Amazon (Minutes): Computing power supply to double again by 2027

The following is compiled by Dolphin Research$Amazon(AMZN.US) 2025 Q3 Earnings Call Minutes, for earnings interpretation please visit "AWS Major Reversal, Has Amazon Finally 'Turned the Corner'?"

I. Review of Core Financial Data

II. Detailed Information from the Earnings Call

2.1 Key Statements and Strategic Outlook from Executives

1. AWS Business Momentum

Growth Trend: AWS growth rate reached its highest level since 2022, with year-on-year growth returning to 20.2%. Achieving 20% year-on-year growth on an annualized run rate of $132 billion is extraordinary.

Customer Base and Order Backlog: As of the end of the third quarter, the unfulfilled contract balance reached $200 billion, and does not include several new contracts announced in October—the total of these contracts exceeds the total transactions of the entire third quarter. Customers prefer to run core and AI workloads on AWS due to its stronger functionality, security, and operational performance.

Leadership: AWS offers a richer variety of services and deeper functionality than any competitor, and continues to innovate rapidly, which is the key reason Gartner has named AWS a leader for 15 consecutive years.

2. AI Research and Development Progress

Models and Development: SageMaker simplifies enterprise construction and deployment of proprietary foundational models; Bedrock offers customers leading foundational model choices and excellent cost-effectiveness.

Agents: The Strands platform allows developers to create agents based on any foundational model; the AgentCore infrastructure building block aids in deploying secure and scalable agents (customer cases include Ericsson, Sony, Cohere Health).

Dedicated Agents: The Kiro IDE preview in the coding field attracted over 100,000 developers within days; the migration tool Transform saved customers 700,000 hours of manual operations; Quick Suite brings consumer-grade AI experience into work scenarios; Amazon Connect contact center service annual revenue exceeded $1 billion.

3. Capacity Expansion and Chip Strategy:

Power Capacity: Current AWS power capacity has doubled compared to 2022 levels, and is expected to double again by 2027. At least 1 gigawatt is planned to be added in the fourth quarter alone.

Custom Chips: Project Rainer AI computing cluster is equipped with nearly 500,000 Trainium2 chips, and Anthropic is expected to run on over 1 million Trainium2 chips by the end of the year. Trainium2 business scale reached several billion dollars, with a quarter-on-quarter growth of 150%.

Partners: Continues close collaboration with NVIDIA, AMD, and Intel.

Future Outlook: We are building Bedrock to become the world's largest inference engine, and in the long term, we believe Bedrock's business scale can rival EC2. The newly added computing power is rapidly being commercialized, which is still in the early stages, creating rare development opportunities for AWS customers.

4. Global Store Business and Customer Experience Innovation

Core Advantages: Continues to focus on product selection, low-price strategy, and convenient experience—product selection increased by 14% since last quarter.

Fresh and Daily Necessities: Daily necessities business growth is nearly twice that of other business segments. Over 1,000 towns can enjoy free same-day delivery, allowing customers to order both fresh and millions of regular items simultaneously.

Delivery Innovation: Introduced the "Add-on Delivery" button, usage has exceeded 80 million times. Three-hour delivery service has been launched in some cities.

Infrastructure Investment: Invested over $4 billion to expand the rural delivery network across the U.S., increasing the coverage rate of same-day and next-day delivery services in rural communities by 60%.

AI Empowerment: Intelligent shopping assistant Rufus monthly active users increased by 140% year-on-year, and customers using Rufus are 60% more likely to complete purchases, expected to achieve over $10 billion in annual incremental sales. Generative AI voice functionality has been extended to millions of products.

Third-party Sellers: Global third-party seller product share reached 62%, an increase of 200 basis points compared to the same period last year; over 1.3 million sellers use generative AI capabilities to accelerate product information release.

5. Advertising Business and Content Ecosystem

Growth Drivers: Full range of image advertising products performed strongly, demand-side platform Amazon DSP supports full-funnel investment.

Partner Expansion: Achieved system integration with Netflix, Spotify, and SiriusXM, expanding premium advertising resources.

Prime Video Live Sports: Exceeded targets in the 2025-2026 annual advertising pre-sale negotiations; NBA event first night double venue attracted an average of 1.25 million viewers in the U.S., achieving double-digit growth compared to cable TV.

AI Tools: The Agentic AI tool and creative studio launched in September can shorten the creative process from weeks to hours.

6. Emerging Business and Future Investment Focus

Alexa+: User and device interaction frequency doubled, conversion rate for completing shopping conversations and ultimately placing orders increased fourfold.

Project Kuiper: Satellite count expanded to over 150, achieving transmission speeds of over 1 gigabit per second.

