Why Amazon is hiring 11,000 junior employees

Thank you. Appreciate you having me here.

Yeah. No, it's actually a really good analogy and it's exactly that. If you roll back 20 years ago when we were first starting AWS, we had to explain to people what the cloud was and why Amazon was a part of that. It was a completely different shift into how work was going to get done. We had to explain that it used to take six months to get a server and now you could get one in five minutes, and if something fails, you can just shut it down and launch another one. That's not how the world worked. So we'd explain this new way of thinking about building applications where you did it differently. You didn't think about having one big server and a couple of servers that you took really good care of and then slowly scaled, but rather you had this disposable infrastructure that was serverless. You didn't worry about the underlying pieces and focused on software and scaling. I view a lot of what's happening right now similarly. I think there are some differences, too, but in a similar way, I do think AI is going to completely change how people build applications. It's not just doing the exact same things a little faster or cheaper. It's fundamentally changing business processes, workflows, business outcomes, customer outcomes, and what was possible. And it's moving so much faster than that last change. It took people through the first five, ten years, and we grew what we thought was really rapidly, but we looked up five, ten years into AWS and still today the vast majority of workloads still run on prem. Even 20 years later, there's a massive amount of workloads that still run in on-premise environments. I don't think AI is going to take that long to transform businesses. I think that's probably the biggest difference: the speed at which this transition is happening. It's a totally different technology, but it has wide-ranging impacts across many industries. Even slow-moving industries are thinking this is not something they have 20 years to get ready for. This is something they have 20 months to get ready for, maybe that long, maybe not, but it's going to happen much faster.

I do actually think these two technology shifts that we've been intimately involved with are quite complementary and compounding of each other. Without the cloud, AI doesn't take off in the same way. If everybody's still running in their own data centers and you had to install the new update to the next model, scale your own GPU capacity, deal with network bottlenecks, that just was never going to happen. But now because of the cloud, enough workloads and data are in a cloud world, and these models are available in a cloud world, that cycle is much faster. So I think those things compound upon each other and have led to it happening at a much faster rate than a previous shift.

Yeah. I would say almost everyone ran a huge number of proof of concepts. That was the common thing two or three years ago. They told everybody, 'Go see what this thing does. It's some magical new technology. We don't really know what it does. Go experiment.' Some people did, and people built a lot of really cool things. They didn't always have a good idea of what they wanted to get out of those experiments. They just wanted to see what the technology could do. So not surprisingly, most of those experiments didn't show great returns because they didn't actually have a plan for what they were going to have coming out the other side. The proof of concept of an interesting chatbot or content generation thing or workflow was cool and it worked, but it wasn't driving real business value because people had to learn about the technology first. Now, we're seeing that we understand directionally where this technology could go. Enterprises and companies are saying, 'Where can I get that return? How do I roll that into production?' It turns out that's not the same thing. Even if the proof of concept was directionally where they want to go, now you have to think about data security, governance, security of agents, compliance, regulations, and real business value. What is the cost? What is the cost savings or revenue increase? Interestingly, that's when we had this hypothesis that when customers started going to production, they would really want this to be in AWS because that's where their data lives, their production applications live, their security policies are, and they trust AWS with their data. They trust us to define guardrails and hold the model inside that. So security, operational excellence, and the existence of all their production applications is when customers said, 'Great, now that I've done proof of concepts everywhere, this is where I want to deploy them.' That's the cycle we're in now. Customers are deploying those workloads into AWS and seeing real business value. I was talking to a room full of CIOs a couple months ago. I asked, 'Raise a show of hands. How many of you either today are seeing materially positive ROI or have a path in the next couple of months to really high ROI on your AI investments?' 90% of hands went up. That's totally different from a year before when they were like, 'No, it's just a cost model for me.' People are starting to see that return. It's the first couple of percentage points of that transition, but they're starting to see real value. There are plenty of other proof of concepts where they're shutting them down because they don't see that. They're doubling down on the ones where they see benefit.

Yeah. I think coding is a clear and obvious one, and I still think we're in the early stages of where that's going to evolve. The whole software development life cycle is seeing massive improvements and efficiency gains. The other ones, if you take that half a step further, those gains in software development are enabling agents to autonomously go do business processes. We're starting to see people roll this out in telcos thinking about network optimization, financial services companies thinking about loan processing, and it goes on in industry-specific pieces. You're starting to see workflows that were successful 20% of the time, now 80%, now high 90s percent successful. Now it's like, 'What can I extend it to?' It used to be a five-step process, now it's a 30-step process. You're seeing real line of business pieces where they say, 'All of a sudden, I used to have a backlog of 60 to 90 days of insurance claims processing where people would just wait. Now I can process them much more quickly, people get their answer faster, they get paid faster, and we're winning new business because we're great at customer support and answering questions quickly.' It's a great example where agents are about driving efficiencies and actually driving top-line revenue.

