Alexa+ Will Be Biggest, Best Use of Generative AI: Amazon CEO Full Interview

Well, you know, we've had Alexa around for ten years. We have 600 million devices in customers' homes and offices. Alexa Plus is our next generation Alexa personal assistant, and she's meaningfully smarter, more capable, and more useful than her prior self. You can do all the things that you used to do, but every single one of those functions is better. I can give you just an example. There are so many examples, but if you have smart home controls with Alexa, now you can say, 'Alexa, I have guests coming over at 7 PM. Could you raise the drapes? Could you raise the temperature by five degrees? Turn on the porch lights, the driveway lights, and put on mellow dinner music.' You can do that all verbally and simply, using conversational language like that. You don't need an app; it just happens. And so Alexa Plus, with what we've just announced and what we're launching, really is the first one that not only is highly intelligent to answer various questions, but she can do so many things for you—play music, play video, control your smart home, make reservations, hire people to fix your oven. It really is the first big, large-scale practical use of generative AI by consumers that they're going to be able to see and use.

Well, I think if you think about what Alexa allows customers to do, it makes shopping easier because it's so much more intuitive to buy products. It makes enjoying music easier, enjoying video and streaming media better. It allows you to control your smart home in a different way. So every single one of our consumer customer experiences gets better with Alexa Plus. And then, of course, Alexa has its own business model. We have a brand new lineup of devices coming in the fall that are beautiful. I think people are going to really like them. We have opportunities for new products and advertising in various interfaces like our mobile and desktop interfaces coming in Alexa. And then we have subscriptions. So I think there's a sustainable business model there as well.

Prime is an incredible value. If you think about getting free shipping on 300 million plus items, we launched Prime with free shipping on about a million items. Today, it's 300 million items, and most of the time you're getting your products now inside of a day. So between that and what you get with Prime Video, Prime Music, the grocery subscription, and our unique selling events, Prime is an incredible value. Adding Alexa Plus on top of that is just great value. So we don't have any plan right now, but I do think Prime is unusual value, and it's why people use it so expansively.

Well, we don't disclose the exact amount, but we're spending a pretty significant amount of CapEx, and the lion's share of it is on generative AI. We've said in our business, even though generative AI for us is a many billion dollar a year business and growing triple digit percentages year over year, that if we had even more capacity, we could use it to monetize it. And we have this very interesting and fortunate flywheel in AI inside of Amazon: if your mission is to make customers' lives easier and better every day, and if you believe that all the customer experiences we know of today are going to be reinvented through generative AI, you're going to be building a lot of generative AI apps. By the way, other companies are too, on top of it. If you ask which is the leading technology infrastructure platform, if there are a lot of generative AI apps being built on top of you, you can't help but get a lot of feedback from people on how they want those building blocks to be better. And if you're willing to invest in those building blocks, which we are, with our own chip Trainium, our own frontier model Amazon Nova, and model building services like SageMaker and Bedrock, and you're getting a lot of feedback, they get better much more quickly, which makes it easier and quicker for people to build generative AI applications, meaning you get more running on the platform. That flywheel is very unique for Amazon.

Most of it, the lion's share, is more than 50%.

We're playing the warmer and colder game!

It's hard to put an exact percentage, but I do think it could be growing faster. I'm confident it could be growing faster. For a long time, there still aren't as many chips as we all want. We're fortunate in that we're very big partners with NVIDIA, but we also have our own custom silicon, Trainium, which we just released and it's 30 to 40% more price performance than the CPU power instances, which is a big deal at scale if you're doing generative AI and on the inference side. So we have maybe more chips than some others might have access to, but we still don't have enough. And there's just not enough power in the world right now, and we're all working really hard on that. I expect that to relieve some in the second half of this year. But right now, we have insatiable demand.

Well, there's a lot of demand for generative AI right now. People are very excited about it. I think all the different providers of chips have been constrained to some extent. Some of the new generations maybe have gone through different evolutions, and when they're going to be released maybe a little later than people thought. There are some components like motherboards that we all use that are in shorter supply than others. So there are some supply chain issues which I expect to get better. People are very excited about Trainium, and we have gone back at least a couple of occasions to make more Trainium than we had intended because we have so much demand. So I expect we'll have customers for as long as I can foresee wanting to run compute on instances with NVIDIA chips, but I think a lot of the demand will also be served by Trainium.

It's always a combination. There are people that just have a very small amount of accelerators they need; they don't usually have a problem with something called capacity blocks, which is kind of like an on-demand way to use accelerators and generative AI chips, and that continues to grow. It's really the folks who have an idea for a new application but need a lot of chips. If we don't have the capacity, we have to give them what we can and push our partners to get it in sooner, and they can't get their initiatives done as quickly as they want because the capacity isn't there.

We have a very close partnership with Anthropic. We have this project called Project Rainier with them. They're building their next version of their frontier model on top of Trainium and our customized silicon. They're going to use over 400,000 Trainium chips. So yes, they have capacity; they're ramping up, and we're excited about that partnership.

We have been talking to administrations, multiple administrations in this country over the years, as well as in other countries. The power shortage really snuck up on people right after the pandemic. The current administration is very receptive to it; they understand the constraints it's having on the economy and are convicted about solving it.

I think we share the concern. I'm going to be curious to see where the AI diffusion act goes. It was enacted pretty quickly at the very end of the last administration. I don't know how this administration feels about it, but it has limitations on certain countries who are natural allies of the US. For their companies to be able to get done what they want on top of technology infrastructure platforms like ours, they're going to need more chips. If we don't do it, we're going to basically give up that business and those relationships to other countries who can provide those chips. I think we're better off being partners with them.

