfillment team spent more time in offices than most other departments. But even so, many people were involved in extensive design, planning, and evaluation. People were spread out, remote, and you had to balance the tension between ‘action-oriented speed’ and ‘maintaining extremely high standards’—the latter meaning delivering according to your planned cost structure, delivery methods, and designs. So you had to make trade-offs.

Our trade-off was to side with the customer, move fast to serve sellers and consumers, and we anticipated some cleanup work later. Did we exactly foresee it evolving into what it eventually became? No. Some things weren’t optimized, and I’m not sure we foresaw they wouldn’t be. But if you have good metrics and a culture of truth-telling—which is another crucial part of our culture—we bring data to the table, surface problems openly, communicate directly, and don’t allow anyone to gloss over issues. If you have such a culture, the indicators become obvious, and discussions are transparent.

Then you just figure out how to improve it, distinguishing between minor efficiency issues and structural problems. Once we deployed that batch of capacity, we went back and completely restructured key parts of the network—something that was initially ranked eighth or ninth on the team’s priority list, but later became much higher priority because it delivered real gains in speed and cost. Yes, fascinating. Kind of like when you first launched AWS—you launched it with the mindset of ‘build it and they will come,’ right? You built the entire infrastructure, then operated it. So they had to come, otherwise you’d be in trouble.

But now this is happening on a vastly larger scale: you and a few other companies—though I must say, mainly you, at least you’re certainly among the largest players—are building data centers and chip capacity at such a rapid pace that financial markets are saying: ‘Oh no, there isn’t enough money in the world to fund this.’ Then recently, we saw that tension in CoreWeave’s debt and equity markets. So how do you judge—given the steep demand curve and your rapid growth—how much is too much for this round of construction? And with supply chain competition—fighting to secure cooling, power, chips, all of it. Yes, that’s an excellent question. I wish there were a formula to simplify it—it would save our team a lot of time.

Well, I’d say this: in the first wave of cloud computing, when we launched AWS, there was a bit of that ‘we don’t know if people will use it, so let’s build it and they’ll come’ mentality. But internally, we had strong signals. Back then, around the early 2000s, many teams were moving far slower than we expected—despite feeling fast ourselves. But we couldn’t understand why it took so long. The reason was, our product teams spent 70% to 80% of their time on infrastructure tasks, which didn’t differentiate their products; only about 20% of their time went toward truly differentiating features.

So we realized every team was repeating the same work, reinventing the wheel. That’s why we initially conceived AWS—to help ourselves run faster internally. Our judgment was: if we built a large-scale, highly available, low-cost, highly flexible foundational service stack—like compute, storage, databases, content delivery, messaging—we could build faster in consumer businesses. Then we simply said, relatively speaking, our tech was good, and other companies likely had similar issues. But even before we launched externally, so many internal teams were interested—we knew we were on the right track. I don’t think anyone among us expected it to grow so fast externally.

In AI, you know, as you know, there have been wave after wave of people saying: ‘Now it’s finally here, AI has arrived.’ Even about six or seven years before the generative AI wave, people thought it was the era of AI. We built a whole suite of AI services—on AWS, it was already a decent business. But nothing compares to what we’re seeing now. In AWS’s early days, we thought we were growing fast and felt good about it—after three years, we hit about $60 million in annualized revenue. In the first three years of this generative AI wave, just our AI business on AWS exceeded $2.5 billion in annualized revenue. So it’s hundreds of times faster. We’ve never seen anything like this. So the question becomes: how much capex should you invest to chase this opportunity? I can tell you we’ve said that for 2026, we expect to spend about $22 billion in capex.

By the way, I remember about 15 years ago at a leadership breakfast meeting, we couldn’t believe a telecom company would spend $50 billion on capex—we said, ‘How can they spend so much?’ So these numbers, you know, are mind-boggling. And even with $22 billion in capex this year, we’re still leaving a substantial amount of demand on the table. What really helps us here is our visibility into customer demand. So if you look at AWS’s capex investment and cash flow cycle, there are two main components: one is data center investment, the other is server and network investment. For data centers, you must, you know, commit about two years in advance before you can place servers and start monetizing. The good news is, you only make this investment once, and it’s a long-lived asset—around 30 years.

