Amazon.com, Inc. (AMZN.US)FY25Q4 conference call: Capital expenditure in 2026 will reach $200 billion, with the majority being invested in AWS.

Recently, Amazon.com, Inc. (AMZN.US) held its FY25Q4 earnings conference call. The company stated that it plans to invest approximately $200 billion in capital expenditures in 2026, with the majority going towards the high-demand AWS to support customer core business and AI workloads, and to quickly convert new compute power into revenue. The company reported a backlog of orders of $244 billion, a 40% year-over-year increase and a 22% increase from the previous quarter. There are many transactions in progress, with significant demand for AWS (including AI and core AWS). The capital expenditures and most of the capacity are consumed by external customers. AWS had a 35% operating profit margin in the fourth quarter, a 40 basis point increase year-over-year. This margin will fluctuate over time, certainly affected by AI investments and corresponding capital expenditure depreciation, but efforts will be made to offset this through efficiency and cost reduction. The company noted that businesses are refocusing on migrating from on-premises to the cloud, with strong continued growth in core non-AI workloads. Additionally, to reduce AI inference costs, the company has developed a custom chip called Trainium. Over 1.4 million Trainium2 chips have been delivered, offering a 30%-40% cost advantage over similar GPUs, generating annual revenue in the tens of billions of dollars. Demand for Trainium3 is strong, with all capacity for the first half of 2026 already booked. Trainium3 has started shipping, with a 40% higher cost efficiency than Trainium2, and market interest is high, with almost all supply expected to be booked by mid-year 2026. Trainium4 is in development (planned for release in 2027), and discussions have begun on Trainium5. In the retail business, the company stated that approximately 300 million customers used Rufus in 2025, with customers using Rufus having a 60% higher likelihood of completing a purchase. Q&A: Q: Can you provide more insights on confidence in strong long-term investment capital returns, including the duration of the current capital expenditure cycle, the profitability level to focus on, and the minimum level of free cash flow that you do not want to fall below during the capital expenditure cycle? A: We are putting all the capacity we get into serving customers, and this capacity will be immediately used by customers. We see a long-term incremental revenue trajectory: from other customers, backlog, and commitments, especially in AI services, which are gradually reflected in our profit and loss statement. This is reflected through CapEx and AWS's operating profit margin. AWS had a 35% operating profit margin in the fourth quarter, a 40 basis point increase year-over-year. This margin will fluctuate over time due to AI investments and corresponding capital expenditure depreciation, but we will also strive to offset these impacts through efficiency and cost reduction. We see strong investment capital returns and strong demand for these services, so we favor investments in this area. The capital we plan to spend this year is mainly directed towards AWS. Some will be used for non-AI core workloads growing faster than expected, but most will be used for AI. We believe that all existing customer experiences will be reshaped by AI and there will be many new experiences we never imagined. Over time, we will see optimizations in the AI field - inference services (which will form the main body of long-term AI workloads), increased service utilization, and gradually normalized prices. Those who not only have excellent infrastructure but also have components that provide better value for customers and better economic benefits for companies (such as our Trainium chips supporting most Bedrock services) will have a financial advantage. Q: Can you discuss how the Project Rainier and Anthropic collaboration has progressed after the first full quarter? Additionally, the press release mentioned 500,000 chips, but a few months ago, it was also mentioned that it reached 1 million chips, can you clarify this? A: We are very excited about the future of Trainium. Compared to similar GPUs, Trainium has a 30% to 40% cost advantage, making it very attractive to customers. Project Rainier is a data center built in collaboration with Anthropic to train its next-generation Claude models using Trainium chips. 50,000 chips are currently in use, and this number will continue to increase. Additionally, Anthropic will also use a considerable number of Trainium2 chips for workloads outside of Project Rainier and its own API. Trainium is already a billion-dollar annual revenue business, and capacity has been fully booked. Trainium3 has started shipping, with a 40% higher cost efficiency than Trainium2, and market interest is high, with almost all supply expected to be booked by mid-year 2026. We are developing Trainium4 (planned for release in 2027), with strong market interest, and discussions have already begun on Trainium5. We have become a strong chip company. Our CPU chip Graviton is about 40% more cost-efficient than similar x86 processors, and 90% of the top 1000 AWS customers widely use it. The combined revenue of Trainium and Graviton has exceeded $10 billion annually and is still in its early stages. Q: You mentioned earlier that the AI market is currently somewhat top-heavy, with spending concentrated in a few AI native labs. Looking ahead to 2026, how do you see this changing? Specifically, how do you plan to expand relationships with companies like OpenAI to help Amazon.com, Inc. in its AI efforts on both the retail side and the AWS side? A: Currently, we see a barbell-shaped market demand in the AI space. On one end are AI labs and some blockbuster applications, which are consuming huge amounts of computing power. On the other end are many businesses that are using AI to improve productivity and reduce the cost of some workloads, such as customer service, business process automation, or anti-fraud measures. In the middle of the barbell are all the production workloads of businesses. Businesses are currently in different stages of evaluating how to migrate, starting to migrate, and investing in production. I believe that this middle part is likely to be the largest and most enduring market. Although we have seen incredible growth in AI demand, as more people with AI backgrounds increase, inference costs continue to decline (which is what we are striving