In the scheme of things, it's not a big swinger. But there are so many countries in the early stages of their economic development who need access to the most cutting-edge sophisticated technology to build the right customer experiences. That could be big geographic markets for companies like ours, and I think it would be a shame to limit them.

Regarding de minimis specifically, we have a certain number of items shipped that way for things like Haul, which is our new low-price offering. We maybe have lots of it, and some other companies like the ones you mentioned. But it's early in this administration. What is encouraging to us is that we have an administration that wants to hear from business. We've been a business through six administrations; every single one, our primary focus is to take care of customers, but we try to build a productive relationship with the administration because we want to help the country. Some administrations are more receptive than others, but this administration cares about what business thinks. I've always been surprised that it isn't obvious that the best economic results for a country are when the public and private sector collaborate. I don't expect the government to kowtow to what companies want, but they should get feedback and input because they'll make policies better together. And I'm encouraged early on that this administration wants to talk to businesses.

I take calls. We talk to people in the administration; we share what's working for us, what's not, concerns we have. As I said, some administrations care more about our feedback than others.

I have spoken to the president. This administration has been pretty busy the first month, but I am encouraged that they are having conversations with businesses and do care about our feedback. We'll see what happens, but it starts with a dialogue.

As we talked about earlier, in terms of the run rate on CapEx, we're spending a lot on infrastructure, and the lion's share of it is not just on AI but also in this country. In the US, we spend elsewhere because we have a global business. We have customers in a couple hundred countries. We have a pretty substantial investment there. I don't expect it to attenuate soon.

At a high level, if you serve as many customers as we do in as many diverse groups as we do, you have to have a diverse team to build products that work for everybody. That has always been and continues to be our intention. There were so many programs we launched and other companies launched during the pandemic. As you've seen over the last three years, we've gone through very thoroughly every single one of our business areas and the programs we had conviction about, we doubled down on. The programs we don't have conviction about, we streamlined and stopped. We did the same thing looking at all our programs on diversity. Some programs we've doubled down on—a good example is our Career Choice program inside our fulfillment network. Our fulfillment center teammates can get an advanced education for free to advance their career. That has been very successful, so we doubled down on that. Other programs just haven't been that successful and haven't moved the needle much, so we moved away from them. We have a diverse group; we're trying to continue to build that out, and that won't change.

You can call it lots of different things, but we have a giant customer base of every imaginable group of people, and we want builders who can build for them.

It's going well. If you have a company where the culture is an important ingredient in your success, which is absolutely true for Amazon, it's not your birthright to keep having a strong culture, especially as you grow the number of people, businesses, and geographies. You have to work at it all the time. When we looked at it as a leadership team last year, there were two areas we wanted to strengthen. One was that we have always hired strong owners who get to own the lion's share of decisions—90-plus percent. But as you add a lot of people, you end up with a lot of middle managers. Those middle managers, well-intentioned, want to put their fingerprint on everything, so you end up with people being in the pre-meeting for the pre-meeting for the decision meeting, not making recommendations the way we want. So we took a goal to increase the ratio of individual contributors to managers by over 15% across the company by the end of this quarter. We've made very good progress and beat that already. The people doing the work will have more ownership and be able to move more quickly. The other thing we saw is that if you're a culture that invents and collaborates a lot, if you don't have people in the office together doing that invention, it's meaningfully worse. You don't collaborate, connect, or learn the culture the same way. So having people back in the office together more frequently, we felt very strongly, will be better for customers and the business.

Quantum computing is very high potential. It has the chance to solve very computationally intense problems, but I still think it's realistically a few years away from having a real shot at solving those problems. You have to solve a bunch of challenges along the way, one of which is error correction on the qubits. That's what Ocelot does; it's a very unique, inventive way to do error correction that makes a meaningful difference. We're excited about that milestone. You can't get to something with real impact unless you get those milestones done and invent along the way.

I don't know for sure. I'm hopeful that it's more in the five-year range than the 20-year range. All these things that are successful are seven, ten, 20-year overnight successes. If you look at generative AI, it's just another evolution of AI, but we've been working on AI for 50-some-odd years, and then it just boom happened. It really shocked us when it actually was more accessible and worked. The same thing could happen with quantum computing—it takes a long time, then all of a sudden it's functional and solves problems you couldn't solve before, and it feels like it happened overnight. But we've been working on quantum computing for ten-plus years.

I don't know. It's funny, I have this feeling that some people will feel a little disillusioned about generative AI because of the investment and whether they're getting the commensurate return. It's still very early; many companies are really doing pilots of their generative AI applications. But the smart companies will figure out which initiatives can really change their customer experience and businesses and keep investing. The slower companies will wait to see if it's safe to go outside and will be behind by two or three years, maybe more. Because even more so than software development generally, AI is very iterative. In software development, you can design something and largely it works. But in generative AI, the models get better at a disproportionate rate, sometimes failing or discovering a new scaling law. If you actually aren't investing in generative AI, you're going to be behind by even more than the time you waited, because there are so many lessons from iterating and building applications today.

What we care most about in all our businesses, but as it relates to AI, we want our customers to be able to change their customer experiences and improve their businesses so they can last over a long period successfully. If we do right by our customers and take the long-term approach, and they're able to run their applications successfully on top of our technology infrastructure services, then they're successful and we ride along with them. We have services we build ourselves in the models area—Amazon Nova, which people are really excited about because it's got comparable intelligence to the leading models but is meaningfully less expensive and has lower latency. But if customers prefer other models, we have partnerships with Anthropic, Llama, DeepSeek, Mistral. We have the largest collection of leading foundation models in the world. If customers are having success with those, we're happy. The truth is, if you build a lot of generative AI applications, you use multiple model types often in the same application. Even Alexa Plus, as we talked about, uses multiple foundation models. We want people to use the right model for their applications, and we make it easy for them to switch between them and run it successfully.