AI computing power sold out as soon as it was packed! Amazon (AMZN.US) Q3 speech: capital expenditure will increase next year, betting on long-term returns

The Zhitong Finance App learned that Amazon (AMZN.US)'s revenue for the third quarter of 2026 reached US$2019 billion, up 12% year on year, and operating profit increased 61% to US$28.1 billion. Chief Financial Officer Brian Olsavsky said that the company's operating margin reached a record 13.9%; AWS revenue increased 39.3% year over year, with an annualized operating rate of US$184 billion and uncompleted orders of US$578 billion.

While AI and cloud demand are driving accelerated growth in performance, the company's cash capital expenditure in the third quarter reached 75.9 billion US dollars, and is expected to be around 220 billion US dollars for the whole year, and expenditure is expected to rise significantly in 2027. Management said in the earnings conference call that the current infrastructure investment cycle aims to expand revenue, free cash flow, and return on investment capital over a long period of time, but the current profit margin of AI services is lower than the core cloud business, and component prices are still putting pressure on it.

AWS accelerates growth driven by AI and core cloud requirements

CEO Andy Jassy said the results were mainly due to the continued acceleration of Amazon Cloud Services (AWS) growth, retail sales growth, and continued investment in artificial intelligence infrastructure. The company also expects net sales of 240 billion to 247 billion US dollars for the fourth quarter and operating profit of 28.5 billion to 33.5 billion US dollars.

Specifically, AWS's total revenue for the quarter was 46 billion US dollars, up 39.3% year on year and 3.8 billion US dollars month on month. The cloud business's annualized revenue operating rate reached US$184 billion. AWS's operating profit was $17.8 billion, and the operating margin was 38.7%.

Jassy said that the AWS growth rate has accelerated for the sixth consecutive quarter and reached the fastest level in nearly five years. He mentioned that both traditional cloud computing and AI services are in high demand, and pointed out that the amount of outstanding orders from AWS reached 578 billion US dollars, a three-digit year-on-year growth rate.

“Almost as soon as we deployed the computing power, customers sold it out,” Jassy said, adding that Amazon deploys an AI accelerator approximately every 14 seconds.

According to the company, Amazon BedRock — its managed service for accessing the basic model and building apps based on it — has become the fastest growing multi-billion dollar service in AWS history. According to Jassy, Fortune 500 companies' spending on Bedrock increased nearly seven times year over year, and total token processing volume increased nearly 17 times. Amazon also recently added the OpenAI model as well as hosted versions of the Kimi, GLM, and SpaceX models to the platform.

Jassy also emphasized the company's need for self-developed chips. Graviton and Trainium's revenue promises increased more than 20 times year over year, and customers using Amazon's self-developed chips saved more than $2 billion this quarter, he said. Amazon plans to make Trainium available outside of AWS starting at the end of 2027.

Looking ahead, Jassy said that Amazon expects AWS to grow by more than 40% in both 2027 and 2028, and believes that the business is expected to become a trillion-dollar business within five to ten years. He said that this outlook reflects the visibility of large-scale customer computing power commitments, the continued advancement of enterprise cloud migration, and the growth in the productive use of AI inference.

Capital expenditure rises as AI infrastructure is built

Amazon's third-quarter cash capital expenditure totaled $75.9 billion, mainly for AWS and AI infrastructure. The company continues to expect a cash capital expenditure of approximately US$220 billion for the full year of 2026, most of which will be invested in AWS. Jassy said the 2027 expenditure is expected to increase significantly.

Management said it sees the current investment cycle as a long-term opportunity to expand revenue, free cash flow, and return on invested capital. Jassy said that the company continues to shorten the time required for infrastructure investment to achieve break-even, but he mentioned that there is a difference in return between core cloud and AI infrastructure.

Olsavsky said that AWS's profitability benefits from capacity optimization, fixed cost management, self-developed chips, and network investments. However, he pointed out that the current profit margin of AI services is lower than that of AWS's core business, putting structural pressure on the overall profit margin. The price of spare parts also remains a negative factor.

Jassy said that Amazon owns or long-term leases for most of its data center production capacity, and the use of self-developed chips, has a structural cost advantage over cloud service providers that rely more on hosting or “new cloud” arrangements.

Retail profit margins, faster delivery, and grocery business expansion

Amazon's North American revenue increased 6% to 112.9 billion US dollars, and international revenue increased 8% to 43.1 billion US dollars after excluding exchange rate effects. Global paid sales increased 12%.

Results are affected by the Prime Day schedule. This year, the promotion was held in the second quarter in most major markets, while 2025 fell entirely in the third quarter. Olsavsky said that excluding this change in time, the year-on-year revenue growth rate for the third quarter will be about 400 basis points higher.

North American operating profit totaled US$8.7 billion, with a profit margin of 7.7%; international operating profit was US$1.6 billion, with a profit margin of 3.7%. Amazon said that despite rising fuel and mainline transportation costs, the company reduced overall contract fulfillment costs by optimizing inventory layout, shortening transportation distances, and package consolidation.

Shipping costs increased 11% year over year, while sales increased 12%. Olsavsky said that excluding fuel and mainline transportation inflation, freight rates rose 8%, which is about 400 basis points lower than the sales growth rate.

Jassy said that Amazon delivered more than 1.5 billion products to Prime members through same-day or next-day delivery this quarter. The company's ultra-fast delivery service, Amazon Now, has now been launched in more than 450 cities, and the number of products ordered through this service has increased tenfold over the same period last year.

