①On Amazon During the third-quarter earnings call, the company's CEO announced that capital expenditures are expected to increase further in 2026, with the vast majority of the investment going towards data centers needed for AI. ① Electricity and chips; ② Amazon's CEO also discussed Amazon AWS infrastructure capacity, chip strategy, retail fresh food business, employee size, and robotics. The article addressed analysts' questions on a number of core topics, including applications.
Amazon released its Q3 2025 earnings report and held a conference call after the market closed on Thursday, Eastern Time.
Financial reports show that Amazon's cloud business AWS revenue growth has returned to 20% year-on-year, reaching a new high in nearly three years, with annualized revenue climbing to $132 billion.
During the earnings call, Amazon CEO Andy Garthy announced that the company expects its capital expenditures to reach approximately $125 billion in 2025, and will increase further in 2026, with the vast majority of the investment going toward data centers , power, and chips needed for AI.
At the same time, he confirmed that Amazon would be laying off up to tens of thousands of employees, saying it was due to corporate culture considerations.
In addition, Jassy responded to analysts' questions on a number of core topics, including Amazon AWS infrastructure capacity, chip strategy, retail fresh food business, employee size, robot applications, and intelligent commerce, and disclosed details of several business progresses and future plans.
Four Key Points from Amazon's Earnings Call
I. AWS: Capacity and Chips as the Dual Drivers, AI as the Core of Growth
Infrastructure capacity: Over 3.8 gigawatts of new data center capacity were added in the past 12 months, with another 1 gigawatt planned for Q4. Current capacity is double that of 2022, and plans are in place to double it again by 2027. New capacity is being built and used simultaneously, with monetization proceeding at the same pace . The current bottleneck in the industry is electricity, but the future may shift towards chips, and AWS is proactively securing supply.
Trainium chip: The Trainium2 business has reached a scale of billions of dollars, with quarterly revenue increasing by 150% and fully booked, serving only ultra-large customers such as Anthropic; Trainium3 will be previewed at the end of the year and mass-produced in early 2026, with a performance improvement of 40%, and will cover more medium-sized customers. Its core advantage is that its cost performance is 30%-40% higher than similar products.
Strategy and Growth: Adhering to the "self-developed + third-party" chip development path, and collaborating with Nvidia. Cooperation deepened and procurement volume continued to increase; AWS's 20.2% growth came from AI business (inference, training, etc.) and enterprise cloud adoption, with outstanding orders of $200 billion, and no new transactions announced in October exceeding the entire Q3 quarter. The AgentCore tool has no similar competitors, and its SDK has been downloaded over 1 million times, helping enterprises deploy intelligent agents.
II. Retail: Breakthrough in same-day delivery of fresh produce, with parallel development of physical retail formats
Fresh produce business: Excluding Whole Foods and Fresh, the annualized transaction volume of the grocery business exceeds $100 billion, ranking among the top three in the United States. The "fresh produce + general merchandise" same-day delivery service covers 1,000 U.S. towns and cities, with a customer repurchase rate twice that of non-fresh produce shoppers. By the end of the year, it will expand to 2,300 towns and cities, reshaping the "weekly stock-up" habit.
Physical stores: Whole Foods is not giving up on exploring physical stores. Its growth rate is faster than the industry average and its profitability is improving. It will expand its stores. The small-format "Daily Shop" has been launched in 3 stores, with a good start. It will be replicated at an accelerated pace.
III. Operations: Improving Efficiency through Flat Structure and Automation
Staffing adjustments: Not driven by finance or AI, but by the fact that previous expansion led to multiple layers and reduced efficiency. The adjustment aims to "reduce layers and increase ownership," returning to a flat organizational structure like a startup, and maintaining a lean approach in the long term.
Robot Applications: The fulfillment network has deployed over 1 million robots to improve safety and efficiency. In the future, it will focus on "human-machine collaboration" rather than pure replacement. It will combine algorithm optimization to reduce costs and support faster delivery and expansion of fresh produce.
IV. Emerging Fields: Unleashing the Potential of Intelligent Agents and Advertising
Intelligent Agent Commerce: Our proprietary tool Rufus (250 million annual active users, 60% increase in conversion rate) and the "buying on behalf" function are proving effective; we recognize the potential of third-party intelligent agents, but acknowledge current pain points in user experience. We will explore cooperation while ensuring user experience and value exchange, with the long-term goal of making the online shopping experience surpass that of offline shopping.
Advertising business: DSP, Prime Video sports ads, and core ads are all performing well. DSP has been fully functionalized over 20 months, and partnerships with Roku, Netflix, and others are unlocking incremental growth; leveraging the advantages of the "full funnel" and AI tools (such as Creative Studio to shorten ad cycles) improves advertiser ROI.
The following is a full transcript of Amazon's earnings call (This article is generated by artificial intelligence) . (Auxiliary translation)
Meeting moderator
Thank you for your patience. Hello everyone, welcome to Amazon's Q3 2025 Financial Results Conference Call. Currently, all participants are in listen-only mode. After the presentation, we will proceed to the Q&A session. This conference call is currently being recorded. For the opening remarks, I will hand over the chairing to Mr. Dave Fildes, Vice President of Investor Relations. Thank you, sir. Please begin.
Dave Fildes – Vice President of Investor Relations
Hello everyone, welcome to our Q3 2025 financial results conference call. Joining me today to answer your questions are our CEO, Andy Jassy, and our CFO, Brian Olsavsky .