But you have to plan years in advance. Servers and network equipment only require a few months’ lead time, depending on the component. So you have strong visibility into customer demand before you pull the trigger. If you find out you don’t actually need that capacity to fill the data centers, you won’t buy servers or network gear. On average, they break even in about three years, and their typical lifespan is five to six years.

Most of our deals are around five years. So when you’re building so many data centers so fast, you burn through a lot of capital, and your free cash flow faces headwinds in those years—until several years later, when all these data centers are live, monetizing, and you recoup your investment in servers and networks. But because of the cash cycle, in the early years of such rapid growth, it does create headwinds. So following that line of thought, you know, going back to when you launched AWS, data centers were probably the last thing anyone would believe could become a political issue. But now we’re in that situation, especially in the U.S.

You know, it’s interesting—I believe you’ve experienced this too—there’s no place in the world where people are more anti-data center than in American politics, regardless of the reason. So how do you think this situation will ultimately affect things? Do the arguments against data centers hold any merit? And how does that feed back into your strategy—what you build here, what you build elsewhere? Well, okay, I can give you some background I’m aware of—some I personally believe. Think about it: everything we do in the world today—stock markets, banks, healthcare, entertainment, shopping, education, business, ride-hailing—they’re all digital-driven, and they all use data centers—plain and simple. No, everyone uses them; I oppose myself—I’d rather not use my phone anymore.

But they all use data centers. And this is even before AI. AI is the biggest technological transformation of our era—considering we’ve lived through the internet, cloud, and mobile internet—this claim carries weight. So the reality is, to support the kind of explosive growth we discussed earlier in AI, the scale of infrastructure investment we need is, in this country and this world, comparable only to what we did building highways and bridges in the 1950s. I mean, what we’re experiencing now is on that scale. And this nation—believe me, the world is similar, but I’m specifically talking about the U.S.—cannot afford the cost of not building data centers. Otherwise, another country like China will establish all leadership positions and all rules there. I believe none of us would like that, whether domestically or globally.

So we must build data centers. Do I think the industry as a whole can do better? Of course. Things are moving fast, and we’re all learning together. But right now, there’s a wave of—let’s call them ‘half-myths, half-lies’—being spread, because it’s a topic where everyone can easily find a ready-made ‘villain.’ Politically, opportunists sometimes enjoy pointing fingers at a villain. I can give you examples. The golf industry or almond farming consumes far more water than data centers.

Just one data point: golf courses consume 200 times more water than our data centers; their two-day water use exceeds our annual consumption. So there are many such examples—this is a good one. Or you can talk about pollution. The reality is, when you register a new site and equip it with backup generators, you must report emissions assuming full 24/7 operation. But of course, that never happens. Backup generators run only 0.1% of the year—say, about 10 hours annually—for maintenance.

So you know, or you can talk about electricity pricing—when you look at different regions with large-scale data center footprints, some places see prices rise, others fall. But the real issue is: as a nation, we haven’t invested in our grid infrastructure at the necessary level. There will inevitably be significant growth, so something will push it to the investment threshold. The result is AI—because it’s growing so fast and at such a massive scale—it’s becoming that catalyst. But once you actually build the capacity, you’ll find that in most regions, electricity prices eventually stabilize or even decline, as seen in Indiana, Georgia, or Mississippi.

