We're on the cusp of a white collar blood bath. This is not a everyone's not going to have a job and just robots are going to be running the world.
Does that mean you're not going to hire anymore?
There's lots of jobs and in general the economy is much bigger and on average everyone is better off.
If you did decide to build kind of a competitive top tier frontier model, the relationship with Anthropic or with Meta would change.
From the very beginning, we knew that people would be building on top of us. Sometimes those would be in conflict.
I want to just touch on the Annapurna acquisition, because I think maybe it's slightly underappreciated outside of
How much would you say of AWS's code is written by AI right now?
Over 80% of our developers are using AI in their workflows.
Do you recommend engineering as a career?
If you spend all of your time learning one specific thing and you're like, that's the thing I'm going to be expert at for the next 30 years, that's not going to be valuable 30 years from now.
Matt, thank you so much for joining me. I appreciate it.
Yeah, thanks for having me on.
Yeah, absolutely. So I wanted to talk first about the future of work and white collar work. There's a large spectrum of folks who either believe we're on the cusp of a white collar blood bath all the way to utopia. Where do you fall on that spectrum and why?
Yeah, I think in the technology world there's never been a more exciting time to be in technology and I think across industry I'm on the much more optimistic side. I'll start there. I think a bunch of the advances that we've seen in AI and technology have an enormous potential to increase the efficiency and effectiveness and enablement in work. But I think a lot of where the technology promises to do is take away the toil of what a lot of people do in their day-to-day jobs. And I think if you look at a lot of white collar jobs today, the vast majority of things that people spend their time doing is not the things that they would ever tell you they get excited about for their job. It's not putting my numbers into this particular system or pulling the report together so that I can report on the thing that I'm doing or trying to collate a bunch of information so that I can actually understand what it is that I want to work on. That's a bunch of the work today that takes up a large percentage of the time and so my view is that AI can really help with that. I think people would love to spend more of their time on that creative aspect, on driving the aspect, on doing interesting analysis, on doing the parts of their job that they love and not the parts that they don't. And so today though it's necessary. It's a necessary part of almost every job. Your job, my job, lots of other people's jobs. And if we can shrink that part down and give people more time to focus on the things that they're excited about, it drives value for companies. It drives value for people, I think people will enjoy their jobs a lot more. So, I'm very optimistic that this is not a, you know, everyone's not going to have a job and just robots are going to be running the world. I just think that's probably not the place that we're going to end up. I'm much more optimistic that companies can be more efficient, people can be more efficient, and people can spend more time on the things that they're excited about.
So, I'm also optimistic. So, I'm very glad to hear you say that. And you know, I think a lot of people think, okay, if you're automating large swaths of people's tasks, day-to-day tasks in their job, you can do so much more as a company, but then does that mean you're not going to hire anymore? Like, how do you think about hiring in this future where every single person on your team is so much more productive?
Yeah, I think it's going to depend. I don't know that there's any one particular answer for any particular company. But look, if you look over the history of industry, there's been massive amounts of efficiency gains. And in every single one of those transitions, there's transition time. And I think one of the more critical things is for people to be flexible, to be willing to learn, and be willing to accept that their job may be a little bit different, and the components that they do of their job may be a little bit different, and that the job that they had two years ago is not necessarily the job exactly. The components of their job won't be the one that they have two years from now. But there's not like there's mass unemployment in the world today because we have computers and because we have automation and because we have robotics. There's lots of jobs. There's actually people that have higher paying jobs. And in general the economy is much bigger and on average everyone is better off. And so I think that that's how you think about it is that not that you know there used to be you know the silly examples are you know lots of people used to spend lots and lots and lots of their time doing calculations. Now that we have Excel it doesn't mean that a bunch of people are out of jobs because there's like a math engine doing all of the work. You just assume yeah that's the tool that does it. And I think that is a good analogy. AI is a very different technology. It's much more disruptive than something like Excel. But it's an analogy where that was a job. There's automation that came into play and now people do more higher value work and I think that that is exactly what's going to be true today. And it's a role that I'm excited about but it does mean that people are going to you know I'm not minimizing that uncertainty that people will have and I know people are worried about it and I think one of the things is embrace that technology. The more you can embrace that technology, be flexible, understand how it can help you do your job faster and better, I think the better off people are going to be as they make that transition because it is going to be a really powerful technology for how it and I tell our teams all the time, I think AI has the potential to transform every single industry, every single company, and every single job. But it doesn't mean they go away. It has transformed them, not replaced them. And so it really is every single one of those. If you don't lean in and take advantage of that technology and learn that technology, then you might be out of a job actually if you don't really think about it and lean into it. But if you do and realize how it can make your job better or faster or you better or faster at your job and frankly push you to do more of what you might like to do, I think that's a world where companies are going to be better off, people will be better off and there will be economic growth.