Andy Jassy – CEO
Thank you, Dave. We achieved strong growth across all our business segments in the third quarter. Excluding the impact of currency exchange rates, revenue for the quarter was $180.2 billion, up 12% year-over-year. Operating profit was $17.4 billion, but excluding two special charges in the third quarter (a $2.5 billion settlement with the Federal Trade Commission (FTC) and estimated severance pay of $1.8 billion), operating profit would have exceeded $21 billion. Free cash flow over the past 12 months was $14.8 billion.
First, let me talk about Amazon Web Services (AWS). AWS achieved its fastest growth rate since 2022, with year-over-year growth rebounding to 20.2%, our highest growth rate in 11 quarters. It's important to note that year-over-year growth is a relative concept. Achieving 20% year-over-year growth on a $132 billion annualized revenue base is fundamentally different from achieving a higher percentage of growth on a much smaller annualized revenue base, the latter of which is where our competitors are.
As of the end of the third quarter, AWS backlog reached $200 billion, and this figure does not include several undisclosed new deals in October—the total value of which already exceeds our total deal volume for the entire third quarter. AWS's momentum continues to grow. With greater capabilities, enhanced security, and superior operational performance, customers are looking to deploy their core workloads and artificial intelligence (AI) workloads on AWS. The potential I see in the current scale of our business gives me confidence in the future. I will now explain the reasons behind this in detail. First, AWS offers a broader range of infrastructure capabilities. Startups, large enterprises, and government agencies are looking to migrate their production workloads to the most comprehensive and deeply integrated platform. AWS offers more services than any other cloud service provider, with richer functionality within each service, and we continue to innovate at a rapid pace.
These services are a critical foundation for customers' various business operations and a key reason why Gartner has ranked AWS as a leader in its Magic Quadrant for strategic cloud platform services for 15 consecutive years. In the field of artificial intelligence , we adopt the same "building block" approach. Amazon SageMaker significantly reduces the difficulty for enterprises to build and deploy their own foundational models. Amazon Bedrock provides customers with a rich selection of foundational models while offering excellent cost-effectiveness in deploying inference to next-generation applications. In the future, a large portion of the value enterprises derive from artificial intelligence will be in the form of agents. AWS is investing heavily in this area and is well-positioned to become an industry leader. Enterprises can either create their own agents or use agents developed by other companies. For those enterprises wishing to create their own agents, the current development process remains overly complex.
It is for this reason that we introduced Strands—a tool that makes it easier for developers to create agents based on any base model they prefer. However, for companies that have already successfully developed agents, they often hesitate to deploy them in production because they lack secure, scalable runtime services specifically designed for agents, as well as corresponding memory support or observability tools. Therefore, we introduced AgentCore—an infrastructure building block that helps developers deploy secure, scalable agents. Ericsson Ericsson provides AI-powered agent services to all its employees using AgentCore; Sony Sony has built an agent-based AI platform with enterprise-grade security, observability, and scalability using AgentCore; Cohere Health has deployed agents using AgentCore, reducing medical review time by 30%-40%. To date, AgentCore's software development... The SDK has been downloaded over 1 million times and has received enthusiastic responses from developers, providing strong support for the application and promotion of intelligent agents.
Enterprises also use agents developed by other parties, and AWS continues to develop various agents that we believe developers will use in the future. In the code development field, we recently released a preview of Kiro, an integrated development environment (IDE) for agent-based code. In just the first few days of the preview release, over 100,000 developers joined, and that number has since more than doubled. To date, Kiro has processed trillions of tokens, with weekly active users growing rapidly, and developers highly appreciating its unique specification definition and tool invocation capabilities.
In the field of migration and transformation, we launched an agent called Transform. This year alone, customers have saved 700,000 hours of human labor using this agent, equivalent to the workload of 335 developers for an entire year. For example, Thomson Reuters... (Thomson Reuters) Transform completes 1.5 million lines of code conversion per month, migrating systems from Windows to open-source alternatives, at a speed four times faster than other migration tools.
Customers have also leveraged Transform to analyze nearly 1 billion lines of mainframe code, facilitating the migration of mainframe applications to the cloud. In the business customer service space, we recently launched Quick Suite—a suite that brings a consumer-grade AI-like experience to the workplace, helping users easily gain insights, conduct in-depth research, automate tasks, visualize data, and take action. Currently, users have used Quick Suite to reduce projects that previously took months to just a few days, saving over 80% of time on complex tasks and achieving over 90% cost savings. In the call center space, we launched Amazon Connect, creating a more personalized and efficient experience for call center agents, managers, and their customers. Recently, Amazon Connect's annualized revenue surpassed $1 billion; in the past year, AI processed a cumulative 12 billion minutes of customer interactions; including Capital One Financial… Capital One, Toyota, American Airlines Several large companies, including American Airlines and Ryanair, are using this solution.
These are tangible results that have brought to customers, and there are many similar examples. With its superior functionality, security, operational performance, and customer-centric service philosophy, AWS continues to win numerous partnerships in cloud transformation projects for large enterprises and government agencies. This is why the vast majority of enterprise data and workloads are deployed on AWS, which is also one of the reasons why many enterprises choose to run AI workloads on AWS.
To support this customer demand, we need sufficient infrastructure capacity. Over the past few months, we have focused on increasing capacity, adding more than 3.8 gigawatts (GW) of computing power in the last 12 months—more than any other cloud service provider. Specifically, AWS's computing power is now double that of 2022, and we plan to double that capacity again by 2027.