So there are definitely areas where we can do better. Another thing people often forget: large data center operators pay substantial taxes in these regions. Take St. Joseph County in Indiana—our taxes have already totaled about $3 billion. The community receives this money and spends it on things truly important to them. So the community benefits too. We’ve launched projects like Built Together—we’ve just committed to investing another $1 billion in communities with data center footprints over the next five years. So I believe we must continue working with communities on this. It must be a public-private partnership. But this round of construction is too critical to the nation’s future—imagine what would happen if we hadn’t built the highways.

a16z 20:44

And bridges—that would determine the nation’s economic prosperity and capability boundaries. So we must view this in a reasonable context, and we must be partners with the communities we’re building in. Let’s return to the capex question. That $22 billion—about $6 million per day. I just want to, you know, put it in perspective. I’m curious how you view AI demand right now—we chatted privately, and you described it as a...

Andy Jassy 21:10

Barbell structure. Perhaps you could clarify or position the current state of AI demand. Well, I’d say the first context 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 AWS’s early days; in the first three years alone, we achieved $2.5 billion in annualized AI revenue on AWS. So it’s truly massive. But simultaneously, if you look at the demand structure, it’s a classic barbell. One end is AI labs consuming massive, unprecedented compute power—scale I’ve never seen before. Then there are a few breakout applications, like ChatGPT, Claude Code, etc.

Andy Jassy 21:59

The other end is enterprises using AI primarily for cost avoidance and efficiency gains—such as automating customer service, fraud detection, or business process automation. The middle of the barbell is current enterprise production workloads—many running on AWS, already using inference today. But I’d say most haven’t yet adopted inference in any meaningful way—yet I believe this will become the absolute largest segment of AI usage. I think you’ll see this.

Andy Jassy 22:38

I can’t predict exactly how fast this will happen. It’s happening to a greater extent each year. But if you zoom out—five to seven years out—what will it look like? And I don’t think it’ll wait until five to seven years to happen; it never happens overnight. But if you zoom out to five to seven years, I believe a large, perhaps even majority, of the enterprise applications we use and exist today will adopt inference in a meaningful way. So we still have massive room for growth.

How do you see this? Because we’re facing an impending shortage of data centers, chips, memory—all of it—really by 2027, since everyone’s building now, but not all will be ready by year-end, and demand isn’t slowing. So how do you think about this—when large labs are willing to pay almost any price for power, cooling, chips; and you have long-time customers paying standard rates to run their businesses. Is there commercial pressure that makes you think: hmm, maybe we should buy out old customers and resell to new ones, since they’ll pay whatever we ask? And for old customers, I can’t double their prices—that wouldn’t be fair. How do you view this? It’s really strange—you have a group of customers who’ve been with you for 20 years, doing something, as you said, that will eventually shift toward inference; then you have new customers who are insatiable, willing to buy anything.

Andy Jassy 24:35

Yes, that’s a good question. You know, we approach this differently than you describe, but I completely understand. If you think about it, today’s data centers, power, and memory are limited—plus I don’t think this scarcity will end by 2027.

I believe—returning to capital investment—I think we can invest heavily without it being poor capital allocation, partly because 2026 is clearly sold out at this point, 2027 is nearly sold out. We always keep some capacity for on-demand use, whether in our inference service Bedrock or EC2—but it’s largely sold out, all major blocks are gone. 2028 also has a significant portion already sold. Demand is that big—you can’t… We have, if you look at AI business and what we call core business—core meaning non-AI parts—they’re both growing at extremely fast rates now.

And you know, if you only have so much memory, you have to decide how much to allocate to HBM—AI accelerators—and how much to DRAM—your core business. These are real trade-offs you must make. But what I want to emphasize is our starting point is customer-first.

For us, we have deep relationships and collaborations with customers on AWS. Like during the pandemic—our competitors, I believe, were thinking about optimizing within a quarter to maximize revenue in a world where everyone was cutting costs; we took a different approach. We proactively sent our teams out to help customers optimize costs, even though we knew it would weaken our short- and medium-term revenue. But you know, we wanted to help customers succeed because we strongly believe: if we help customers succeed long-term, our business will follow. The same applies to core business. You could theoretically calculate and conclude: due to scarcity and customer willingness-to-pay, you could extract more money from core customers. But they’re running their businesses on AWS—everything in their core applications runs on it.