Yeah. And there's been a number of historical examples of world changing technology. And I think the pessimists will point to well the speed of change that AI is affecting right now is very different from previous technological shifts, right? So do you think the speed is going to really affect the white collar market? Do you think that's the differentiating factor compared to previous technological shifts?
I think the speed is, it is a really rapidly evolving technology space and people are just going to have to move faster with it and so that is a challenge. People are going to have to move with it and embrace that. I'll use an example for your audience from a developer audience. You know, developers I've heard from developers worried about these coding tools are somehow going to make developers not need it anymore. And what I would say is my view is we're going to need more software developers, not fewer software developers. The job of a software developer may be a little bit different. I don't know that the archetype of a software developer that stays in a basement and codes by themselves for weeks on end. I don't know if that's the job, but we're going to need somebody who really knows software. We're going to need lots of people who know software where they might have agents that can actually go and do some of the coding. In fact, I do think the part of the job that you probably are not going to have to do two or three years from now is authoring Java code. That is probably not a job that's going to exist because these tools are going to be really good at authoring Java code. Deconstructing a problem, deciding what to go in and build, pulling it together, looking at the Java code that comes back and deciding it's not quite exactly what you want and you have it go and have agents go do that, coordinating a bunch of agents, that is going to be more a job that a software developer is. And that person is going to be able to drive so much more value because they're enabled. Like how many times have you been frustrated at the speed of your own brain and how fast it is to actually do something? You can think of ideas, you want to go do, you want to actually put them into action and it just takes a long time to do stuff. If you can kind of unlock that ability for creative people to go solve problems and develop interesting applications or interesting results, the amount of the value of having somebody that's good at that is much higher. Why would you want to get rid of those people? You're very excited about having more people that can do that. Now you have to have people that can transition into thinking about how they use these tools. But my experience is that most people once they get over that initial hump actually quite get excited because it is really enabling. It does actually enable you to create and go build faster.
Yeah. Probably also the mindset of a business leader like yourself. You're looking at somebody and all of a sudden they are 5x, 10x more productive. The last thing you're
I want to get rid of that person.
I need less of that type of no. You want to invest even more heavily because the ROI on that dollar comes back 10-fold.
You got it. So, it's just I don't understand the math of why we want fewer of those people. So, yes.
Okay. I love the optimistic view. Let's continue on engineering specifically. How much would you say of AWS's code is written by AI right now? And maybe define what written by AI is because that can change depending on who you're talking to. I look I think for a while for the last year or two people have been excited about bragging about the number of lines of code that have been written by AI. It's a silly metric like okay I don't like I bet it can actually write infinitely more lines of code. They might be bad code by the way but if like that it's not really that lines of code.
Yeah, measuring lines of code has never actually been the best metric. Often times fewer lines of code is way better than more lines of code. So I'm never really sure why that's the exciting metric that people like to brag about. For us, I would say the last metric I saw was something like over 80% of our developers are using AI in their workflows in some way, right? And sometimes it's writing unit tests, sometimes it's helping write documentation, sometimes it's writing code, sometimes it's an agentic workflow where we have something called QCLI or our Nuro IDE that allows them to actually have agentic things going and writing and collaborating with them. And that number is going up every single week. So, I admit I may not have the latest number, but it was north of 80% of our developers are using AI in some way, shape, or form to go build whatever they're building for Amazon customers.