In the last quarter of this year alone, we expect to add at least another 1 gigawatt of capacity. This capacity includes power, data center, and chip resources, primarily our own Trainium chips and NVIDIA chips. Recently, we launched Project Rainier—a large-scale AI computing cluster spanning multiple data centers across the United States, equipped with nearly 500,000 Trainium 2 chips. Anthropic is currently using this cluster to develop and deploy its industry-leading AI model, Claude, and we expect the number of Trainium 2 chips running Claude to exceed 1 million by the end of the year. Trainium 2 adoption continues to climb and is currently fully booked, with the business reaching billions of dollars and quarterly revenue growth of 150%. Currently, Trainium chips serve only a few very large customers, but we expect to serve more customers starting with Trainium 3.
We are committed to making Bedrock the world's largest inference engine, and in the long term, we believe Bedrock has the potential to become a core AWS business comparable in scale to Amazon Elastic Cloud Computing (EC2). Currently, the vast majority of token usage in Amazon Bedrock is based on Trainium chips. At the same time, we continue to work closely with chip partners such as NVIDIA —we are still sourcing chips in large quantities from NVIDIA; in addition, we are working with AMD. (AMD) and Intel Our collaboration with Intel is also progressing. These are all very important partners, and we look forward to further deepening our partnerships with them in the future. Given the current strong market demand, we will continue to actively invest in expanding our infrastructure capacity. Currently, our rate of capacity addition is roughly in line with the rate of capacity monetization. The development of artificial intelligence is still in its early stages, which presents unique opportunities for AWS and its customers.
Next, I'll talk about the Stores business. The Stores team continues to create value and innovate for customers, focusing on core growth priorities—product selection, low-price strategy, and convenience (especially fast delivery). Compared to the previous quarter, we added 14% more product selection , including items from popular brands such as The North Face and Charlotte Tilbury; year-to-date, we've added hundreds of thousands of popular brand items. The Everyday Essentials category is growing rapidly, almost double the overall business growth rate this year. We are continuously optimizing our services to make it easier for customers to buy low-priced perishable foods on Amazon; currently, customers in over 1,000 towns can purchase millions of Amazon items while also ordering fresh food and enjoying free same-day delivery. This service is revolutionary for customers—they can now order milk while buying electronics, using a bank card... The card payment is completed, and all goods are received within hours.
The retail team also innovatively launched the "Add-to-Delivery" button, allowing customers to add items to pre-ordered deliveries. Since its launch (a relatively recent addition), this button has been used over 80 million times. This seemingly simple yet highly impactful innovation has genuinely improved customer convenience. We remain committed to maintaining competitive pricing, ensuring prices are at or below those of other major retailers. This July, we hosted the largest Prime Day in our history, saving customers billions of dollars across more than 35 categories. Regarding delivery speed, we continue to break records. This year, we are on track to achieve the fastest Prime delivery speed in history globally again; we have also begun rolling out 3-hour delivery in select U.S. cities. Furthermore, we continue to invest in infrastructure to accelerate delivery in rural areas, serving customers in more communities.
This includes investing over $4 billion to expand delivery networks in rural areas of the United States. Residents in many rural towns also need fast delivery services, but other companies are withdrawing from these areas and scaling back their services. In contrast, we have increased the number of rural communities with same-day and next-day delivery services by 60%, and by the end of the year, we plan to cover approximately half of our target number of rural communities.
Our retail team is also rapidly innovating in the field of artificial intelligence. For example, our AI shopping assistant, Rufus, has reached 250 million active users this year; monthly active users have increased by 140% year-over-year, and interactions have increased by 210% year-over-year; customers who use Rufus during the shopping process are 60% more likely to complete a purchase than those who don't . Rufus is expected to bring us over $10 billion in annualized incremental sales. These are the key highlights of our retail business.
Our generative AI-powered audio feature integrates product summaries and reviews to enhance shopping. This feature has expanded from covering hundreds of products at launch to millions today; millions of customers use it, accumulating nearly 3 million minutes of playback time. Amazon Lens is an AI-powered visual search tool that allows customers to find products using their phone's camera, screenshots, or barcodes. The tool now includes Lens Live, which instantly scans products and displays real-time matching results in a scrollable carousel. Tens of millions of customers use Amazon Lens every month. Next, let's talk about Amazon Ads. We are pleased with the continued strong growth in advertising, with revenue reaching $17.6 billion this quarter, a 22% year-over-year increase. Our full-funnel advertising portfolio performed strongly, helping advertisers reach an average of over 300 million ad-supported audiences in the U.S. alone.
We are also very excited about the development of Amazon DSP, the demand-side platform that helps advertisers plan, execute, and measure the effectiveness of their full-funnel advertising investments. Last quarter, I mentioned our partnership with Roku; building on that, we've now partnered with Netflix—through this partnership, advertisers using Amazon DSP will have direct access to Netflix's premium ad inventory. In addition, we announced integration plans with Spotify and SiriusXM. With Spotify, advertisers will be able to directly reach over 400 million monthly ad-supported listeners globally programmatically; with SiriusXM, brands can reach 160 million monthly digital listeners through services like Pandora and SoundCloud. We are also excited about the huge potential of advertising related to Prime Video live sports events. In our 2025-2026 advertising pre-sale negotiations, live sports events attracted significant advertiser attention, with pre-sale commitments exceeding our expectations and showing significant growth across all sectors. Finally, we are also leveraging artificial intelligence to provide innovative services for advertisers.
