The Amazon CTO returns to its strategy in terms of artificial intelligence. He deciphers the impact of this technology on customer relations and the group’s application developments.
JDN. How does AI integrate with Amazon long-term vision, beyond the Cloud AWS?
Werner Vogels. With 300 million products, more than 200 million members prime and 9 billion packages delivered daily, we had to automate a maximum of processes. There is what is seen from the point of view of the customer. This is the case of Rufus, a tool you can ask for example: “My nine-year-old daughter is celebrating her birthday next week, we are in France, the weather is nice, what should I bring as a gift?” And the tool will offer you adapted product recommendations. We also use AI to identify counterfeits. Last year, we detected about 15 million ads that sellers wanted to sell on our marketplace but who turned out to be linked to counterfeit products.
And beyond the customer experience?
It is in the supply chain that most of our AI technologies are used. This concerns four areas: supply forecasts, the identification of suppliers, the optimization of the flow of products, and all the work necessary to keep our promises, in particular in terms of delivery. Large -scale forecasts are always a challenge. For example, when marketing the new version of a television, the supplier will probably lower the price of the previous model. There is the elasticity of prices, some products sell faster than others, some items are regional, there is seasonality … AI must take into account all these elements to determine which products we should have in stock.
Then, we must decide in which of our 175 warehouses deliver these products. Preparation, sorting and transport are widely automated thanks to robotics. We have 750,000 robots with maneuver in our distribution centers.
So you use robotics a lot?
Yes, we recently presented one of our new robots which has a sense of touch, which allows it to perform complex and sensitive tasks. It is mainly used to improve safety in our warehouses. We have seen the number of incidents decrease by around 35% from one year to the next thanks to the increased use of robots.
We also use AI to decide the best packaging for each product. We have created a packaging decision -making engine that combines different types of AI to minimize the quantity of packaging. We also score each product before it gets out of our warehouses to check the expiration dates or any damage, all of this in an automated manner.
What about delivery?
When I started at Amazon, delivery spanned 5 days for a price of 25 dollars. Now we deliver in 15 minutes, in an hour or in one day depending on the country. In the Netherlands or in France, you receive the packages within 24 hours, in Dubai in two hours. The intelligent optimization of the routes is crucial because the delivery of the last kilometer can represent more than 50% of the total distribution costs.
We have also developed the automated vehicle inspection. It is like when you rent a car at the airport and go around to take photos at the start, except that our system can take 500 images in 10 seconds and identify exactly the differences between the entry and exit of the vehicle.
So AI is omnipresent at Amazon?
AI is everywhere. I have not even talked about Amazon One digitization for automated boxes, or Amazon Scribe who can transform a handwriting into precise text, or intelligent research in Amazon Prime Video.
With the rapid growth of large models, how does Amazon manage energy consumption and the environmental impact of AI?
We were committed to using 100% renewable energy by 2025. We achieved this objective from 2023. Amazon is the largest group in terms of renewable energy consumption in the world. We have more than 600 wind, geothermal and solar energy projects to ensure that the electricity we use is renewable.
Then we seek to reduce the cost of each operation. We have been investing in our own chips for a long time, such as graviton which is 30 to 40% more efficient than comparable general processors. For AI, we have developed two different chips: trainium for training and infrentia for inference, which considerably reduce energy consumption.
How do you optimize the cost of AI models on your cloud?
On AWS, where you pay for each resource and its duration of use, the cost is a good sustainability indicator. We observe the same with AI. Companies often use the biggest models, but it is extremely important that our customers experience and determine which model is suitable for their needs, because a unique model is not suitable for all use cases.
Some models have 700 billion parameters, other 7 billion. The cost of some can be one penny per billion tokens, while others can cost several dollars for the same number. That’s why Bedrock (The AI IA environment available on Amazon’s cloud, editor’s note), offers a wide variety of models to allow customers to experiment and assess the value for money in Serverless mode.
This is one of your main differentiators compared to your competitors, right?
There are two ways to approach the question: if you focus on your competitors, you do not accompany your customers. However, you must focus on your customers.
Customers are overwhelmed by the number of models that arrive every week, so they are looking for a trusted partner. We want to help them navigate in this offer, new versions, larger or smaller, the different supported languages, culturally adapted or not. This is the objective of Bedrock, and it reflects our approach centered on the customer in a field where we do not know exactly how technology will evolve.
We collaborate with many partners, including French partners like Hugging Face, Mistral and others. Via Bedrock, you can compare Deepseek to Mistral, or even compare Claude’s coding capacities to those of other tools.
Werner Vogels is technological director (Chief Technology Officer) at Amazon. He is one of the pioneers of Amazon’s approach to Cloud Computing. Before joining Amazon in 2004, Werner Vogels was a researcher in systems distributed at Cornell University. He holds a doctorate of the Vrije Universiteit d’Amsterdam and is the author of numerous articles on technologies of systems distributed for business computer science.