Zoox Autonomous Driving: Already operational in Las Vegas, Washington D.C. becomes the eighth test site.

Investment Strategy: Continues to increase investment, especially in the AI field, due to its enormous opportunities, expected to achieve strong capital returns in the long term. Also investing in logistics and delivery networks to support business growth, improve delivery speed, and reduce service costs.

2.2 Q&A

Q: Regarding AWS, how do you currently feel about capacity levels, are you currently constrained by capacity? You mentioned Trainium3 demand and potential expansion of the customer base, can you talk about what demand for Trainium you see beyond major customers?

A: In terms of capacity, we have added a significant amount of capacity, last year we added 3.8 gigawatts, and in the fourth quarter we will add over 1 gigawatt, and we expect total capacity to double by the end of 2027. We are currently introducing a large amount of capacity, and industry-wide, the bottleneck may be power. I think at some point, it may shift to chips, but we are introducing a large amount of capacity. And, the speed at which we introduce capacity is as fast as we monetize it.

Regarding demand for Trainium beyond major customers, first, Trainium2 is performing very well, it is a multi-billion dollar business, with revenue growing 150% quarter-on-quarter. We are now seeing truly large-scale projects, such as the Rainier project with Anthropic, where they are training the next version of Claude on 500,000 Trainium2 chips, expanding to 1 million chips by the end of the year.

Currently, the number of Trainium2 customers is not large but the scale is significant. Because Trainium's price-performance is 30% to 40% higher than other options, and when customers start considering moving their production workloads to AI at scale and using inference, they are very concerned about price-performance. Therefore, we have great demand for Trainium.

Trainium3 should be previewed by the end of this year, with larger-scale mass production coming in early 2026, we have many customers, including large and medium-sized ones, who are very interested in Trainium3.

Q: Regarding Trainium chips, there are questions about its positioning compared to other third-party chips in the market. What key obstacles do you think Trainium3 needs to overcome to achieve broader adoption and continue its development, rather than relying on third-party chips in the short term to meet broader demand?

A: We will always offer customers multiple chip choices, consistent with AWS's approach to every major technology building block. We have a very deep and long-term partnership with NVIDIA, with large procurement volumes, and expect to continue procuring in the future.

But what sets us apart from other tech companies is that we have a very strong in-house chip team. This is first reflected in the CPU aspect with Graviton, which has a price-performance ratio about 40% higher than other x86 processors.

Now in the AI custom chip aspect, Trainium also offers customers a similar magnitude of price-performance advantage compared to other GPU options. Customers need better price-performance to use AI freely, and they care deeply about this.

Trainium2 is doing well, I expect Trainium3 to improve efficiency by about 40% over Trainium2, and Trainium2 already has a significant price-performance advantage. Of course, we must successfully deliver chips, quickly achieve large-scale mass production, and continuously improve the software ecosystem. As successful cases like the Rainier project on Trainium2 with Anthropic increase, Trainium's credibility will continue to rise, customers are very optimistic about this, and I am personally optimistic as well.

Q: Can you talk about the architecture and differentiated advantages of the Rainier project, what does this mean for customers and AWS? Do you expect the Rainier project to expand beyond Anthropic? And how to replicate the Rainier project with Trainium3 chips?

A: The appeal of the Rainier project to Anthropic lies in the Trainium2 chips. We have built a very large cluster for them to use extensively. Building a cluster with 500,000 and soon to be 1 million chips is not easy, this kind of large-scale infrastructure achievement is difficult to achieve in the industry. Therefore, part of the advantage lies in our extraordinary infrastructure capabilities built over the long term at AWS. Of course, the performance and price-performance of the chips themselves are also crucial, both are important. The Rainier project is specifically built for Anthropic, but we also have many other customers interested in using large Trainium chip clusters, and we hope to provide them with such opportunities through Trainium3.

Q: Regarding the fresh grocery business, does this mean Amazon Fresh physical stores are no longer needed? Do you have enough scale and speed to change consumer habits and make Amazon the preferred grocery channel?

Regarding employee numbers, given the efficiency improvements brought by AI, how do you view employee numbers in the future?

A: We have a large grocery business, with total product sales exceeding $100 billion in the past 12 months, making us one of the top three grocers in the U.S. This includes consumer goods, pet food, etc., growing rapidly. Whole Foods is a pioneer in organic food, with strong profitability, and will expand physical store scale in the future. We are also launching small city stores "daily shop," which have started well.

What we are most optimistic about is the fresh same-day delivery service. Customers can add fresh items like milk and eggs while ordering daily necessities, delivered within hours. This service has expanded to 1,000 cities in the U.S., and will reach 2,300 by the end of the year, significantly changing the trajectory of the grocery business. We believe the traditional habit of weekly concentrated shopping is changing, and we have great potential in this area.