We won’t squeeze our customers; we’ll ensure they can scale. Then our job is to secure as much capacity as possible so not only our core customers, but all those wanting to use AI on us, and those wanting to use us for AI, can access it—in a way that’s economically sustainable for us. So I really wish I could tell you supply tightness will only be a 2027 issue; I believe we’ll see it across multiple components for several consecutive years. We’ve always been aggressive, and I believe we’re deploying more capacity than anyone else because we’re deeply committed, very dedicated to remaining a key leader in infrastructure and AI—just as we’ve always been.

But as I said, we’re spending $22 billion this year, and we could have spent much more. That wouldn’t be poor capital allocation. And from mid-to-long term, I believe we’ll still fall short of meeting all demand. The current demand is truly staggering. Then continuing this thread: you did something in 2015—you acquired Annapurna Labs. At that time, you probably weren’t anticipating the massive AI demand we have now, since AI demand was minimal in 2015. What drove the decision for that acquisition? And how has it generated returns in this new world? Yes, I think the Annapurna team was one of the most talented groups I’ve had the privilege to work with—one of the best acquisitions in Amazon’s history—if not the best.

Yes, we foresaw some aspects, but I don’t think we acquired Annapurna specifically targeting AI chips. Here’s how it worked: when you reach a certain scale, you gain visibility into customer behaviors at massive scale, and you realize that if you can vertically integrate and build certain capabilities in-house, you can deliver more value to customers and achieve better economics for the business itself. So Annapurna’s initial focus at Amazon was around core EC2 servers: we realized we were spending too much server resource on virtualization, networking, security, and storage, and we wanted to allocate nearly 100% of server resources to providing compute for customers.

So even before acquiring them, we were already collaborating with Annapurna on the NIC—offloading those components to the NIC so we could keep as much server capacity as possible for computation. We loved the team and loved the product we co-created. We realized the thing we were buying would eventually be integrated into almost every EC2 server—indeed, it ended up in all our servers.

So our initial reason for acquiring them was that it would become a core capability in our compute capacity; also because the team was exceptionally talented at chip design, and we believed they could build other types of chips. At the time, we had a very deep relationship with Intel—we still do. But the problem was, Intel held almost all the CPU market share. When you have a partner with near-total market dominance, they often don’t focus on improving cost-performance; instead, they focus on features that increase their market share and revenue.

By the way, this is understandable—this is part of how markets operate and contributes to the ‘innovator’s dilemma.’ So we wanted to push cost-performance, but it was hard to get their attention. But our customers wanted to move more local workloads to the cloud and demanded better cost-performance. So we tasked the Annapurna team with building a CPU chip. They created Graviton—today, 99% of the top 1,000 EC2 customers use it extensively. Its cost-performance is about 40% better 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 space: there was indeed a dominant, excellent supplier. We had a deep collaboration with them—we still do, and we’ll continue for the foreseeable future.

But they have massive market share and many different products and priorities. And we knew our customers needed better cost-performance—this is why we formed a separate team within Annapurna to build AI accelerators, called Trainium, whose cost-performance is about 30% to 40% better than comparable processors. And you know, when we talk about the middle of the barbell—we just discussed that—what could change the middle and enable more current enterprise production applications to adopt inference? Several things. Part of it is talent: some people are used to post-training and fine-tuning, and also need the organizational ability to prioritize, know what to attack, and how to execute.

But a major part lies in: over the past few years, inference costs have dropped significantly, but in my view, they’re still too expensive by an order of magnitude—insufficient to enable widespread adoption. Because if you require enterprises to rewrite their current CPU-based applications to run on AI accelerators, and the cost difference is significant, they’d need extremely high incremental gains from AI accelerators to justify the rewrite—creating a much higher barrier. But if you can lower inference costs, that barrier drops dramatically, and all applications become smarter, more powerful, and better thanks to AI accelerators and inference. So a very critical factor for customers widely adopting AI is lowering inference costs—and our chip business is a major piece of that. Yes, indeed. I can tell you from our own company—it truly gets very expensive. One employee spent $2 million over the past few years—on valuable work.