Groceries and daily necessities remain the focus. Fresh groceries and daily necessities are growing at about three times the rate of the rest of the Amazon store business, accounting for about half of sales this quarter. Management said that customers using Same-Day Fresh Delivery bought about seven times as many daily necessities as those who did not use them.

Advertising, Pharmacies and Fourth Quarter Outlook

Amazon Ads third-quarter revenue was $19.6 billion, up 12% year over year. Jassy said that the growth rate was affected by Prime Day moving to the second quarter, the NBA program schedule on Prime Video, and the comparison base for search engine marketing project adjustments over the same period last year. He said that Amazon expects the growth rate of advertising revenue to rise to a high of close to 20% in the fourth quarter.

Jassy said that Amazon Pharmacy customers increased by more than 60% year over year, and repeat customers increased by more than 50%. The company now provides a three-hour prescription drug delivery service in more than 350 cities.

For the fourth quarter, Amazon expects the exchange rate to reduce the year-on-year increase in net sales by about 60 basis points. Olsavsky also said that fuel and main line transportation inflation is expected to be higher in the fourth quarter than in the third quarter.

Amazon Leo's commercial services are still expected to begin initial operations before the end of the year. The company expects to begin capitalizing part of Leo production and launch costs in the first quarter of 2027; therefore, it is expected that the fourth quarter will continue to cost most Leo-related costs.

Full summary content

(1) Host (operator)

Thank you all for your patience. Hello everyone, welcome to Amazon's Q3 2026 earnings conference call. (Operator Reminder) This meeting is being recorded. In the opening remarks, Mr. Dave Fildes, Vice President of Investor Relations, was invited. Thank you, please speak up.

(2) Dave Fildes -- Opening Address by Vice President of Investor Relations

Hi everyone, welcome to our Q3 2026 earnings conference call. Our CEO Andy Jassy and CFO Brian Olsavsky joined us to answer questions today. As you listen to today's conference, we recommend that you also prepare our press release, which includes this quarter's financial results, metrics, and reviews. Please note, unless otherwise specified, all comparisons in this meeting are for the same period 2025.

Our statements and answers to questions represent management's views as of today (October 29, 2026), and include forward-looking statements. Actual results may vary significantly. For more factors that may affect our financial results, please refer to today's press release and our filings with the US Securities and Exchange Commission (SEC), including our most recent 10-K Annual Report and subsequent filings.

We may discuss certain non-GAAP financial measures during this session. In our press release, slides accompanying this webcast, and documents submitted to the SEC (all posted on our investor relations website), you will find additional disclosures about these non-GAAP measures, including adjustments between these metrics and comparable GAAP measures. Our guidance incorporates the order trends we have seen so far and the assumptions we currently feel are appropriate. Our performance itself is unpredictable and may be substantially affected by many factors, including fluctuations in foreign exchange rates and energy prices, changes in global economic and geopolitical conditions, tariffs and trade policies, resource and supply fluctuations (including memory chips), customer demand and spending (including the impact of recession concerns), inflation, interest rates, regional labor market constraints, world events, growth rates in internet and online commerce, cloud services, and emerging technologies, as well as factors detailed in our submission to the SEC.

Our guidelines assume (among other things) that we will not enter into any additional business acquisitions, restructuring, or legal settlements. It is impossible to accurately predict demand for our products and services, so our actual performance may differ materially from our guidance.

Next, I would like Andy to speak.

(3) Andrew Jassy - President, CEO and Director's keynote address

Thanks Dave. We reported revenue of $2019 billion, up 12% year over year. Operating profit was US$28.1 billion, up 61% year over year. The third quarter was another strong quarter for Amazon. On the AWS side, our plans are working. Although we added a large amount of production capacity during the same period last year, AWS achieved a year-on-year growth of 39.3%, accelerating growth for the sixth consecutive quarter, and the fastest growth rate in the past five years. The backlog of orders reached US$578 billion, achieving three-digit year-on-year growth and is highly diversified.

We install an AI accelerator approximately every 14 seconds, and our customers run out of it almost as soon as we install it. Our core compute and storage revenue recorded the fastest growth rate in years. The operating margin of our core business has been strong for a long time, and now the profit margin of the AI business is also improving rapidly, clearly ahead of the level of the core business at a similar stage of development.

Despite rising costs of components such as memory and significant depreciation from our capital expenses, we continue to accelerate the recovery of capital expenses, and are more confident than ever: our investment in AWS will result in a much larger revenue and free cash flow business than we could have imagined two or three years ago, with a strong return on invested capital.

We are forming a very attractive flywheel at AWS. To borrow Jim Collins's definition of a flywheel in “From Excellent to Excellent” — if you do A, you'll almost certainly do B; if you do B, you'll almost certainly do C, and so on.

On AWS, leading functionality, security, and operational performance will inevitably drive customers to run applications and store data on AWS; this will inevitably drive AWS' AI requirements (because customers want their production reasoning capabilities to be close to the rest of the data and applications); this will also drive core AWS requirements (because AI needs to use core services for enhanced learning, post-training, and tool calls); and this in turn will inevitably drive reinvestment in our broader AWS functionality.

This is a different model from mainly hosting GPUs or renting data center capacity for a short time. It is a long-term platform strategy that can broadly help customers invent and evolve customer experiences, stabilize operations, and establish reliable partnerships that can last for many years. At the same time, whether supply is scarce or not, it can bring us very attractive long-term economic benefits.