For example, this September, we launched an AI-powered tool in Creative Studio that can complete the entire advertising creative process in hours, whereas traditional processes typically take weeks. Furthermore, we are actively innovating and making good progress in several other areas, to name just a few. Regarding Prime Video's live sports broadcasts, the NBA Prime broadcasts kicked off last week, with the first two games averaging 1.25 million viewers in the US, a double-digit increase over last season's cable television viewership. We will apply the same innovative approach we use in NFL broadcasts to NBA broadcasts. In 2026, we will add Masters Tournament golf broadcasts; this Black Friday, we will also partner with the PGA Tour to launch a new skins competition. In the US market, Prime Video has over 100 add-on subscription channels, and we recently added Peacock and FOX One. The market response to Alexa+ (the upgraded version of the smart voice assistant) is very encouraging.
Compared to the classic Alexa, Alexa+ doubles the frequency of user interactions with Alexa, significantly extends the interaction duration, and covers a wider range of topics. On Fire TV (Amazon's TV box), Alexa+ usage is 2.5 times higher than the classic version; users discover audio content through natural conversation three times more often, interact with the photo function three times more frequently, and complete shopping conversations and place orders three times more often. We have increased the number of Project Kuiper satellites in orbit to over 150; through enterprise-grade client terminals, we have achieved over 1 gigabit in testing. With a transmission speed of Gbps, this is the first commercially available phased array terminal we know of to reach this threshold. Finally, Zoox Robotaxi is now available to passengers in Las Vegas, and we've announced Washington, D.C., as our eighth testing location. We look forward to expanding these services to more areas and serving even more passengers.
The fourth quarter is our busiest and most exciting time of the year. Demand for AWS remains strong, we're excited about the innovations we'll be announcing at re:Invent in December, customers are responding positively to our AI-driven experiences, we'll be shipping a large volume of goods during the holiday season, and there are many other exciting developments to look forward to. I'd like to thank our employees around the world in advance—they're working tirelessly to once again provide excellent service to our customers. Now, I'll hand over the meeting to Brian, who will provide an update on the financials.
Brian Olsavsky – Chief Financial Officer
Thank you, Andy. First, let's look at the overall financial results: Global revenue was $180.2 billion, representing a 12% year-over-year increase after excluding a 90 basis point positive currency effect. In the third quarter, we reported global operating profit of $17.4 billion. This operating profit included two specific expenses that resulted in a combined decrease of $4.3 billion in operating profit.
The first expense, $2.5 billion, relates to a legal settlement with the Federal Trade Commission (FTC). This expense affects the North American business segment and is recorded under "Other Operating Expenses." The second expense, $1.8 billion, relates to severance costs related to layoffs, affecting all three of our business segments. Severance costs are primarily recorded under "Technology and Infrastructure Expenses," "Sales and Marketing Expenses," and "General and Administrative Expenses." Excluding these two expenses, global operating profit would be $21.7 billion, $1.2 billion higher than the upper end of our guidance range. Moving on to the performance of each business segment, we are encouraged by the innovation delivered to our customers by our teams across all business segments. North American revenue for the third quarter was $106.3 billion, up 11% year-over-year. Excluding the impact of currency exchange rates, International revenue was $40.9 billion, up 10% year-over-year.
Global Paid Units grew 11% year-over-year. We consistently prioritize what matters most to our customers. In the third quarter, our competitive pricing, extensive product selection, and rapid delivery services continued to be recognized by our clients. They appreciate our ability to quickly deliver everyday essentials, including perishable goods, and offer same-day delivery. Millions of third-party sellers worldwide are a vital support for our extensive product selection, helping clients find the products they need at competitive prices. We are committed to creating innovative services and features for sellers, including our ongoing efforts in generative AI. Currently, over 1.3 million sellers are using our generative AI capabilities to create high-quality product listings (product detail pages) more quickly. Better product detail pages lead to better sales results – in the third quarter, third-party sellers accounted for 62% of global product sales, an increase of 200 basis points compared to the same period last year.
Let's look at profitability. The North American segment generated $4.8 billion in operating profit, with an operating margin of 4.5%. Excluding the $2.5 billion in expenses related to the Federal Trade Commission legal settlement, the North American segment's operating profit would have been $7.3 billion, with an operating margin of 6.9%. The North American segment's operating profit margin also includes some severance costs. The International segment generated $1.2 billion in operating profit, with an operating margin of 2.9%. Excluding severance costs, the International segment's operating profit margin improved year-over-year. Globally, our progress in core business elements has not only delivered a better customer experience but also driven a more efficient cost structure. For example, we have made significant progress in optimizing inventory distribution to accelerate customer delivery. 2025 will mark the third consecutive year that we have achieved the fastest global Prime member delivery speed in history.
We continue to optimize and improve our fulfillment operations, and our regional fulfillment network has achieved scale. Improvements to the inbound process have brought us numerous benefits; for example, lead time in the US has been reduced by nearly four days compared to last year. This improvement allows us to conduct inventory procurement more efficiently, positively impacting working capital management. Simultaneously, we have also more rationally distributed inventory across our fulfillment network. By fully leveraging our existing infrastructure, we can now offer perishable food ordering services to US customers, achieving same-day delivery in as little as five hours. Early feedback on this service has been positive. Since its launch in January of this year, customers making their first purchase of fresh food on the Amazon platform have visited the website more frequently, and the repeat customer rate is twice that of customers purchasing non-perishable goods. Looking ahead, we believe there is still significant room for improvement in the efficiency of our global fulfillment and transportation network.