Do you find that the engineers are proactively seeking how to upskill themselves and use these new tools or do you also have educational processes in place and programs that help them?
Yeah. It's both. I mean, Amazon is a very large company and so we don't have a homogeneous set of employees. I think the vast majority of Amazon developers are curious though and so I think the number of them who have not used an AI coding tool is zero or rounds to zero. Figuring out how many are really using it to completely transform their job just to use it as some of the pieces. I think that that is where education is also important because I think there is some number of our employees that have really dived in head first and actually understood what works because there is some learning curve there. It's the kind of understanding how you change work. What are the things that you do that can actually really accelerate versus slow you down. There's plenty of examples where you can be using some of these first generation coding tools and you kind of get down a rabbit hole and it's actually quite hard to go back because it's like actually a very linear process where the tools are kind of when you're doing your vibe coding and they're giving you the code and you're like that's not what I want to do. There's not like an easy way to go back to here. It's like this is now the code you have and so you have to explain this the other day.
Yeah. With Curo and part of what we've done in agentic coding first kind of mentality is that you actually start with a spec of the thing that you want to build and then you work with the tool to actually go and build parts of that spec and as you're vibe coding you it can automatically change parts of that spec but you still have that spec as the core thing you can always go back to and change aspects of it or functions of it or whatever it is. And we have seen the light bulb go on because it actually one of the cool things about that is that you can actually guide more junior developers as to like what are great coding practices, how do we think about this? And so one of the things I actually am very passionate about I was at a leadership group and people were telling me they're like we think that with AI we can replace all of our junior people in our company. I was like that's the dumbest thing I've ever heard. They're probably the least expensive employees you have. They're the most leaned into your AI tools and like how's that going to work when you go 10 years in the future and you have no one that has built up or learned anything. And so it's you know I my view is like you absolutely want to keep hiring kids out of college and teaching them the right ways to go build software and decompose problems and think about it just as much as you ever have. And so, part of the benefit of some tools like Curo is it actually can guide you into some best practices and kind of really learning how you construct and actually work together with maybe some of your more seasoned engineers who can kind of teach the right ways of going and building some of these systems.
Yeah. I mean, one of my questions was going to be, do you recommend engineering as a career to somebody who is entering college? I think I know the answer to that.
Yeah. My take is, you know, my oldest is going into senior year of high school and first of all I think the kids should go and study whatever they're passionate about and learn. But what I would say is you know sometimes a lot of the skills that I think there that should be emphasized is how do you think for yourself? How do you develop critical reasoning of solving problems? How do you develop creativity? How do you develop a learning mindset that you're going to go learn to do the next thing? Because with the pace at technology that's advancing right now, if you spend all of your time learning one specific thing and you're like, okay, that's the thing I'm going to be expert at for the next 30 years. The most thing I can promise you is that's probably not the thing you're going to be like, that's not going to be valuable 30 years from now. But if you learn how to learn or you learn how to think, I think that's actually where schools can be incredibly valuable. So I think engineering is really great not at teaching you the blocking and tackling of the specific thing but really kind of how do you think and how do you decompose problems and I think that's where kids coming out of school if they really focus on that then you're going to be in great shape.
And so when you look internally at your engineering organization, how do you measure success with respect to using artificial intelligence?
Yeah, I won't pretend that we have that magic measure today. I don't know that it's any different than it was before. Some of its productivity. We are still trying to learn best practices as the technologies are changing and so we encourage experimentation. We encourage people to be testing different tools, different methodologies. We're also experimenting with different setups. Like sometimes, you know, one of the interesting things is that when you used to have to go build a really large complicated system, you needed a lot of people because everybody could focus on their one little piece.