There are many examples showing the progress of AWS; I'll just share a few. On the core business side, it's important to remember that the vast majority of global IT spending (around 85%) is still deployed locally, and we're in a 10-20 year cycle, and this pattern will move to the cloud. The traditional view is that critical workloads such as mainframes or storage area networks will never migrate to the cloud. But these loads are migrating, as AWS has built unique cloud alternatives and AI-driven migration capabilities to make this transformation easier. Businesses are getting lower capital expenses, lower operating costs, faster innovation, higher developer productivity, and better data accessibility, making it easier to use AI.

Our core business margins are attractive because we've learned how to operate efficiently over the past 20 years. As our core business grows, we continue to reinvest in core AI capabilities. For example, as customers generate more and more data and want to derive value from analytics and AI, the complexity of moving data between data storage and analytics engines is burdensome and expensive. With the acquisition of DuckLabs, DuckDB will become the generic SQL dialect for AWS storage and analytics, simplifying data processing with a modern analytical engine — fast, easy to use, and ubiquitous. Customers only need to write an SQL query once and execute it on any AWS analytics service.

For the first time in the field of analytics, customers will achieve zero migration or zero portability costs between search engines. This is a big deal. As you can imagine, this will also help agents make efficient queries.

On the AI side, we continue to expand the capabilities of the inference service Amazon Bedrock, which has led to significant growth — Fortune 500 customers spent nearly 7 times on Bedrock, and the overall amount of Bedrock tokens increased nearly 17 times year over year. Bedrock is the fastest growing multi-billion dollar service in AWS history. We recently introduced the OpenAI model to Bedrock, and usage is growing rapidly. We've also just launched Bedrock hosted versions of the Kimi, GLM, and SpaceX models.

Today, when we talk about AI, we can't help but talk about security. This is another reason why customers continue to choose Bedrock. We designed Bedrock to ensure that customer data remains secure and private. Inputs and outputs are never shared with model providers, and are never used to train the underlying model. As customers fine-tune the underlying model, we create private copies specifically for their use, and the data is fully encrypted both in transit and at rest.

As cutting-edge models become more powerful, they are discovering previously undiscovered software vulnerabilities faster than ever before. We built the AWS continuation to discover, sort, verify, and fix code bugs at machine speed. Customers such as Capital One, IBM, Infor, Luminate, and Robinhood are already using it. In Q3, we announced the integration of it into our coding tools—Anthropic's Claude Code, OpenAI's Codex, and AWS's Kiro. Customer interest in this new security service is beyond anything we've ever seen.

Agents are being deployed at an increasingly rapid pace, and many of them are built on top of the Bedrock Agent core building block. Tens of thousands of customers are currently using AgentCore, and the number of agents deployed in the third quarter was double that of the previous three quarters combined. In addition to building their own dedicated smart devices, most companies also use out-of-the-box smart body services, and we provide a number of services that customers love. Kiro for coding, Amazon Connect for contact center agents, and Amazon Transform for migration. They're all rapidly iterating and winning over customers. The same is true of Amazon Quick (our smart AI work partner), which is growing rapidly. Large companies like Mondelez, EchoStar, 3M, Southwest Airlines, and DXC choose Quick because it's built for work and doesn't compromise on customer data.

Finally, our chip business continued to maintain three-digit year-over-year growth, and Graviton and Trainium's revenue promises each increased more than 20 times year over year. AI continues to be Graviton's tailwind, because post-training reinforcement learning and intelligent tool calls mainly run on the CPU (not the AI accelerator), and Graviton is the most powerful CPU chip, which is 30% to 40% more cost-effective than similar products.

In the third quarter alone, customers using our custom chips saved over $2 billion. We've also just open-sourced TorchNeuron, so developers can run the existing PyTorch code most of them already use on Trainium without having to rewrite any code. As we launch a new generation of chips, demand for previous generation chips remains strong. After the launch of Graviton 5, the backlog of orders for previous-generation Graviton instances still more than doubled month-on-month. Compared to the first half of the year, the short-term capacity of the previous generation training instances was sold at a premium.

Due to this strong demand for training, customers continue to encourage us to offer this capability outside of AWS — we plan to do so by the end of 2027. Overall demand for AWS capacity remains unusually strong. Businesses are still in the early stages of using inference at scale, yet inference accounts for more than half of AWS AI revenue and is still growing. We also have a lot of training income. We still expect to spend approximately $220 billion in cash capital in 2026, the majority of which will go to AWS, and expect to invest more in 2027. As I mentioned on my last conference call, we have a clear plan for a great financial return.

AWS is a business with an annual revenue operating rate of $184 billion, yet the opportunity is still extraordinary. With our continued pace of innovation, new core and AI capabilities, the will of enterprises to run their core and AI workloads on AWS, and our new production capacity every year, we believe AWS will grow by more than 40% each year in 2027 and 2028, and become a $1 trillion business in revenue over the next 5 to 10 years. We are in the very early stages of this inflection point, our strategy is working, and I expect it to continue to work for the foreseeable future.

In terms of retail (store) business, I think some people underestimate the progress this business has made in recent years and its current scale. As far as I know, no other retail business can reach the size of Amazon (hundreds of billions of dollars, double-digit operating profit margin including advertising) and grow at this rate. Since the beginning of the year, the number of units sold has increased nearly 15% year over year, and is based on the huge base of sales of more than 30 billion products last year. In absolute terms, we've added more than $90 billion in total sales over the past 12 months. We are growing significantly faster than other retail competitors, adding about twice as much total sales each quarter as our second-largest competitor in the US, and continuing to seize retail market share.