But yes, the feeling at the time was: well, this inference really isn’t cheap. Then you did something: made significant investments, especially in Anthropic and OpenAI—huge investments, I recall $18 billion for Anthropic, $500 billion-ish for OpenAI. What drove you to do this? And such investments, along with Nvidia and others, are making Wall Street nervous, reminding them of the bad old days of the dot-com bubble. How do you see this—especially considering they’re also your customers? Why is this rational? Both companies are exceptional teams and exceptional companies; their business growth rate—revenue-wise—at least—well, I don’t think we’ve ever seen anything like it. Really, it’s incredible.

So for us, it’s a combination of factors: what kind of teams they are, the technology and models they’ve built, some of the applications they’ve developed; our belief in AI’s transformative potential; our belief that they’ll remain two of the most successful companies in this field for a long time—of course, not everyone wins, but many very successful companies will emerge, and we believe they’ll be two of them. Then there’s our deep collaboration. Let me start with Anthropic—we’ve collaborated longer. We have many customers running Anthropic on our inference service Bedrock. And I believe the most successful use cases are customers running Anthropic on Bedrock. We have deep chip-level collaboration with them—they’ve been building on Trainium for some time. I believe this is great for Anthropic, great for our Trainium team, and both sides have learned a lot from each other. So our deep collaboration—whether it’s our belief in AI’s future, in these companies, or the partnership itself—makes investing as part of that collaboration reasonable.

Then with OpenAI, we simply had too many customers wanting to use OpenAI on AWS, but couldn’t due to prior commercial arrangements. And for a long time, I believe both companies had mutual respect and a desire to collaborate. So this was one of those situations: after months of conversations—enabling OpenAI to run on Bedrock; we co-developed a unique product, the stateful runtime environment, which I believe will be how all intelligent agents are built in the future—you’ll need state; and like Anthropic, they made multi-gigawatt, multi-year 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 deepening our collaboration and investment reasonable. Switching topics, talk about how you use AI to build a company. I remember you said the Bedrock team rebuilt your inference engine in 76 days with just six engineers—something that would have taken 40 people a full year. That’s a huge difference.

And you know, Amazon is famous for the ‘two-pizza team’ concept. So what’s your current organizational design philosophy? Considering this technology—it seems to truly change everything. Over the years, we’ve become extremely specialized: designers, product managers, frontend engineers, backend engineers, etc. Now it seems everyone is becoming generalists again. From an organizational design perspective, how do you see this? Yes, I agree with what you said. I think the Bedrock experience you mentioned was eye-opening for everyone here.

We’ve spent a lot of time sharing that story internally, to inspire and remind people what’s possible. Because over time, eventually, this happens—me, I joined the company 29 years ago. I started as an assistant product manager, responsible for, I don’t know, five products. Today, as the company succeeds and grows, and products grow larger, you reach a state where a product manager handles just one product—and only one slice of it—say, pricing. Everyone becomes more specialized, and everyone accepts: if you want to scale and succeed, this is the necessary way to work. But that’s not, and when you look at what AI is doing, it absolutely isn’t.

Take just the past week: we’re doing something around intelligent agents in consumer business. We have a lot of investment there, but we had a favorite idea—we decided Thursday to try building a real prototype. Result? Four or five people worked over the weekend, and by Sunday, it was running at 80% release-ready state. This is real—it’s a completely different world. So I think what this means for the organization’s future: first, I believe the number of humans doing today’s work in the future will be fewer than today. Just look at how much code is now written by agents, how many customer service, business intelligence, and research tasks are being automated.

