Amazon (Minutes): AI computing power still has bottlenecks, both merchants and consumers are hoarding goods under tariffs

Below is the$Amazon(AMZN.US) 2025 Q1 earnings call Minutes. For earnings analysis, please refer to《AI Heavy Investment, Tariff Turbulence: Is Amazon Re-entering a "Squatting Period"?》

1. Key Earnings Highlights:

II. Detailed Earnings Call Content

2.1. Core Messages from Management:

1. Delivery Service Optimization: Over the past few years, significant progress has been made in improving the efficiency and cost-effectiveness of our fulfillment network. A key milestone was dividing the national fulfillment network into multiple regional hubs. By storing goods closer to customers, we can deliver more orders faster and reduce shipping costs.

Our inbound network, which is how we transport goods to each fulfillment center, previously did not fully utilize this new regionalized structure. We redesigned it and recently launched a new inbound architecture, increasing the proportion of products we can place in each fulfillment center, improving delivery speed, and reducing our service costs.
Looking ahead, we will continue to refine the newly designed inbound network, expand same-day delivery stations, and add more robots and automation equipment in warehouses. We will also increase the number of delivery stations in rural areas of the U.S., enabling faster delivery to people living in less densely populated areas.

2. Views on Rising Tariffs for Retail Business: It’s unclear what the final tariff levels will be or when they will be finalized. So far, we haven’t seen any signs of weakening demand. To some extent, we’ve observed increased purchases in certain categories, which may indicate consumers are stockpiling ahead of potential tariff impacts.

We also haven’t seen significant increases in the average selling price of retail goods. This reflects some advance purchases by our first-party sales and arrangements made by third-party sellers. Most sellers haven’t adjusted prices yet. Of course, this could change depending on the final tariff situation.

In times of uncertainty, customers tend to choose suppliers they trust most. Given our wide product selection, low prices, and fast delivery, our market share has grown during past periods of uncertainty, and this could happen again.

3. Essentials: In Q1, our daily essentials grew more than twice as fast as other businesses. On Amazon U.S., 1 out of every 3 items sold is a daily essential. Even excluding Whole Foods Market and Amazon Fresh, Amazon is one of the largest grocery retailers in the U.S., with total sales exceeding $100 billion last year.

4. Advertising Business: We’re working to become the best platform for brands of all sizes to grow their businesses. Our full-funnel advertising products are performing strongly, reaching an average of over 275 million ad audiences in the U.S. alone. Amazon Ads provides brands with tools to reach target audiences on our entertainment platforms (e.g., Prime Video, Twitch, IMDb), live sports events (e.g., NFL, NBA), audio content (e.g., Amazon Music, Wondery), and of course our online store, as well as external sites like Pinterest and BuzzFeed.

All our audience analysis and performance measurement capabilities apply to ads placed through Amazon DSP on premium third-party publishers. Our secure "data clean room" allows advertisers to analyze data, generate core marketing metrics, and understand their cross-channel performance.

5. AWS Business: Currently, over 85% of global IT spending remains on-premises, meaning it hasn’t migrated to the cloud. In my view, this is likely to reverse over the next 10 to 20 years.
Before this generation of AI technology emerged, we believed AWS could eventually achieve annual revenues in the tens of billions. Now, we think its potential is even greater. We believe every customer experience will be reshaped by AI, as seen in the 1,000+ AI applications built internally at Amazon, including Alexa+, fulfillment networks, robotics, shopping, Prime Video, and advertising.
6. Trainium2 Chip: For developers building models, our new custom AI chip, Trainium2, is now in large-scale deployment and has been widely welcomed with strong demand. Trainium2 offers 30% to 40% better performance-price ratio compared to GPU-based instances, which is highly attractive.
For AI to achieve its expected success, inference costs need to drop significantly.

7. Other Software Services: In the middle layer, Amazon Bedrock’s fully managed service provides access to high-performance foundation models and a range of attractive features, making it easier to build high-quality generative AI applications.
We continue to iterate quickly on Bedrock, adding Anthropic’s Claude 3.7 Sonnet hybrid leasing model (the smartest model to date) and Meta’s Llama 4 series models. We’re also the first cloud provider to fully offer DeepSeek R1 and Mistral AI’s Pixtral Large as fully managed models. Bedrock also includes our own state-of-the-art foundation model, Amazon Nova.

8. Agents: We launched Amazon Nova Sonic, a new voice-to-voice foundation model that enables developers to build highly accurate, expressive, and realistic voice AI applications. For rapid interactions, Nova Sonic has lower word error rates and higher success rates compared to similar models.

So far, almost all agent use cases have been Q&A-based. Our goal is to enable agents to perform broad, complex multi-step tasks, like planning a trip, setting up lighting, temperature, and music for dinner guests, or handling complex IT tasks to improve business efficiency. The technology to build these agents is still quite primitive, not very accurate, and requires ongoing human supervision.
We recently released a research preview of Amazon Nova Act, a new AI model trained to perform actions in web browsers. It allows developers to break down complex workflows into reliable commands, like searching, checking out, or answering questions about on-screen content. Developers can also add more detailed instructions to these commands, such as rejecting insurance upsells. Nova Act aims to improve the accuracy of multi-step agent operations from 30%-60% to over 90% by providing the right building blocks for action-oriented agents.
9. AI Revenue Scale: Our AI business has reached billions in annualized revenue and continues to grow at triple-digit YoY rates, though it’s still in its early stages. But we remind everyone that a significant amount of on-premises infrastructure hasn’t migrated to the cloud yet. While infrastructure may sound less exciting, it’s the foundation of any company’s tech innovation, developer productivity, speed, and cost structure. To fully realize AI’s potential, they need to move their infrastructure and data to the cloud.
10. Alexa+, the Next-Gen Alexa Personal Assistant: Prime members get it for free; non-Prime members can access it for $19.99/month. We’ve just started rolling it out in the U.S. and will expand to other countries later this year.