Over the past few years, our North American and international division operating margins have each increased by nearly 10 percentage points, which has enabled us to reinvest and make the customer experience better. Keep in mind that it's still early days. The vast majority of retail sales (nearly 80%) still take place in physical stores. Over time, this pattern will eventually shift to e-commerce. We are driving this transformation by rapidly building an efficient and fast fulfillment network, providing a delightful customer experience while bringing significant free cash flow and return on invested capital.

From faster delivery, to unit service costs, to groceries and pharmacies, to AI shopping, every one of our priorities is a growth amplifier, and we're making full efforts. As we invest in the Same-Day Delivery network and ultra-fast Amazon Now and Amazon Prime Air services, our delivery speed continues to accelerate. We had the fastest third quarter in global history, delivering more than 1.5 billion items to Prime customers with same-day or next-day delivery in the third quarter. We are encouraged by our customers' enthusiasm for rapidly promoting Amazon Now around the world. Amazon Now is now online in more than 450 cities, and the number of products ordered through Amazon Now has increased tenfold year over year.

We are expected to reduce unit service costs for the fourth year in a row, which allows us to continue to meet or beat our competitors' prices and provide a wide range of fresh food and everyday necessities in a cost-effective manner. Faster speed is driving our grocery business, including fresh produce and daily necessities, with total sales growing to nearly $250 billion over the past 12 months. In the third quarter, the growth rate of fresh food and daily necessities was about 3 times that of the store's other businesses, accounting for about half of the total number of units sold. Customers who buy fresh food with Same-Day Delivery buy about 7 times as many daily necessities as customers who don't use it. Active grocery customers buy non-grocery categories such as home and clothing nearly 4 times more often than non-grocery customers, and spend more than 2 times as much on these categories.

Amazon Pharmacy (Amazon Pharmacy) is also growing rapidly. The number of customers using the service increased by more than 60% year over year, the growth rate continued to accelerate in each quarter of 2026, and the number of repeat customers increased by more than 50% year over year. J.D. Power recently rated Amazon Pharmacy as the best mail order pharmacy for the third year in a row. We're continuing to improve and now offer 3-hour prescription delivery in over 350 cities.

Over the past 12 months, more than 400 million customers have used our smart shopping assistant, Alexa for Shopping, and customer interactions increased 570% year over year. US customers who use Alexa for Shopping spend 50% more per order than customers who don't. It's responsible for researching products, comparing options, reviewing a full year's price history, providing personalized gift recommendations, and tracking offers and price drops for customers. Although we anticipate partnering with horizontal agents, we believe Alexa for Shopping will be the primary shopping intelligence that customers will use when shopping on Amazon.

As you browse, get inspired, and check product details, a shopping assistant is always around, and you can let it place orders directly based only on your previous purchase history — it already has your history, delivery address, and payment information — this is a unique and convenient experience for customers.

Next, look at Amazon ads. We achieved revenue of $19.6 billion in the third quarter, up 12% year over year. Despite some timing anomalies last quarter, we're still happy with the overall momentum of our entire advertising product portfolio. I would like to mention three factors affecting the growth rate. First, we moved Prime Day in most countries to the second quarter, and moved our impressive sales and advertising revenue from the third quarter of 2025 to the second quarter of 2026 — this was a smooth wind for our growth in the second quarter of 2026 and a headwind for the third quarter of 2026. As far as advertising is concerned, this had a month-on-month impact of 850 basis points on the year-on-year growth rate in the third quarter.

Second, the second quarter of 2026 includes the first season of NBA on Prime (not in the second quarter of last year), while the third quarter does not include any NBA (the NBA starts in the fourth quarter). Third, in the third quarter of 2025, we adjusted our search engine marketing programs to drive more traffic to search pages on Amazon. It's a better experience for our customers, and we're also monetizing this traffic at a higher rate and boosting ad revenue growth a year ago — we're starting to contrast that now. We expect the year-on-year growth rate in the fourth quarter to return to a high level of 10 (high teens, or 10 percent).

A natural question is whether smart purchases from external horizontal agents are significantly affecting Amazon advertising. The answer is: no. Currently, few customers do a lot of shopping this way because the experience isn't good enough, and it's not as convenient as going directly to a retailer to research or place an order with one click.

Our advertising team achieved great results in the third quarter and is expected to have a positive impact in the next few months, with the launch of Amazon Ads Agent — which integrates our complete advertising products under a single AI-first platform, supported by trillions of shopping, browsing, and streaming signals, to provide advertisers with insight, answers, and the ability to optimize and act on their behalf every step of the way.

We've also launched a full-funnel campaign in Amazon Ads Agent, combining search ads, display ads, video, and streaming TV into a single ad campaign. Advertisers provide initial input (including budget and delivery time), and Amazon's AI is responsible for planning, execution, and ongoing optimization of the rest.

We are also seeing continued growth in a number of other areas. I'll just mention a few. Prime Video has become an important business, running at tens of billions of dollars in revenue, and is already profitable itself. With hit series such as “Carrie,” “The Summer I Turned Pretty,” “The Boys,” “Off-Campus,” “Reacher,” and “Fallout,” and hit movies such as “Project Hail Mary,” plus engaging live sports (including Thursday Night, NBA on Prime, NASCAR, Masters, and Champions League soccer), we are gaining momentum.

