Amazon(AMZN.US)FY25Q4 Earnings Call: Capital Expenditures to Reach $200 Billion in 2026, Mostly Allocated to AWS

Recently, Amazon (AMZN.US) held its FY25Q4 earnings call. The company announced plans to invest approximately $200 billion in capital expenditures by 2026, with most of the investment directed toward the booming AWS to support customers’ core businesses and AI workloads, and to quickly convert additional computing power into revenue.

The company stated that backlog orders reached $244 billion, up 40% year-over-year and 22% quarter-over-quarter. Many transactions are underway, and demand for AWS (including AI and core AWS) remains very high. Most of the capital spent and capacity owned are consumed by external customers.

AWS’s operating profit margin for Q4 was 35%, up 40 basis points year-over-year. This margin will fluctuate over time and will certainly be affected by AI investments and the depreciation of related capital expenditures, but efforts will be made to offset this through efficiency improvements and cost reductions.

The company pointed out that enterprises are refocusing on migrating from on-premises to the cloud, with continued strong growth in non-AI core workloads. Meanwhile, to reduce AI inference costs, the company has developed its own custom chips, Trainium. Over 1.4 million Trainium 2 chips have been delivered, offering 30-40% better cost-performance than comparable GPUs, generating billions of dollars in annual revenue. Demand for Trainium 3 remains strong, with all capacity booked before mid-2026.

Trainium 3 has begun shipping, offering 40% better cost-performance than Trainium 2, with strong market interest. Almost all supply is expected to be booked before mid-2024. Trainium 4 is under development (scheduled for release in 2027), and discussions about Trainium 5 have already begun.

In retail, the company stated that approximately 300 million customers used Rufus in 2025, and customers using Rufus are about 60% more likely to complete a purchase.

Q&A Session

Q: Can you provide more insights into your confidence in long-term capital returns from strong investments? For example, how long will the current capital expenditure cycle last, what profitability levels should be focused on, and what is the minimum free cash flow level you aim to maintain during this cycle?

A: We are investing all the capacity we acquire into serving customers, and these capacities are used immediately. We see a long-term incremental revenue trajectory driven by other customers, backlog, and commitments customers are eager to make with us, especially in AI services, which are gradually reflected in our income statement.

This is reflected both through CapEx and AWS operating profit margins. AWS’s Q4 operating margin was 35%, up 40 basis points year-over-year. This margin will fluctuate over time and will be influenced by AI investments and depreciation of related capital expenditures, but we will also work to offset this through efficiency and cost reductions. We see strong investment returns and robust demand for these services, which makes us favor investments in this area.

Our planned capital expenditures this year will mainly go toward AWS. Some will support faster-than-expected growth in non-AI core workloads, but most will be dedicated to AI. We firmly believe that all existing customer experiences will be reshaped by AI, and many new experiences we haven’t yet imagined will emerge.

Over time, in AI, inference services (which will constitute the main part of long-term AI workloads) will continue to optimize, service utilization will increase, and prices will gradually normalize. Those with excellent infrastructure and components that can offer better cost-performance to customers and generate better economic benefits for the company (such as our Trainium chips powering most of Bedrock services) will have a financial advantage.

Q: Can you update us on Project Rainier and your collaboration with Anthropic after the first quarter? Also, the press release mentioned 500,000 chips, but a few months ago, 1 million was mentioned. Could you clarify?

A: We are very excited about the growth and future of Trainium. Compared to similar GPUs, Trainium offers a 30-40% cost-performance advantage, which is very attractive to customers. Project Rainier is a data center built in partnership with Anthropic to train their next-generation Claude model, based on Trainium chips.

Here, we mentioned 500,000 chips, and this number will continue to grow. Additionally, Anthropic is using a significant number of Trainium 2 chips for workloads outside of Project Rainier and for their own API.

Currently, Trainium is a business generating billions of dollars in annual revenue, and capacity is fully booked. Trainium 3 has started shipping, offering 40% better cost-performance than Trainium 2, with strong market interest. Almost all supply is expected to be booked before mid-2024.

We are developing Trainium 4 (scheduled for 2027), with strong market interest, and discussions about Trainium 5 are underway. We have become a powerful chip company.