Regarding employee numbers, recent adjustments are mainly due to cultural considerations rather than financial or AI-driven. As the business expands rapidly, organizational layers increase, potentially weakening frontline employees' decision-making power and execution. The leadership team is committed to maintaining the "world's largest startup" operating model, which means reducing layers, enhancing employee autonomy, and accelerating innovation. In the current period of technological change, maintaining a lean, flat, and fast-moving approach is crucial.

Q: Can you talk about the opportunities brought by promoting more robots, automation, and physical AI in operations? How should we view them as potential efficiency drivers and long-term business reinvestment capability drivers?

A: Robotics is a very important investment area for us. Currently, there are over 1 million robots in our fulfillment network. Although this number is already considerable, we are still innovating significantly, so we expect more in the future. Robotics is crucial for us, our customers, and team members because they improve safety, enhance productivity, speed up processes, and allow our team members to focus on problem-solving and what they do best.

We expect employees to always be at the core of our fulfillment network, just as when we first started using robots. Over time, we expect to build a fulfillment network where robots and humans complement each other. But we believe you will continue to see us invest heavily in this area, which will help improve safety, productivity, speed, and ultimately optimize some costs, allowing us to continuously improve the customer experience.

Q: How does Amazon view the future development of shopping assistants? How will Amazon use agents to help customers purchase products in the future?

A: We are very excited about the long-term prospects of agent commerce. It has the opportunity to benefit customers and is very favorable for e-commerce. When customers know exactly what they want to buy, Amazon can provide a top-notch experience; but when customers are uncertain, physical stores currently have an advantage. AI and purchasing will change the online experience, making the filtering process under uncertain demand better than in physical environments.

We have our own shopping assistant efforts, such as Rufus, which is constantly improving and being more widely used, and features like "Buy for Me," which can purchase products from other merchants even if we don't stock them, have been successful. We also look forward to collaborating with third-party agents, but this requires finding ways to improve the customer experience, as there are currently issues with lack of personalization, shopping history, delivery estimates, and frequent price errors. We believe AI and agent commerce solutions will increase the total volume of online shopping, which is good for customers and Amazon because ultimately customers will choose merchants that offer the widest selection, high value, and can consistently deliver quickly and reliably, which is beneficial for us.

Q: How much of AWS's accelerated growth this quarter is driven by core infrastructure business, and how much is driven by AI workload monetization? How important has new services like AgentCore become in attracting enterprises to use AWS to build agents?

Additionally, regarding the accelerated growth of the advertising business, can you differentiate the contributions of core advertising, DSP, and Prime Video?

A: We are very pleased with AWS's performance this quarter, achieving 20% year-on-year growth on an annualized run rate of $132 billion is extraordinary. Our growth momentum is reflected in both the AI field and core infrastructure business.

In AI, we see growth in inference, training, and the use of Trainium custom chips, with Bedrock and SageMaker also continuing to grow rapidly. Many companies are currently building agents, and we believe the value enterprises will derive from AI in the future will largely come from agents.

However, building agents is currently more difficult than expected, and enterprises lack the secure, scalable building blocks needed to support their business, customer experience, and data security. This is precisely why we developed AgentCore—to provide a set of primitive modules similar to AWS's early foundational components. Customer response has been strong, and there are currently no similar products, which is changing their timeline and acceptance for building agents. Additionally, we see many enterprises restarting migration from local infrastructure to the cloud, and we have won most of these transformation projects. Based on current momentum, we believe this growth rate can be sustained for some time.

In advertising, we are pleased with the significant growth of all our advertising products this quarter. We have an extraordinary full-funnel solution. From Prime Video and live sports events for brand awareness and large-scale coverage, to product advertising for sales conversion, combined with our unique audience planning and effectiveness measurement advantages, providing advertisers with excellent advertising investment returns.

We expect multiple growth areas in the future: first, our store business still has huge opportunities, as 80% to 85% of global retail sales are still in physical stores, this proportion will reverse in the future, and AI will accelerate this process. Second, video advertising, although just starting, has already contributed significant advertising revenue and is still in the early stages, will continue to be a growth focus. Finally, our demand-side platform Amazon DSP is also growing very rapidly. Over the past 20 months, the team has addressed key functionality gaps pointed out by customers, and now our DSP is considered fully functional. Combined with integration with Roku (which has the largest connected TV coverage in the U.S.), Netflix, Spotify, and SiriusXM, our DSP platform is strong and growing rapidly, and we are very optimistic about this.

<End of Text>

Risk Disclosure and Statement of this Article:Dolphin Research Disclaimer and General Disclosure