We recently launched our most powerful device product line to date, spanning 34 new products spanning Echo, Kindle, Alexa Tablet, Fire TV, Ring, and Bee. It's still early, but we're off to a very strong start and look forward to more customers getting started with these products when the holiday season arrives. Finally, Amazon Leo is still on schedule to launch the initial service this year.

We are about to enter the most exciting period of the year — both the re:Invent conference and the holiday season. I want to thank our team for their hard work and innovation day in and day out to make our customers' lives better and more convenient. The business we've built at Amazon is unprecedented. There has never been a business with annual revenue of 1 trillion US dollars. Amazon is on the path to having two such businesses (AWS and retail stores). It's rare, and we're so inspired by everything that's to come.

Please Brian.

(4) Brian Olsavsky - Senior Vice President and Chief Financial Officer Financial Report

Thanks Andy. Let's first look at overall revenue performance. Global revenue was US$2019 billion, up 12% year over year; excluding exchange rate effects, up 13% year over year. The global operating profit was US$28.1 billion, and the operating margin was 13.9%, the highest operating margin in our history.

Let's take a look at the North American and international division results first. Let me start by reminding everyone that the timing of Prime Day has changed this year, and sales activities took place in the second quarter in most of our largest countries (including the US). Prime Day 2025 falls entirely in the third quarter. Excluding the impact of the Prime Day time change, the year-on-year growth rate for the third quarter of 2026 will be about 400 basis points higher. We saw strong customer engagement in the third quarter, as well as a continuation of the trend in the first half of the year. The North America segment's third-quarter revenue was US$112.9 billion, up 6% year over year. The International Division's revenue was US$43.1 billion, up 8% year over year after excluding exchange rate effects. The number of global payments increased 12% year over year.

Prime remains a key pillar of our business. Hundreds of millions of Prime members are responding to our combination of huge selection, low prices, and record delivery speeds. Over the past year, Prime members' monthly fresh food and grocery orders for free same-day delivery have more than tripled.

Next, let's look at profitability. The North America segment had operating profit of $8.7 billion, with an operating margin of 7.7%. The International Division's operating profit was US$1.6 billion, with an operating margin of 3.7%. In terms of the fulfillment network, the team dealt with rising transportation costs due to fuel inflation and higher mainline transportation rates. Despite these headwinds, we reduced overall unit service costs in the third quarter through better inventory layout, shorter shipping distances, and package consolidation. Transportation costs increased 11% year over year, while the number of items increased 12%. Excluding the impact of fuel and mainline inflation, transportation costs increased by 8%, which is about 400 basis points lower than the growth rate of the number of items.

Robotics and AI are delivering measurable benefits through productivity and employee safety, and we expect these benefits to continue to accumulate in the future. In 2026, we will double our holdings of robotic arms such as Cardinal and Sparrow. We are also developing basic models of robots to enable faster programming and teach robots increasingly complex physical tasks, with the goal of delivering useful robots at scale. Looking ahead, we see significant opportunities in our global fulfillment network to increase productivity, reduce unit service costs, and deliver products to customers faster. Our operating margins fluctuate from quarter to quarter, but we focus on continuing long-term improvements in unit service costs.

Next, let's take a look at the AWS division. Revenue was $46 billion, up 39.3% year over year, driven by a combination of core services and AI services. We added $3.8 billion in revenue month-on-month. AWS currently has an annual revenue operating rate of $184 billion. Customers continue to migrate to the cloud and expand their use of AWS core services. Also, more and more customers who want to reap the full benefits of AI are moving to the cloud at an accelerated pace.

We're seeing a strong correlation between AI spending and revenue growth from our core services. As customers invest in AI, we see a corresponding increase in core consumption, and we expect this relationship to further strengthen as more AI workloads enter production and drive new demand for computing, storage, and networking. AWS's operating profit was $17.8 billion, and the operating margin was 38.7%, reflecting strong growth and our focus on improving efficiency. Our investments in custom chips and custom networks, as well as software and process improvements, are helping us optimize capacity and build a more cost-effective infrastructure. As we have always said, we expect AWS operating margins to fluctuate, in part due to our level of investment at any point in time. Our AI products are being favored by customers, and AI revenue has increased by three digits year over year. Bedrock has been an important driver of growth.

Now let's look at our cash capital expenditure, which was $75.9 billion for the third quarter. This mainly relates to AWS and AI as we are investing to support strong customer demand. We continue to make large-scale investments, particularly in AI, because we believe this is a huge opportunity that promises long-term revenue and free cash flow and return on invested capital.

Finally, I concluded with financial guidance for the fourth quarter. Net sales for the fourth quarter are expected to be between $240 billion and $247 billion. Based on current exchange rates, we estimate that the year-on-year impact of exchange rate changes was about 60 basis points of headwind for the quarter. Operating profit for the fourth quarter is expected to be between $28.5 billion and $33.5 billion. I would like to mention one factor that influenced the fourth quarter guidance.

Amazon Leo's commercial services are scheduled to begin initial service this year, and we expect to start capitalizing some of the costs, including production and launch costs, in the first quarter of next year. This is later than my previous predictions in previous earnings conference calls. As a result, we will continue to fee-based most of the costs associated with Leo in the fourth quarter. We're ready to serve our customers this holiday season and support millions of customers building on AWS.