We will continue to optimize our inventory layout, reducing package delivery distances and handling frequency. Simultaneously, we will further solidify the advantages of our regionalized network through algorithmic improvements and the introduction of robotics and automation. While operating margins may fluctuate quarter by quarter, we have a clear strategy committed to long-term, sustainable improvement. Next, let's talk about advertising. Advertising revenue reached $17.7 billion, marking the third consecutive quarter of accelerated growth. Our full-funnel advertising model helps brands connect effectively with customers, driving strong growth even as the revenue base expands. Moving on to Amazon Web Services (AWS): AWS revenue reached $33 billion, a 20.2% year-over-year increase. This represents a 270 basis point improvement compared to the previous quarter, driven by strong growth in AI services and core services, as well as increased infrastructure capacity to meet customer demand.
AWS revenue increased by $2.1 billion quarter-over-quarter, bringing its annualized revenue to $132 billion. AWS operating profit was $11.4 billion, reflecting both continued business growth and our efforts to improve operational efficiency. We are significantly expanding our data... The scale of the centers is primarily to meet the needs of generative artificial intelligence (Gen AI); the depreciation expenses associated with these new assets will have some impact on profit margins. As we have stated for a long time, AWS operating margins may fluctuate over time, partly due to differences in the scale of our investments at different times. Next, let's look at capital expenditures (CapEx) cash outflows: third-quarter capital expenditures totaled $34.2 billion, bringing the year-to-date total to $89.9 billion.
These expenditures are primarily for AWS operations—investing in custom chips such as Trainium to support AI and core service needs, as well as building the technology infrastructure to support our North American and international operations. We will continue to make significant investments, particularly in AI, as we see it as a tremendous opportunity with the potential for substantial long-term returns on our capital investments. In addition, we continue to invest in our fulfillment and transportation networks to support business growth, improve delivery speed, and reduce service costs. These investments will support our business growth for many years to come. Looking ahead, we expect capital expenditures of approximately $125 billion in cash for the full year 2025, with this figure expected to increase further in 2026. Finally, I'd like to address our net profit situation.
While our discussion has primarily focused on operating profit, it's important to note that the $21.2 billion in net income for the third quarter included $9.5 billion in pre-tax gains related to the Anthropic investment. This investment is unrelated to Amazon's continuing operations and is recorded as "non-operating income." With the peak season off to a good start, we are well-prepared to provide excellent service to our customers in the coming months. I want to thank all Amazon employees for their hard work and dedication in preparing for a delightful holiday season. Improving customer experience is the only reliable way to create sustainable value for our shareholders. Now, let's move on to the Q&A session.
Meeting moderator
Now, we will open the meeting to questions. Please ask only one question per person. The first question is from Bank of America. Justin Post of Bank of America. Please ask a question.
Justin Post
Okay, thank you. I'd like to ask a question about AWS. Could you talk about your current view on AWS capacity levels and the current level of capacity strain? In your prepared remarks, you mentioned the demand for Trainium 3 and plans to potentially expand the customer base. Could you talk about the demand for Trainium from customers other than large clients? Thank you.
Andy Jassy – CEO
Regarding capacity, as I mentioned in my opening remarks, we have added a significant amount of capacity—3.8 gigawatts in the past year, with over 1 gigawatt to be added in the fourth quarter. We plan to double our total capacity again by the end of 2027. Currently, we are continuously expanding our capacity. From an industry-wide perspective, the current bottleneck may be power supply; I believe that at some point in the future, the bottleneck may shift to chip supply, but we are still continuously adding capacity. Moreover, our current rate of capacity addition is roughly in line with the rate of capacity monetization. Regarding the demand for Trainium among non-large customer groups: First, as I mentioned, Trainium2 is currently performing exceptionally well, fully booked, with a business scale reaching billions of dollars and quarterly revenue growth of 150%.
Several large-scale projects are currently underway, such as Project Rainier, a collaboration with Anthropic. Anthropic is developing its next-generation Claude model based on 500,000 Trainium 2 chips, and by the end of the year, the number of Trainium 2 chips running this model will reach 1 million. As I mentioned before, Trainium 2 currently only serves a few very large customers. However, because Trainium chips offer 30%-40% better performance than other similar products, and as customers begin to consider scaling up production workloads, shifting to AI-centric business models, and using inference capabilities, their focus on performance will significantly increase. Therefore, market demand for Trainium is very strong. Trainium 3 is expected to release a preview version by the end of this year, with larger-scale mass production expected in early 2026.
Currently, numerous customers (including very large and medium-sized customers) have expressed strong interest in Trainium3.
Meeting moderator
The next question comes from Morgan Stanley. Brian Nowak of Morgan Stanley.
Brian Nowak – Managing Director
Thank you for taking my question. Congratulations on your excellent performance this quarter. I have two questions for Andy. The first is a macro-level question about chip strategy. There are currently many questions in the market regarding Trainium and its positioning compared to other third-party chips. In your opinion, what key obstacles does Trainium3 need to overcome to achieve wider adoption of Trainium (as you mentioned in the previous question)? Also, in the short term, how do you balance the supply of Trainium with that of third-party chips in response to broader market demand?
Andy Jassy – CEO
Okay. First of all, we will always provide our customers with a variety of chip options.