Yeah. One of the cool enabling things about this technology is you can have a broader vision or a scope over what you're driving. And so what we're finding is a lot of times we would prefer to have more smaller teams that are focused on problems. And actually you can kind of decompose these products into areas where there's really these focused teams that really go after something and that small team can operate really fast. I know you were at a startup before. One of the benefits that startups have over big enterprises is you can move fast. Well, it's not like an inherent thing that you have to organize in a different way. Part of what we're learning is these tools enable you that even in a large organization, you can organize in small pods of people that can also move really fast. And if you can move fast and the tools enable you to move fast, you can go deliver outstanding things for our customers. And we think as even a relatively large business, that's going to enable everybody to move faster.
Okay. So let's switch topics for a second. I want to talk about capacity and infrastructure buildout. When you are thinking about the next two to five years from now, what are the bottlenecks for your infrastructure buildout? Is it the silicon? Is it the energy? What is it?
Yeah. You know, when I was in business school, we read a book called The Goal. Did you ever read The Goal?
Okay. It's an old book. But the idea is when you're optimizing a production line, there is not like one bottleneck. There is a bottleneck until it's solved and then the next thing is a bottleneck. And in many ways like that is this thing like so all of those are bottlenecks at some point. And I don't think that there's one thing I guess you Yeah. And there will just be a next one. And so you have to solve all of those things. I don't think there's one that's like okay if we have an infinite number of chips there's no problem because you will need power. And if you have an infinite number of power and you don't have enough networking devices, that's problem. And so you have to have all of those pieces. I think one of the challenges is the challenges and opportunities for us as an infrastructure provider and a cloud provider is the opportunity is massive. And so there's a lot of need for compute. There's a lot of need for electricity. There's a lot of need for servers and chips and HPM memory and networking devices and discs and all of those pieces. And so, the shortage for sure for the last couple of years has been in Nvidia chips and chips in general. And I think as that eases, you're going to find that it may be power and it may be chips. Again, it's really hard to get all of these things exactly in sync to the growth. Because a lot of these things require lots of capital to go invest in. So, our job is to think about all of those. And so, we think one year out, 3 years out, 5 years out, how do we make sure that our customers, our business and our customers have enough power to support the things they're looking for? How do they have enough access to a broad swath of capacity that they have access for? How do they have network connectivity that they're going to need? And that requires us to think about what are the amounts of capacity that people are going to need that many years out. Sometimes we're wrong and sometimes we don't have enough, but that's part of what we try to take on for customers so everybody doesn't have to think about all of those things.
Yeah, I'm sure part of planning the infrastructure buildout is understanding where the usage is coming from. And so what are you seeing in terms of growth? Is it reinforcement learning? Is it inference? Is it training? Where is the most growth coming from?
Well, I think most of the growth today is at the end of the day it comes from inference, right? And so it's customers using the end product and so everybody likes to talk like it's fun to learn about all these new techniques of model building and whether it's reinforcement learning or fine-tuning but the vast majority of the usage actually comes from the end users, right? It is the inference of the end result and then there's a bunch of pieces that go into how do you create better and better models that are cheaper and more efficient to use, more capable, more cost effective. All of those different pieces and that is in that other bucket of great we have a bunch of startups that are very focused in this space. We have a bunch of larger companies that are thinking about how do I create custom models and put a bunch of models together. But the vast majority of compute is frankly driven by that once it's all put together, customer comes in and asks a question or interacts with an application or tries to run a workflow or something like that. And that's what's driving the most of that usage. But we're focused on all of those pieces. And I think there's, from an infrastructure or power perspective, it doesn't really matter. Like the computer doesn't care if you're using it for reinforcement learning or for fine-tuning or for inference. Frankly, most of these platforms are actually converging. There's some differences around networking that you might need for big pipelines, but the same silicon can serve all of these different needs.
Yeah. Absolutely. Interesting.