I'd like to thank the Amazon teams for their hard work and customer-centric focus. Our commitment to improving the customer experience remains the only reliable way to create long-term shareholder value.

Let's move on to the Q&A session.

Q & A session

(1) Q1 — Justin Post, Bank of America Securities Analyst

Q: Andy, I'm excited to see the cloud business continue to accelerate, which is great. However, the month-on-month growth rate has slowed. I want to know if this is due to factors such as customer side or token costs? And what do you think of the increase in production capacity in the fourth quarter? And Brian, what impact will fuel costs have on the fourth quarter guidance?

Andrew Jassy responded:

OK, I'll answer the growth question first. We're very happy with this growth — a business with an annualized revenue running rate of $184 billion achieved a 39.3% increase, which is quite rare. This is our fastest growth rate in the past five years, and it is also the sixth consecutive quarter of acceleration. We are also increasing our production capacity by a large amount. I think if you look at the current drivers of growth, it's a combination of several aspects. First of all, when it comes to AI, we have extraordinary full-stack capabilities — whether you're using SageMaker AI to build your own models (the easiest way to build your own models), or if you want to take advantage of leading-edge models, high security, and a very competitive price/performance ratio, you can use Bedrock.

Whether you're building your own intelligence (we provide capabilities like Strands and AgentCore to build and run securely and scalable) or use an out-of-the-box smart device (most companies' choice), just take a look at Kiro (coding), Connect (contact center), Transform (migration), and Quick (a rapidly growing AI work partner), people are adopting these very fast.

We are increasing our production capacity every quarter. The increase in production capacity is not the same every quarter. But if you look at our increased production capacity this year and our plans for the next 8 quarters (we can usually look forward to 8 to 10 quarters), we still have a lot of production capacity to come. On this basis, the influence of chips must also be added. We are in a very good position in the AI wave because of the growing fact that most customers — every customer that uses AI on a large scale — especially companies that are now beginning to deploy inference on a large scale in production, are extremely eager to save on price performance and training costs, and Trainium's significant advantages in terms of cost performance over other solutions are very appealing.

As more and more people use AI in a very wide range of ways, this has driven the need for a number of core capabilities, such as reinforcement learning, post-training, and tool call, and we have the leading CPU chip Graviton, which is about 30% to 40% more cost-effective than other x86 providers. These factors are really driving very significant growth. Our quarterly increase in production capacity will vary, but at this scale, predicting annual year-on-year growth of over 40% in 2027 and 2028 is significant, and we think this can become a $1 trillion revenue business within 5 to 10 years. This is an extraordinary and rare opportunity, and we are very excited to seize it.

Brian Olsavsky added:

Justin, on the question of fuel and mainline inflation, I don't have exact figures, but the fourth quarter will be higher than the third quarter. Remember we introduced fuel surcharges for FBA sellers at the beginning of the year to recover some (but not all) of the fuel costs we saw for these shipments and our remaining partners for retail inbound and outbound shipments. So it's included, which is slightly higher than in the third quarter.

(ii) Q2 — Morgan Stanley analyst Brian Nowak

QUESTION: I have two questions. Andy, your passion for AWS has been conveyed over the phone. You rarely make statements like “Achieve $1 trillion in revenue in the next 5 to 10 years,” which is beyond normal. Can you please explain how you reached this trillion dollars? Is it top down or bottom up? What makes you confident that you can achieve this kind of revenue growth and reach this number in the next 5 to 10 years?

Second question, with regard to robotics, which key milestones are you focusing on in terms of robotics in retail business looking ahead to 2027?

Andrew Jassy responded:

Regarding AWS, Brian, just look at the growth rate — look at the growth rate of the operating rate, and then look at our predictions for 2027 and 2028. As I've said many times, no matter how much production capacity we provide, the bulk capacity that customers acquire (hundreds of megawatts, sometimes up to 1 gigawatt) is reserved in advance because you can't leave 1 gigawatt idle. As a result, we have a very clear view of these bulk capacity requirements for 2027 and 2028 and part of 2029.

We always reserve a significant portion of capacity for on-demand use on EC2 or Bedrock, and work closely with customers to ensure they have sufficient capacity available in AWS. However, we have a very clear view of what AWS will look like in the next few years. When we calculate revenue and current operating rates, this growth rate will soon push up the numbers.

Building on that, I don't think much of the use we're seeing today will decay any time soon. The general deployment of reasoning by companies in production is still in the very early stages, really early. Part of the reason is that companies need enough talent that knows how to operate AI, conduct post-training, and fine-tune models for appropriate applications; partly because of inference costs. I think if businesses want to use reasoning as we believe, they must continue to work to reduce the cost of reasoning. This is a major AWS mission, and one of the key reasons we built our own custom chips and Trainium—we believe it's extremely important that customers can use inference in a more economical way. The cost of inference has dropped drastically in the past two years, and I think it will drop drastically in the future.

Then think about what you see. Take security as an example. This alone is now a very advanced model, and the vulnerabilities they have discovered have never been detected by humans. If you can use these models, you can use them to find bugs and ultimately help with patching and deployment. Some people think that a significant portion of the total amount of tokens consumed in the next year or two may end up being used for security. This is a relatively new use case, superimposed on the growth in coding and various AI partners (in consumer spaces and workspaces, such as Quick) we have seen. These are just the first few big use cases.