在AWS 的发展历程中，每一个重要的技术构建模块或组件，我们都坚持这一原则。回顾AWS 的历史，在任何一个市场领域，长期来看都不会只有一家供应商能满足所有客户在所有维度上的需求。我们与英伟达（NVIDIA）有着深厚的合作关系，这种关系已持续很长时间，而且在我可预见的未来，这种合作仍将继续。我们采购了大量英伟达芯片，在采购英伟达芯片方面不存在任何限制。我预计，无论是明年还是未来，我们采购英伟达芯片的数量都将继续增加。
与大多数科技公司不同，我们拥有自己实力雄厚的芯片研发团队—— 也就是我们的Annapurna 团队。大家最初在中央处理器（CPU）领域看到了他们的成果—— 我们开发的Graviton 芯片，其性价比比其他x86 处理器高出约40%。如今，在人工智能定制芯片领域，我们再次看到了他们的实力——Trainium 芯片的性价比，相较于其他图形处理器（GPU）产品，也具备约40% 的优势。客户希望能更广泛地应用人工智能，而且要记住，目前人工智能的发展仍处于相对早期阶段，他们需要更高的性价比，并且对此高度关注。
正如我之前提到的，Trainium2 目前发展势头良好。在我看来，对于Trainium3，我们预计其性能将比Trainium2 提升约40%，而Trainium2 在性价比方面已经具备显著优势。当然，我们需要确保Trainium3 芯片能够顺利交付。
我们需要实现芯片的规模化量产，并快速交付给客户。同时，我们还需要持续完善软件生态系统—— 目前该生态系统正不断优化。随着我们获得更多像Project Rainier 这样的成功案例（例如Anthropic 在Trainium2 上取得的成果），Trainium 的市场认可度也会不断提升。我认为客户对Trainium 的前景非常乐观，我个人也对其充满信心。
会议主持人
下一个问题来自摩根大通（JPMorgan）的道格安穆斯（Doug Anmuth）。请提问。
道格安穆斯（Doug Anmuth）—— 董事总经理、互联网分析师
谢谢你们接受我的提问。安迪，我还是想围绕同一个主题提问。能否请您详细介绍一下Project Rainier 的架构特点，以及它的差异化优势？这对客户和AWS 分别意味着什么？您是否预计Rainier 项目将扩展到Anthropic 之外的其他客户？另外，如何利用Trainium3 芯片复制Rainier 项目的成功模式？谢谢。
安迪贾西（Andy Jassy）—— 首席执行官
好的。我认为，Project Rainier 对Anthropic 具有吸引力的核心原因，在于我们开发的Trainium2 芯片。首先，我们构建了一个规模庞大的计算集群，Anthropic 可以充分利用该集群开展业务。构建一个包含50 多万块芯片（未来将增至100 万块）的集群并非易事，这需要具备大规模的基础设施建设能力—— 这种能力是我们在AWS 长期发展过程中逐步建立起来的，在行业内并不常见。其次，Trainium2 芯片的性能和性价比也极具竞争力，这两点对客户来说都至关重要。我认为，Project Rainier 是专门为Anthropic 定制的项目，但目前已有许多其他客户表示有兴趣使用大规模Trainium 芯片集群，我们希望能借助Trainium3 芯片，为这些客户提供类似的服务。
会议主持人
下一个问题来自Evercore ISI 的马克马哈尼（Mark Mahaney）。请提问。
马克马哈尼（Mark Mahaney）—— 高级董事总经理
谢谢。我想请教两个问题：一个关于食品杂货业务，另一个关于未来员工人数规划。在食品杂货业务方面，关于易腐食品品类，我记得上一季度您提到，约70% 的用户此前从未在亚马逊购买过易腐食品。您之前用“革命性” 来形容这项服务，这是否意味着亚马逊可能不再需要运营Amazon Fresh（亚马逊生鲜）实体店了？亚马逊一直拥有DVD 配送货车密度方面的优势，您认为目前公司是否已达到足够的规模和配送速度，能够改变消费者的购物习惯，让他们在购买食品杂货时首先考虑亚马逊？您是否认为公司已达到这一阶段？
第二个问题关于员工人数。结合近期的相关新闻，想请教您对未来员工人数规划的看法。目前人工智能带来的效率提升，是否足以让公司在可预见的未来保持员工人数相对稳定？请谈谈员工人数调整的利弊，或者说可能带来的得与失。谢谢。
安迪贾西（Andy Jassy）—— 首席执行官
好的，马克。我先回答食品杂货业务相关的问题。亚马逊的食品杂货业务规模非常庞大。如果仅看我们的食品杂货业务（暂不包含Whole Foods Market（全食超市）和Amazon Fresh），过去12 个月的商品交易总额（Gross Merchandising Sales）已超过1000 亿美元，这一规模足以让我们跻身美国食品杂货零售商前三甲。
其中，很大一部分商品是超市中间货架上的品类，例如日用品、罐头食品、宠物食品以及健康美容产品等，这部分业务规模非常可观，且持续保持良好增长势头。此外，我们还拥有Whole Foods Market—— 它是有机食品领域的开拓者，其增长速度也快于大多数食品杂货公司，盈利前景良好。未来几年，我们将继续扩大Whole Foods 的实体门店布局。同时，我对我们推出的全新业态Daily Shop 也非常期待—— 这是Whole Foods 在城市地区推出的小型门店模式。目前我们已开设了三家，运营初期表现良好，未来大家有望看到更多Daily Shop 门店。正如你所提到的，我们一直致力于扩大实体业务布局。