Yes, imagine forward: how will spreadsheets be made, how many lawyers will you need—just, I believe, you won’t disappear, but you won’t need nearly as many humans. In the past, as business scaled, we kept adding people; in the future, it won’t be like that. New jobs and different jobs will emerge—some we know, some we don’t. With AI, robotics, and space development, new roles will appear. There will be all kinds of new jobs, but fewer humans doing today’s tasks. I believe teams will be much smaller. You’ll have smaller teams, people capable of cross-functional work; you won’t assign one person just one tiny slice, because models can help you achieve 80%. In many of these different functions, I believe every team will start from a much higher baseline than before, so I think you’ll be able to invent more broadly and more rapidly.

I truly believe a lot of things must change. Startups today are mostly founded from the beginning as agent-native or AI-native. And all established mature companies must figure out how to build at the same speed. I believe some will figure it out. And you always have this situation: when you have a large business with strong capabilities and loyal customers, some companies move very slowly to figure it out—and in my view, they’ll be more vulnerable when facing startups.

And I believe, I predict, companies will have to rewrite almost all their code to enable agent capabilities; they must allow agents to update different code segments, and you need to establish constraints there. I find the security aspect particularly interesting. If you look at the recent tensions around models, they’re focused on more fantastical scenarios.

Incidentally, those scenarios are possible: could a malicious botnet take over everything? I suppose it’s possible; could you inject malicious intent into robots? Also possible. Personally, I believe as an industry, with public and private sectors working together, we’ll solve this. But you have to start by focusing on certain things. Yes, but interestingly, I think discussions about these more immediate risks are still insufficient—they’re mainly centered around cybersecurity. These models are so powerful they can find far more vulnerabilities than humans ever found before. Maybe you can argue whether humans can find them, but the fact is, they haven’t. So looking ahead, I believe application responses to urgent security threats must be measured in minutes.

Today’s situation is: you discover a vulnerability, someone classifies it, someone develops a patch; even if that part is fast, most companies take a long time to deploy due to dependencies. In the future, you must deploy in minutes. So if you don’t truly restructure your applications to be updatable by agents, you’ll have no chance in that new world. So I believe all organizations will be re-invented, and companies prioritizing this will fare far better than those that don’t.

Yes, and it seems, as an industry, we haven’t built security architecture well into everything we design—from large systems to embedded devices. The reckoning is coming. But the good news is, this presents a huge opportunity for new companies, new attempts, and people doing things. But yes, as we said, this threat—those theoretical, thought-experiment-type threats—well, maybe they’ll happen, maybe not. But this one is clear and real.

Yes, almost every partner and customer we’ve talked to 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 identifies numerous such vulnerabilities and generates patches. But this, you know, is easy to say, hard to do. Actually, because many companies and many engineers sometimes find focusing on security unexciting, they often wait until problems arise. But if you’re protecting critical data, that’s a bad idea. And I believe 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 working with other engineers, they’re scarce resources, etc.

As a result, in most companies, most software companies, their security practices when writing software are generally poor. I believe this must change. The good news is, we have tools that can help solve this problem. Following your point, the good news is: you need to hire penetration testers, but you can only hire so many security professionals, and they can only cover so many applications. Models plus automation will allow you to perform penetration testing across your entire attack surface better, and you’ll find things you might never have discovered otherwise. So one thing we observe in the market is quite interesting: with AI, you can do many things. One is using it to cut costs, reducing the need for accountants or lawyers, etc.

Indeed, many venture capital firms now say: okay, let’s buy old companies and cut their costs. This has never really interested us much because we think: well, if that’s your idea, your idea is too limited. Another is: okay, we’re going to do something the world has never seen before. That’s very exciting. But there’s a third: doing something like Alexa—having a product—a successful product—and saying, okay, let’s completely rethink this from scratch because we have this new platform. How do you get to that point? When is it rational? Yes, I think you’ve drawn a nice three-part division. And I believe, as we discussed earlier, organizations will truly change and become leaner.

If you do it right, it will empower your truly outstanding builders and small teams to create the second type of thing—things people have never seen before—you can do it. And when you have that kind of experience—getting a new idea to 80% completion in four days—it’s incredibly motivating for people; they’ll genuinely spend more of their spare time and mental cycles thinking of new ideas. To me, this is one of the most exciting aspects of what AI is doing.