Add to that the corporate trends you'll see — I believe almost every modern enterprise's production application will be rewritten as intelligent, because if you want to act fast, or if you want to prevent some security risks in the right way, you have to deploy it this way. The wave of AI is just too big. So, you can't accurately predict the next 5 to 10 years, especially this kind of growth rate, but for us, it's not an exaggeration to imagine that this could be more than a trillion dollar business.

What I want to say about robots is that what's happening in the robotics world right now is pretty amazing — in general, and at Amazon in particular. We now have over 1 million robots in our fulfillment network. They have had a very significant impact on our cost structure, productivity, and security—look at our compliance network's security data, which is at an all-time low. We've made tremendous progress over the past four or five years, and robots are an important part of it.

But I'd say we're still in the early stages in terms of what we've built and deployed, and what we're developing. I think you'll see that our fulfillment centers will always have a large number of employees. But I think much of the more repetitive work we currently do by humans will be done by robots. We're investing in robots of all kinds. We also have the opportunity to help complete the “last 100 yards” delivery to customer stores. We have lots of results coming soon.

For us, a robot is measured by whether it provides equal or better productivity, safety, cost structure, and delivery times. When the economic mix is agreed upon, we will deploy it on a large scale in the fulfillment network. If you look at the upcoming new designs in our fulfillment network—including the expansion of Same-Day Delivery facilities—they're investing heavily in robotics. So we think robots have had a huge impact, and this is only the tip of the iceberg compared to the next five years.

(3) Q3 — Evercore ISI analyst Mark Mahaney

Q: Two things, first of all, thank you very much for the outlook for AWS growth. Could you also give me an outlook on profit margins? I know profit margins will fluctuate from quarter to quarter, but in the next two years, do you think the current profit margin structure is relatively stable? Will we maintain an operating margin level of more than 30% or 40% for AWS?

Also, regarding fresh checkout, I'm super excited to see this opportunity come to fruition in my community and other regions. How much fresh checkout coverage do you have? In other words how many Prime families do you guys think it covers? For customers who have used it, how much do you think this is a big increase in spending per Prime household? Is it a 20%, 50% or 100% increase? Considering that the fresh food category is so big?

Brian Olsavsky responded:

Mark, let me first talk about the AWS operating margin outlook. We do not provide guidance on AWS profit margins. But I can give you my opinion. The operating margin for the quarter was 38.7%, which looked very strong. I'd like to add that the profitability you've seen recently is not a random result, but rather a product of significant disciplinary efficiency improvements, capacity optimization, and tight control of fixed costs. As Andy said, over time, more and more custom chips have entered our cost structure. However, there are other factors at play. In fact, the unconventional growth of AI will put pressure on overall business margins because profit margins are lower there.

We are pleased with the progress of both core and AI profit margins. As Andy said earlier, they are ahead of the core business's performance at a similar stage, but are still lower. So there's this kind of structure (mix) problem. There's also the headwinds we've talked about with component pricing. So it's impossible to predict, but there will always be multiple factors intertwined, and we work very hard on the cost side and efficiency side while handling the profit margins of new products — as they launch, launch, and scale.

Andrew Jassy added:

I'm as excited as you are about fresh checkout, Mark. As you know, we've been in the big grocery business for a while now. It mainly focuses on products on the middle shelf of a grocery store — daily necessities, canned food, pharmaceuticals, beauty products. We have the leading organic grocery store Whole Foods Market, which is also growing faster than the rest of the grocery business. In terms of revenue and profit, the new business format and daily shopping performance in the urban environment is excellent, and we are planning to promote it over the next few years.

We've always known that if we want to serve our customers the way we think they want and we want, we have to provide popular products, and it all starts with fresh food. If you look at the total expenses of each household, when we compare the total amount of products we think consumers and customers can spend with us and what they actually spend with us, the biggest gap is groceries. And a lot of shopping and a lot of groceries start with fresh food. So, we're really excited because we think we've found the right direction when it comes to delivering fresh food from Same-Day Delivery facilities.

Our daily necessities are growing at about 3 times the rate of the rest of the store, and currently account for about half of our total number of items. We found that customers who bought fresh food with Same-Day Delivery bought about 7 times more items of daily necessities than those who didn't. So if you're trying to close the gap between total consumer spending and what we can provide, fresh food has a huge impact on this gap.

We've also seen that our grocery customers buy non-grocery categories such as home and apparel about 4 times more, and spend about 2 times more on these categories, compared to customers who don't buy groceries from us. These are all huge differences, and we believe that progress in fresh food and groceries will ultimately make us more relevant to customers in more shopping scenarios.

(4) Q4 - Goldman Sachs analyst Eric Sheridan

Q: Ask two questions about AWS market opportunities. First of all, about Bedrock, what are your key perceptions of Bedrock's market opportunities? How will it continue to expand over the next few years? Second, with regard to the decision to sell Trainium externally, how do you view the balance between the market opportunity for external sales and the computing power required for this opportunity, as well as the balance between it and internal use cases and other AWS opportunities?

Andrew Jassy responded:

You've probably heard us talk about Bedrock more than once in the past two years. But over 20 years of experience providing technology to customers has taught me that no one tool can rule the world. People always thought that the early leader would be the only thing used, but that was never the case. This is true of databases, and so is the field of analytics. In the early days of AI, everyone said that the only framework was TensorFlow, but now PyTorch is much larger.

The same goes for models. In the next 2 to 3 years, I think there will be 6 to 8 models at the same level as each other, or at least enough to cover more than 95% of the workload. Some are better at reasoning, some are better at coding, some are better at visuals, some have lower latency, and some are cheaper. Most customers use multiple models, and many use multiple models in the same generative AI application.