for through Trainium and our hardware strategy), and businesses further succeed in migrating workloads, the middle part of the barbell will be the main contributor to demand. Regarding how to expand relationships with companies like OpenAI, we have very important relationships with many different companies. We announced an agreement with OpenAI in November, which is a significant agreement, and we hope to continue expanding partnerships over time. But this AI movement will not be limited to a few companies; it will involve thousands of companies in the future. Q: In terms of the retail business, how do you view the impact of intelligent agents? As consumers get better answers, this seems to potentially compress the consumer funnel, thus affecting the retail business and the in-store advertising section. A: I am very optimistic about the intelligent shopping experience that customers will eventually use, which is beneficial to customers and makes shopping more convenient. This is also why we are investing heavily in our proprietary shopping assistant Rufus. Rufus has improved a lot and continues to improve every month. Approximately 300 million customers used Rufus in 2025, and customers using Rufus have a significantly higher likelihood of completing a purchase, with significant usage and growth. Furthermore, we will also establish partnerships with third-party horizontal agents who can facilitate shopping. We have to explore better customer experiences together because these horizontal agents currently have no shopping history for any users and often get product details and prices wrong. We need to find a mutually meaningful customer experience and value exchange method, although currently this traffic and sales percentage is relatively small, I am hopeful about this. But we need to consider how many consumers would prefer to use a horizontal agent rather than a retail provider that has their entire shopping history and accurate data for precise shopping or exploration. I believe many customers will ultimately choose to use excellent shopping agents provided by retailers because what consumers really need in retail is wide selection, low prices, fast delivery, and a retailer they trust to take care of them. Horizontal agents may be good at aggregating choices, but retailers are much stronger in these four things. Therefore, I am very optimistic that people will use our shopping agents, and I also hope that over time, we can work with third-party agents to address the aforementioned issues. Q: Concerning the global retail business, what are the sources of efficiency improvement expected this year? Additionally, what are the investment areas driving sustainable growth (such as Siasun Robot & Automation)? A: In terms of investments driving continuous growth in the retail business, the core demand drivers remain the same. We will strive to expand product selection, including introducing more luxury brands (such as L'Oreal's rapidly growing business and high partner satisfaction) and continually increasing daily essentials. Daily essentials currently account for one-third of our total product sales volume, with significant growth. The more customers rely on us to purchase daily essentials and low-priced items, the more they tend to shop with us for subsequent purchases. The significant improvement in delivery speed over the past three years has been crucial in winning more business in daily essentials and groceries. The rapid business service (Amazon Now) has seen rapid growth in countries like India, the UAE, and Mexico, for example, in India, customers trying fast business shopping have tripled their shopping frequency. We are also continuing to expand in the fresh daily delivery sector, currently achieving same-day fresh delivery in thousands of cities globally, where nine out of ten top-selling products are fresh. Customers buying fresh produce then double their shopping frequency. In terms of efficiency improvements, we have ongoing optimization projects. For example, we are deepening regionalization of our fulfillment network in the U.S., increasing the number of regions from 8 to 10, and making products closer to customers faster by optimizing inbound logistics. Siasun Robot & Automation technology is another focus, with over 1 million Siasun Robot & Automation units in the fulfillment network handling various functions. This technology helps free up employees from repetitive tasks, thereby increasing business productivity, ensuring employee safety, and realizing real cost benefits. Q: Regarding AWS, can you discuss the revenue backlog situation as of the fourth quarter and how you view the situation between internal use cases and external customer demands, and whether there is a supply-demand imbalance in AI, and how do you plan to close these gaps as more capacity comes online in 2026? A: Our backlog of orders reached $244 billion, a 40% year-over-year increase and a 22% increase from the previous quarter. There are many transactions in progress, and there is currently significant demand for AWS (including AI and core AWS). The capital expenditures and most of the capacity are consumed by external customers. Amazon.com, Inc. has always been a very important large customer of AWS, but it has always been a very small part of the overall business (including AI). Internally, we are using AI extensively, with over 1,000 deployed or in-progress AI applications covering everything from shopping assistants like Rufus, large-scale generative AI applications like Alexa +, to predictions in the fulfillment network, customer service and chatbots, tools for helping brands create and optimize full-funnel advertising campaigns, and defensive alerts in live sports (like Thursday Night Football). External demand, as mentioned earlier, sees AI labs consuming a lot of computing power for training, inference, and research; while businesses are investing in various workloads such as customer service automation, business process automation, anti-fraud measures, complete application reconstruction, intelligent coding applications, and legal applications. In the past 12 months, we have added 3.9 gigawatts of power capacity, double the capacity when our annual revenue was $80 billion in 2022, and we expect to double it again by the end of 2027. Just in the fourth quarter, we added 1.2 gigawatts sequentially. Our team is actively, innovatively increasing capacity at the fastest pace and will add more in 2026, 2027, and 2028. We are very optimistic about being able to sustain growth levels similar to this over time.