But I also believe that for established companies with successful customer experiences already scaled, if they view AI merely as incremental improvements to the experience, they’ll likely not have those winning experiences in three to five years. Because I believe every experience we know—even those that seem hardest to imagine changing—will be significantly different in a few years due to AI.

So this means, if you don’t just want incremental improvement, you must find a way for teams to completely reimagine these applications from scratch—assuming you’re building them today with all available new technologies. In some cases, take Alexa—my view is, we must do this quite clearly. Alexa’s ‘brain’ was indeed creative when it was built. But with generative AI, the capabilities of that ‘brain’ have changed entirely. So if you want to provide an outstanding customer experience and maintain relevance long-term, you must rewire Alexa’s brain. So that’s quite clear. Incidentally, although quite clear, doing it is actually hard—because teams and the things they’ve built have emotional attachment; they’re proud of what they’ve built, they have hard-earned scars from what worked and what didn’t.

Getting them to completely start from scratch and reimagine requires time, because teams will say: well, we don’t do it that way; we tried that before, it didn’t work. Or similar. But I think the harder part is: take our other consumer experiences—our retail business, or Prime Video, or Fire TV, or Prime Music—these experiences now have much better technical solutions, but that’s not immediately obvious. But I actually believe every one of these businesses needs to be reinvented; and we already have teams working on completely reimagined versions of these businesses.

So do I think we’ll get it right the first time? Probably not. We’ll iterate our way in, but it’s already generating all sorts of new ideas and customer experiences we find very compelling. Then the interesting race will be: do you want a completely different experience—almost like a new app—or do the new things move fast enough that you can add them incrementally into the existing experience, gradually and substantially changing the overall customer experience over the next two to three years, as you identify what works? But I do believe, if you want to be a successful leader, each of these experiences must become different in the next few years.

Yes, this is a very challenging leadership dilemma. It reminds me of when I was a young engineer. I argued with someone about something. My boss said: listen, Ben, you must understand one thing—before you write the first line of code, everything is open to discussion; afterward, it becomes faith. I guess you’ve experienced this with Alexa too. But interestingly: all that knowledge, all that experience, all those elements used to build the original thing—now you’re saying, okay, we’re willing to abandon a large part of it, to enter this new world.

And incidentally, it’s the same for customers: you’ll have a large group of customers who love the current experience, right? Yes, every major transition—certainly based on my experience, maybe you can find exceptions—but almost every major transition requires a massive shift: technology, product, customer experience, and business elements must be compelling enough to drive change. But even so, it doesn’t happen automatically. This is a massive change management effort—like migrating from on-premise to cloud: similar kind of transformation. Many infrastructure leads are very proud of what they’ve built, know it intimately, are at the center of action, don’t want to learn new things.

It will be the same thing—exactly what I mentioned earlier. For re-inventing all your customer experiences, it will be the same. We’re all proud of what we’ve built in these experiences over the past decade, we each have 100 things we know how to improve, we’re excited about it, we spend all our waking hours on it. To think, ‘how do I completely start from scratch?’—that’s hard both intellectually and emotionally. Yes, no doubt.

a16z 53:06

Also, Andy, we recently incubated Harwood’s and Jason Academy, helping young people prepare for the transformation we’re discussing. You said AI makes building easier, but young people still need to learn, you know, genuine professional skills. I’m curious: if you were 18 in 2026, what advice would you give yourself to prepare for the transformation we’re entering?

Andy Jassy 53:29

Well, I have a 25-year-old and a 22-year-old child. So I think the 22-year-old was really taking off at age 19. I quickly realized: you can’t tell them not to use AI in education. The ship has sailed, or the train has left the station. And I believe universities should wisely not fight it. I think it’s just a tool people will use.