11. "Project Kuiper" reached a major milestone as we launched our first satellites into orbit. More satellites will follow soon, and we expect to begin customer service later this year.

2.2. Financial Commentary and Guidance:

1. Global revenue was $155.7 billion, up 10% YoY excluding FX impacts. FX headwinds negatively impacted revenue by $1.4 billion this quarter. Global operating income was $18.4 billion, about $400 million above the high end of our guidance range. This performance includes one-time charges affecting North America and International operating income, which we’ll discuss later.

2. This quarter, we recorded one-time charges related to historical customer returns and costs from preemptively receiving inventory due to anticipated tariffs. Excluding these charges, operating margins for North America and International would have been ~90 bps and ~70 bps higher, respectively—7.2% for North America and 3.7% for International.

3. AWS’s annualized revenue now exceeds $117 billion. In Q1, as enterprises shifted focus to new initiatives, migrated more workloads to the cloud, and restarted or accelerated existing cloud migrations—while also leveraging generative AI—we saw growth in both generative and non-generative AI products.

4. AWS operating income was $11.5 billion, reflecting our continued growth and focus on driving efficiency across the business. We expect AWS operating margins to fluctuate over time, partly due to varying investment levels. We plan to add more capacity in the second half of the year.

5. Cash CapEx was $24.3 billion in Q1. Most of this spending went toward meeting growing demand for tech infrastructure, primarily related to AWS to support AI services. We’re also increasingly investing in custom chips like Trainium, as well as tech infrastructure for our North America and International segments. We’re investing in fulfillment and transportation networks to support future growth, improve delivery speed, and enhance cost structures. These investments will support business growth for many years.

6. Our guidance considers various scenarios, incorporating Q1 performance, current quarter trends, and macroeconomic expectations. We expect Q2 net sales of $159-$164 billion. FX is expected to be a ~10 bps headwind in Q2. Note that global currencies may fluctuate this quarter. We expect Q2 operating income of $13-$17.5 billion. This estimate includes seasonal increases in stock-based comp due to the annual compensation cycle.

2.3. Q&A with Analysts

Q: Since February, the company seems to have introduced more GPU instances to support all these new AI workloads. How would you describe the supply-demand imbalance for AI workloads? When will AWS start seeing enough AI revenue to drive business acceleration? Could it happen this year, or are capacity constraints more likely to ease next year?

A: We’ve been adding a lot of P5 GPUs (a type of Nvidia chip) while deploying Trainium2 as quickly as possible. Our AI business has reached billions in annualized revenue and is growing at triple-digit YoY rates. The capacity we’ve added is being consumed almost immediately. With more capacity, we could serve more customers and drive more revenue. In the coming months, we’ll have more Trainium2 and next-gen Nvidia chips coming online.

I think some other parts of the supply chain, like motherboards and components, are also somewhat congested due to strong demand. But I do believe supply chain and capacity issues will gradually improve as the year progresses.

Q: Strategically, given potential changes in the global trade environment and uncertainties, how should the company position itself for the medium term? What key priorities will help the company remain profitable under different outcomes? How are you prioritizing investments over the next few months?

A: It’s hard to predict tariffs right now—what they’ll be or when they’ll be finalized. So, our short- and medium-term focus aligns with our long-term focus: how to offer customers the widest selection at the lowest possible prices. In recent memory, keeping prices low has never been more important, and we’re obsessively focused on this, along with fast delivery and great service. That’s our core.

We’ve taken various steps to support these priorities. As a first-party seller, we’ve made some advance inventory purchases. Our third-party sellers have also stocked up early, so they have inventory too. We encourage this to help keep prices low for customers.

I think our 2M+ sellers won’t all react the same way if tariffs rise. Some will pass costs to consumers, but others won’t. This diversity gives us a better chance of some sellers choosing to gain market share without passing on tariffs.

Compared to the past, we’ve meaningfully diversified our sourcing for AWS and device components beyond China. We think this is wise. These are some ways we’re trying to protect customers in the short, medium, and long term.

Q: On the Q2 EBIT guidance, are there any one-time costs or tariff-related costs like the $1B you mentioned in Q1?

A: Stock-based comp is higher—this follows historical patterns. There are also additional Project Kuiper launch costs. Specifically, we’re paying tariffs on retail purchases under current rates. Q2 tariff impacts are small since we did a lot of advance inventory purchases in Q1. But given uncertainties, we’ve widened our guidance range. We see some positives in April, like advance purchases and strong ad growth, but uncertainties remain.

Q: On AWS, what’s the backlog situation? Regarding core workloads not yet migrated to the cloud, what are you seeing? Is AI causing any confusion about migration timing?

A: Q1 backlog was $189B, up ~20% YoY, with a weighted average remaining term of 4.1 years.

On unmigrated workloads: Pre-pandemic, enterprises were methodically moving to the cloud for innovation, developer productivity, and cost benefits. During the pandemic, everyone optimized costs. Then, as that trend eased, generative AI emerged, and everyone wanted to run AI workloads given their potential and public excitement.

So over the past ~16-18 months, enterprises realized they need to do two things at once: adopt AI (with various pilot projects underway) and resume cloud migrations. But migrations take years—some faster, some slower, but always phased and careful. Apps can’t break during migration. We’ve had great success helping enterprises modernize their infrastructure on AWS.

End of Text

Risk Disclosures & Disclaimers: Dolphin Research Disclaimer & General Disclosures