Our CPU chips, Graviton, are about 40% more cost-effective than comparable x86 processors, with 90% of the top 1,000 AWS customers using them extensively. Combined, Trainium and Graviton form an over $10 billion annualized business, still in early stages.

Q: You previously mentioned that the AI market is somewhat top-heavy, with spending concentrated among a few AI-native labs. How do you see this changing by 2026? Specifically, how do you plan to expand relationships with companies like OpenAI to support Amazon’s retail and AWS AI efforts?

A: Currently, we see the AI market demand as a bell-shaped curve. On one end are AI labs and some blockbuster applications consuming massive compute power. On the other end are many enterprises using AI to boost productivity and reduce costs in areas like customer service, business automation, and fraud detection.

The middle of the bell curve comprises all enterprise production workloads. These companies are at various stages of evaluating, migrating, and deploying workloads. I believe this middle segment will be the largest and most persistent market.

Despite the incredible growth in AI demand, as more AI-skilled personnel emerge, inference costs continue to decline (which we aim to achieve through Trainium and our hardware strategy), and enterprises succeed further in migrating workloads, the middle segment will generate the majority of demand.

Regarding expanding relationships with companies like OpenAI, we have important relationships with many different firms. We announced a significant agreement with OpenAI in November and hope to expand our partnership over time. But this AI movement will not be limited to a few companies; it will involve thousands of firms in the future.

Q: Regarding retail, how do you view the impact of intelligent agents? As consumers get better answers, it might lead to a compression of the shopping funnel, affecting retail and on-site advertising.

A: I am very optimistic about end customers using intelligent shopping experiences, which benefit customers by making shopping more convenient. This is a key reason we are heavily investing in our own shopping assistant, Rufus. Rufus has become much better and is continuously improving every month.

In 2025, about 300 million customers used Rufus, and those customers are approximately 60% more likely to complete a purchase, with usage and growth being very significant.

We are also establishing partnerships with third-party horizontal agents capable of enabling shopping. We need to jointly explore better customer experiences because these third-party agents currently have no shopping history, often misrepresent product details and prices.

We need to find a meaningful way to exchange value and create a good customer experience. Although this traffic and sales share is still relatively small, I am optimistic about it.

However, we should consider how many consumers would prefer to use a third-party horizontal agent instead of the retail site’s own, which has complete shopping history and accurate data, making precise shopping and discovery easier.

I believe many customers will ultimately choose to use the retail site’s excellent shopping agents because what consumers truly need in retail are broad choices, low prices, fast delivery, and a trusted retailer that cares for them.

Horizontal agents may be good at aggregating options, but retailers are much better at delivering these four core aspects. Therefore, I am very optimistic that people will use our shopping agents, and I look forward to collaborating with third-party agents over time to address these issues.

Q: Regarding global retail, what sources of efficiency improvements do you expect this year? Also, which areas will drive more sustained growth, such as robotics?

A: The core drivers of retail growth through investments remain unchanged. We will continue expanding product selection, including adding more luxury brands (L’Oréal’s business is growing rapidly and partners are satisfied), and increasing daily essentials.

Daily essentials now account for about one-third of our total sales, with significant growth. As customers rely more on us for daily necessities and low-priced items, they tend to shop more with us afterward.

The significant improvement in delivery speed over the past three years has been key to winning more daily essentials and fresh grocery business. Fast delivery services (Amazon Now) are expanding rapidly in India, UAE, Mexico, and other countries. In India, customers trying fast delivery shop three times more often than before.

We are also expanding in fresh produce same-day delivery, now available in thousands of cities worldwide. In these cities, nine of the top ten products by regional sales are fresh items. Customers buying fresh produce tend to double their shopping frequency afterward.

In terms of efficiency, we are continuously optimizing. For example, our fulfillment network in the US is becoming more regionalized, increasing from 8 to 10 regions, and inbound logistics are optimized to bring products closer to customers faster.

Robotics is another focus. Currently, over 1 million robots operate within our fulfillment network, handling various functions. They help free employees from repetitive tasks, improve productivity, ensure safety, and deliver real cost benefits.