So what customers really want — we've proven time and again in our conversations with customers — is a service with a wide selection of these cutting-edge models, good value for money, strong security and access control. They wanted this service to decide what combination of models to use and switch back and forth as the models exceeded each other and applied constraints changed. This is exactly what we saw on Bedrock. People are so excited about it.

We have both closed source models such as Anthropic and OpenAI, as well as open source models such as Kimi and GLM. I think the Western world is very hungry for more Western open source models, and I'm sure this will happen over time as well. But I think Bedrock's opportunity is huge; it could become a service as big as EC2, and EC2 is a very huge service in AWS today. So I think it's pretty huge.

Regarding the balance of chips — how much is sold externally and how much is used internally in the cloud, this is a great question, and one we are continuously evaluating. I would say that there are actually only 2 companies that have significant cloud businesses and enough to sell chips abroad at the same time. So there aren't many customers who really need to face this kind of problem on this scale. Currently, everyone is using the same foundry to produce their advanced chips. The foundry can only supply that many chips each year, and everyone gets a quota. You have to decide how much of your quota to sell externally and how much to spend on your cloud.

There's a trade-off here. If you sell abroad, you can achieve more revenue in the short term, because you can confirm the revenue as soon as the chip leaves the warehouse. Conversely, if you decide to use it in the cloud in the form of an EC2 instance or Bedrock, you'll actually drive more absolute revenue and free cash flow from these chips because you can continue to use them for 5 to 6 years and generate revenue over a longer period of time. So, you really have to figure out how to do it. So far, I'd say we've taken a different approach to chip use than another provider facing the same balance issues.

As I mentioned earlier, we believe that selling chips abroad will be a very lucrative business, and we will launch at the end of 2027. We probably rethink this balance every year. But both are going to be very important. Our ability to build cost-effective, expandable chips will be our major strategic advantage in terms of scale — whether in our own cost structure or as an independent business selling chips overseas — selling chips overseas is indeed a completely different business from the cloud infrastructure business.

(5) Q5 — Douglas Anmuth, J.P. Morgan Analyst

Q: When you consider AWS's revenue growth of more than 40% in the next two years and the multi-year path thereafter, can you talk about the capital expenses needed to drive this revenue growth in the next few years and beyond? Also, Andy, do you have any updated ideas about the payback cycle of servers, networks, and data centers?

Andrew Jassy responded:

OK, I answered in reverse order. Regarding infrastructure break-even, as I mentioned in my opening remarks, we're continuing to move it to the left — making it shorter. The core business is slightly different from AI, and the funding method is different, depending on what part you're talking about, but we're continuing to make progress. As we've always been good at AWS history, we're constantly looking for ways to extend our lifespan and speed up our break-even, and we're progressing well in this regard.

As for the capital expenses required to grow the way we say, this is pretty impressive. I think, Doug, our take on this matter is that this is a once in a lifetime opportunity. The AI we're seeing is the biggest transformation of our lives. If we can achieve 39% growth with an annual operating rate of 184 billion US dollars, and increase by 40% every year over the next two years, and eventually become a business of more than 1 trillion US dollars, we will invest capital to continue leading the way. Today, we are leaders in infrastructure and AI. If this acceleration and opportunity were spread over 25 years, would it be more convenient if you could spread capital expenditure over 25 years? Yes, it might be more convenient, but the world we live in isn't like that, and so are opportunities.

We are able to fund this program. I think it's going to change the size of Amazon forever, forever change our revenue, forever change our free cash flow. We will invest and persevere in our pursuit to continue to be the important leaders we are today.

(6) Q6 - Barclays analyst Ross Sandler

QUESTION: OK, I'd like to dive a little deeper into Bedrock's question. The token trading market on Bedrock seems to be one of the hottest services in the AI field right now. As far as we know, most of these workloads run on Trainium. As it becomes a more important part of the overall story of AWS — as you mentioned, Andy — how will this impact revenue growth, particularly AWS's profit margins, as you shift from selling bare metal to more selling reasoning as a service? Can we talk about it?

Andrew Jassy responded:

Yes. You're right, most of the content on Bedrock currently runs on Trainium. One important reason we do this is that we love the value for money we get, which not only keeps our prices as low as possible for our customers, but also allows us to have the associated profit margin. If you think about profit margins in the cloud sector and look at the cost structures of the different players, there are a few key points worth paying attention to. The first is data centers, and how many of them are purchased or long-term leased rather than using colo (hosting) or neocloud (new cloud). Neocloud's operating profit margin is about 3,500 basis points lower than long-term leases.

Next, look at the hardware they run on. If you're using hardware that everyone uses, your cost structure will be similar to everyone else's. If you have the ability to use different hardware, you have an advantage. So I think, look at what we do — most of our capacity is owned or leased for a long time, which has a better operating margin than mainly using colo or neocloud, or a significant portion of it. Moreover, we run on our own chip, which not only provides an advantage in terms of cost performance, but is also more cost-effective for ourselves because we own it — I think this gives us some structural cost advantages, and when we return the savings to our customers, it will be very beneficial to the customer, and it will also be very beneficial to the overall economy of AWS.

(7) Dave Fildes

Thank you all for attending today's call and asking questions. The replay will be kept on our investor relations website for at least 3 months. Thank you for your interest in Amazon and we look forward to speaking with you again next quarter.