Andy Jassy 53:54

But at the same time, I believe all young people—and I’d say this applies to us too, we have a writing culture, many people use AI to write—can tell what’s AI-generated: it’s just off, those parts don’t connect well. And you know, so my advice to young people—also to my kids—is:

Andy Jassy 54:18

You can’t avoid it, you know, you can’t expect it to disappear. It will have a massive impact on your life, on every experience you try to use, on your work; and it will let you invent in ways we could never do before. So you really need to understand it, you need to teach yourself it, you need to use it extensively. You need to experiment. Try doing things beyond your comfort zone.

Andy Jassy 54:48

My two kids aren’t developers, but they’ve built apps using AI. Treat it as the focus of how you structure your day and use tools at work. But I don’t think it will replace critical reasoning and thinking. And I believe this applies to universities too. I think of—take business school as an example. You get three cases to analyze, you do a lot of reading, take notes, identify the key problems. Then you get a bunch of attachments, you also take notes on them. Then you integrate what you think the company should do. And you know, often there are people who don’t do these assignments—don’t read, don’t research, don’t take notes. And all of this, AI can help you with—it can get everyone faster to that higher starting point.

Andy Jassy 55:52

But I also say, truly connecting those dots, those fragments—whether forming a strategy, an idea, a plan, a new product, or a new customer experience—I don’t believe AI will do this for most people, most of the time. I’m not saying it never will.

Andy Jassy 56:10

I still believe you need human judgment, creativity, inventiveness, experimentation, and the ability to think about how all the pieces fit together to ultimately create stunning experiences and businesses. So I believe you should use these tools—they’ll be tremendous enablers—but I still believe critical reasoning and thinking will remain vital in everything we do. Wrapping up. Your business is like…

a16z 56:43

From satellite connectivity to robotics to healthcare—so diverse. I mean, this ambition is unbelievable. Think about how this business started—selling books online—and now it’s this. How do you see, well, what ties these different…

Andy Jassy 57:00

business lines together? Or how do you decide which ones are worth investing in for years? Or how do you ensure you’re not spreading yourself too thin? Or how do you understand this extremely broad landscape?

Andy Jassy 57:11

Well, I’d say a few things. First, I believe we’re living through the largest three transformations of our lifetime—perhaps in human history: AI, robotics, and space exploration. They’re all massive and will fundamentally change many customer experiences we all encounter. And we’re right in the middle of these three major transitions.

Andy Jassy 57:40

If you look at what we do in retail, and how we use robotics, and what we do on the AI side—both enabling internal and external builders to create successful AI applications and agents, and re-inventing all our consumer experiences.

Andy Jassy 57:58

If you think about Amazon Leo—we’re building a low Earth orbit satellite constellation for the 500 million homes worldwide without broadband connection—there’s a series of incredible transformations happening there, and we have the opportunity to help customers leverage them to improve their lives. And we do this in a way that helps our customers and builds long-term, sustainable business for us. And you know, I believe we have the opportunity to be more relevant to more customers in more areas. Not to mention what we’re doing in grocery, autonomous ride-hailing with Zoox, healthcare, robotics, and so on.

Andy Jassy 58:47

Regarding our organizational approach, over the years we’ve had many internal debates, and many outsiders have asked the same questions: is this too spread out? I think it depends, to some extent, on how you organize yourself and how much you’re willing to invest. Because we don’t have the same people managing multiple businesses—we organize in a 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 levers, and as long as you have differentiated ideas in areas underserved, you have the opportunity to pursue many such things simultaneously—and we’ve done this for a long time. Then it comes down to how much you’re willing to invest.

Some companies—completely understandable—say: we’ll stick to our core, focus on our core business, and grow it. By the way, that can also be a very successful strategy. And we’re very focused on continuing to grow and strengthen our existing businesses, and I believe we’ve done quite well, with good speed. But with so many other customer experiences, we know customers really want us to help solve them, and we can—these will help our customers and long-term transform our business. So we’ll continue pursuing them. A motivating close—endless ambition. Andy, thank you so much for joining us. Yes, thank you for having me. Thank you, everyone.

Source: Wall Street Journal