我们仍在持续尝试不同的门店业态。目前我们最看好的，正是你所提及的—— 提供易腐食品当日送达服务的模式。试想一下，有多少客户每周会在亚马逊平台购物多次，购买洗发水、洗涤剂、纸杯、饮用水等商品；如果他们能在下单这些商品的同时，添加牛奶、鸡蛋、酸奶等易腐食品，将其放入同一个购物车，并且在几小时内收到所有商品，这种体验对客户来说极具吸引力。大约一年前，我们在少数几个市场推出了这项服务，客户的接受度远超我们预期—— 不仅迅速吸引了大量新用户购买易腐食品，而且这些用户后续再次购买易腐食品和其他食品杂货的频次也非常高。
目前，我们已将这项服务扩展到美国1000 个城市，到年底将覆盖2300 个城市。这项服务正显著改变我们食品杂货业务的发展轨迹和规模。我认为，多年来形成的“每周一次大规模采购食品杂货” 的消费习惯也在发生变化，而亚马逊在这一转变过程中发挥了重要作用。我们认为食品杂货业务具有巨大的增长潜力。这并不意味着我们会停止尝试其他实体业态，但目前我们在易腐食品当日送达服务方面取得的进展确实意义重大。
关于员工人数的问题，我想说明的是，我们几天前宣布的相关举措，并非出于财务考量，至少目前来看也与人工智能无关，其核心原因在于企业文化建设。
在过去几年里，亚马逊业务增长迅速，业务规模、员工数量、办公地点以及业务类型都大幅扩张，这导致员工人数远超以往，管理层级也相应增加。在这种情况下，我们可能会在不经意间削弱一线员工的主人翁意识—— 这些员工承担着实际工作，负责做出大多数“双向门” 决策（指那些可逆、可快速调整的决策），而这类决策本应在一线快速做出。层级过多会导致决策效率下降。作为管理团队，我们致力于将亚马逊打造成“全球最大的初创企业”，这意味着要减少管理层级、增强员工的主人翁意识、鼓励创新并加快决策速度。
无论是在亚马逊的历史上，还是在整个商业领域，当前技术变革正以前所未有的速度推进，在这样的时期，保持精简的组织架构、扁平化的管理模式以及快速的行动能力至关重要。这正是我们未来要坚持的方向。
会议主持人
下一个问题来自高盛（Goldman Sachs）的埃里克谢里丹（Eric Sheridan）。请提问。
埃里克谢里丹（Eric Sheridan）—— 董事总经理
非常感谢你们接受我的提问。安迪，我想请您谈谈在推广机器人与自动化技术方面面临的持续机遇，以及“实体人工智能”（Physical AI）在运营中的广泛应用前景。您认为这些技术在提升效率、以及为业务长期发展提供再投资能力方面，将发挥怎样的推动作用？非常感谢。
安迪贾西（Andy Jassy）—— 首席执行官
机器人技术是我们重点投资的领域之一。目前，我们的履约网络中已部署了超过100 万台机器人。尽管这一规模已相当可观，但我们仍有多项机器人创新项目正在推进中，因此预计未来机器人数量还将进一步增加。机器人技术对我们、客户以及员工都具有重要意义：它能提升安全性、提高生产效率、加快配送速度，并让人类员工能够专注于解决问题等他们擅长的工作。我们始终认为，人类员工将继续是履约网络的核心—— 从我们开始投入机器人技术研发以来，这一立场就没有改变过。我们预计，未来的履约网络将实现机器人与人类的互补协作。我相信，我们会继续在机器人技术领域进行大规模投资。
这项技术将在提升安全性、生产效率和配送速度方面发挥重要作用，最终也将有助于降低部分成本，从而让我们能进一步优化客户体验。
会议主持人
下一个问题来自TD Cowen 的约翰布莱克利奇（John Blackledge）。请提问。
约翰布莱克利奇（John Blackledge）—— 高级股票研究分析师（互联网与新媒体领域）
好的，谢谢。亚马逊如何看待未来智能体化商务（Agentic Commerce）的发展前景？您认为亚马逊未来将如何利用智能体，帮助客户在平台上购买商品？谢谢。
安迪贾西（Andy Jassy）—— 首席执行官
从长期来看，智能体化商务的发展前景让我和整个公司都感到非常兴奋。它有望为客户带来更优质的体验，也有望为电子商务行业注入新的活力。我认为，如果客户明确知道自己想买什么，那么亚马逊平台能提供数一数二的购物体验。
但如果客户不确定自己的需求，那么拥有实体销售人员的实体店在某些方面仍具有优势。显然，很多客户还是会在亚马逊上完成这类购物，但他们往往希望能通过提问获得帮助，从而缩小选择范围。在不断提问的过程中，他们希望能看到更多不同的商品选项。我认为，人工智能和智能体化商务将改变线上购物体验—— 未来，当客户不确定需求时，线上购物体验甚至会超越线下实体购物体验。当然，我们在智能体化商务领域也有自己的布局。正如我在开场发言中提到的，我们的人工智能购物助手Rufus 正不断优化，使用范围也在持续扩大。
我们还推出了“代买”（Buy for Me）功能—— 即使某些商品亚马逊平台没有库存，但其他商家有售，我们也会为客户展示这些商品；如果客户希望我们从这些商家的网站上代为购买，我们会提供这项服务。这两项功能目前都取得了不错的效果。同时，我们也在与第三方智能体开发商进行沟通，预计未来会与他们建立合作关系。这让我想起了多年前搜索引擎刚刚兴起时的场景—— 当时搜索引擎成为了商务发现的重要渠道，行业内需要探索合作模式。如今，搜索引擎为我们带来的引荐流量占比已非常小，第三方智能体带来的流量占比则更小。但我相信，我们最终会找到合适的合作方式。不过，合作的前提是要确保客户体验的优质性。
目前来看，第三方智能体提供的客户体验还不够理想—— 缺乏个性化服务、无法关联客户购物历史、配送时间预估常常不准确、价格信息也经常出错。我们必须找到方法来改善客户体验，并建立合理的价值交换机制。我认为，智能体化商务的令人兴奋之处和发展潜力在于，人工智能和智能体化商务解决方案将推动更多购物行为转移到线上。这对客户来说无疑是好事，对亚马逊来说也是如此—— 因为归根结底，客户会选择那些商品选择丰富、性价比高、配送快速可靠的平台进行购买。我认为这一趋势对我们非常有利。
会议主持人
最后一个问题来自贝尔德（Baird）的科林塞巴斯蒂安（Colin Sebastian）。请提问。
科林塞巴斯蒂安（Colin Sebastian）—— 董事总经理、高级股票研究分析师（互联网与数字媒体领域）
谢谢。下午好。首先想继续请教AWS 相关问题：AWS 此次增速回升，有多少来自核心基础设施业务，又有多少来自人工智能工作负载的变现？这其中，我想了解AgentCore 等服务在吸引企业到AWS 构建智能体方面发挥了多大作用。第二个问题关于广告业务增速回升：能否分别说明核心广告业务、DSP（需求侧平台）以及Prime Video 广告业务对增速的贡献？这将对我们的分析很有帮助。谢谢。
安迪贾西（Andy Jassy）—— 首席执行官
好的。我先回答AWS 相关问题。我们对本季度AWS 的业绩表现非常满意。在1320 亿美元年化营收基数上实现20% 的同比增长，这一成绩实属难得。目前AWS 发展势头良好，这一点有目共睹。我们看到，增长既来自人工智能领域（包括推理、训练以及Trainium 定制芯片的应用），也来自核心业务领域。
Bedrock 平台增长势头持续强劲，SageMaker 平台同样保持快速增长。我认为，目前有大量企业正在投入智能体开发工作。我相信，从长期来看，企业从人工智能中获取的价值，很大一部分将来自智能体。但目前，智能体开发过程仍过于复杂。企业需要更简便的工具，这正是我们推出Strands 的原因—— 这是一款开源工具，能让开发者基于任意模型创建智能体。更重要的是，当我们与那些重视安全性和规模的企业交流时发现，他们虽然已经开始开发智能体，但却缺乏能让他们放心将业务、客户体验和数据托付的基础构建模块，无法部署安全、可扩展的智能体。
这正是我们开发AgentCore 的初衷—— 就像AWS 早期为客户提供计算、存储、数据库等基础构建模块一样，我们为智能体部署打造了一套核心基础构建模块，帮助客户实现智能体的安全、可扩展部署。从客户反馈来看，AgentCore 的推出引起了强烈共鸣，目前市场上尚无同类产品。它改变了客户开发和部署智能体的时间规划，也提高了他们对智能体技术的接受度，市场反响非常积极。我认为，除了智能体相关业务，AWS 增长的另一个重要来源是，越来越多的企业重新启动了从本地数据中心向云端的迁移项目，而AWS 在这些迁移项目中持续赢得了绝大多数合作机会。综合来看，目前AWS 的发展势头良好，我相信这种强劲增长态势还将持续一段时间。
再来看广告业务。我们对广告业务的进展整体感到满意，本季度所有广告产品都实现了显著增长。我认为，广告业务的增长得益于我们独特的全漏斗广告产品组合。
从漏斗顶端（通常是品牌认知和大规模触达）来看，我们可以利用Prime Video 及其直播体育赛事资源；而到了漏斗底端（销售转化环节），我们可以利用赞助商品（Sponsored Products）等工具。很少有企业能像我们这样，拥有如此完善的全漏斗广告产品组合。再加上我们在受众筛选与开发方面的能力，以及在广告效果衡量方面的优势，最终为广告主带来了极高的广告投资回报率（ROI），这是其他平台难以比拟的。我认为，广告业务未来还有多个增长方向。首先是零售业务相关广告—— 目前全球零售市场中，线下零售仍占80%-85% 的份额，未来这一比例将逐步向线上转移，而人工智能将进一步加速这一趋势。因此，我们在现有零售业务中仍有巨大的广告增长空间。
其次是视频广告业务—— 我们进入该领域的时间不长，但目前已实现可观的广告营收，且仍处于发展早期阶段，未来增长潜力巨大。正如你所提到的需求侧平台（Amazon DSP），其增长速度也非常快。部分原因在于，我们之前在DSP 平台功能建设上存在一些短板—— 尽管我们早已具备受众定位、效果衡量（如亚马逊营销云Amazon Marketing Cloud）等核心功能，但在其他一些客户关注的功能上仍有欠缺。在过去20 个月里，团队大幅完善了这些功能，如今客户认为我们的DSP 平台已具备全面的功能。再加上我们近期达成的一系列合作—— 例如与Roku 的合作让我们获得了美国最大的联网电视（Connected TV）受众群体；此外，我们还为DSP 客户提供了接入Netflix、Spotify 和SiriusXM 广告库存的机会—— 这些都为DSP 业务增长提供了有力支撑。目前，DSP 业务增长迅速，我们对其未来发展充满信心。当然，我们仍有很多工作要做，但可以肯定的是，DSP 业务远未达到增长天花板。
会议主持人
感谢各位参加本次会议并提出宝贵问题。会议重播将在我们的投资者关系网站上保留至少三个月。感谢大家对亚马逊的关注，期待下一季度与大家再次交流。
(Article source: CLS)