I will say we spend a lot of time as a leadership team thinking about this. It's not a random number we pick out of the hat. It's driven by a lot of data. We get a ton of data from customers: what they're looking at, what they're looking to spend, what workloads they're moving. We have an enormous amount of data on what customers are doing now and what their pipeline looks like. We're working deeply with customers as a strategic goal to think about what they want for the next one, two, three, five years. So we have a pretty good view of what customers are looking to do on us. I think it's probably an undersell of what they're going to want to do, but we have a good view. Then we come from Amazon, which is inherently an operations business. You order something, it shows up at your door the next day. That's a complicated operations process of inventory, inventory turns, loading on the truck at the right time. That ethos we took into AWS, making parts of AWS an operations problem: supply chain, long-term land, data centers, power, which are often three-year, five-year investments. Those are further out, with less certainty on demand, but they're two-way door investments. I can buy land, and if the land goes down 10%, but if I don't need it in five years, I can resell it. It's a durable asset. Power is similar. The chances of the world needing less electricity in five or ten years is so small. Even if we invested in the wrong way, it's not wasted investment. Those long-term investments we feel good about. Even in the worst case, we can offload them to others. As we get closer in, like a server, we have a good idea of demand three to six months out. Customers often give us three- or five-year commitments, which minimizes risk. We take a portfolio approach. Our business model is that we have upfront capital investment. It's been true for 20 years of AWS: the faster we grow, the more upfront capital investment, but we like the returns. Someone told me recently, if you like the ROIC of a business, you want the C to be as high as possible. That's where we are. It's not speculative. We have mitigations and guardrails, and we think intentionally about reducing risk. But we love ROI, and if you like that, you want the seed to be big. That's how we think about it.

Yeah. I think the world is not all homogeneous. We have more data than a random investment company that's going to go build data centers. We have more certainty of customers. We have a pipeline of customers interested in purchasing from us. For people going to build a cloud from scratch today and invest $10 million hoping customers show up, that's the first place I think demand will go away from. There may be demand for all these places, but for us, we have a lot more. We've built up the customer base. We have that flywheel. We have their production applications and data. I don't think the JP Morgans, Netflixes, or Airbnbs are going away. We have a broad, diverse set of customers building inference into their applications, making them better and delivering positive returns. I have a hard time thinking they'll stop spending on things with positive return. For us, the more speculative things that happen only in a GPU cloud somewhere, maybe those continue forever. And by the way, if you're a startup with a billion-dollar valuation, maybe you turn into a $10 million

Well, look, I think there's a broad swath of things that people are appropriately worried about. I think they're worried about, you know, I have a big ugly building next to me that I didn't used to want to have, or is it going to be really loud, or is it going to be bad for the environment, or is it going to... And actually, one of the most common things you hear now is, is it going to make my electricity rates go up, because it's a natural conclusion. It's not right, by the way, but it is a natural conclusion to say, if there are 100 units of energy, and they used to cost a dollar, and now somebody comes in and says, I want 50 units of energy, supply and demand says the price goes up. Practically speaking, it's not quite how it works, and actually, we've because there's a number of things, price actually depends on your peak-to-average load. So, actually, if you bring in more average load, which data centers bring in a really stable average load, it actually shrinks your peak-to-average, and actually reduces the cost of an incremental unit of electricity to a consumer, if you're operating on that same grid. And so, in fact, we have one of our very largest data center implantations or construction projects is in Indiana. We've talked about Project Rainier as we've talked about publicly. It's a very large project. And we've announced with the governor there that we actually reduced the average cost of energy for the citizens of Indiana because of this project. So their cost of energy actually went down because of that. And I think that's, you know, it's harder for people to understand that because you have to understand actually how energy is charged and things like that, but that is what it is. And we've committed, you know, with the administration that we'll pay for all of the power that we, we'll make sure that we bring power, we'll make sure that we bring cost up, upgrade grids when necessary. There's a bunch of those things that we've committed to. We just announced this week actually from an environmental perspective that we use 7x less water in our data centers than most data centers do, and we're well on our path to being water positive by 2030. So from an environmental perspective, people worry about it appropriately so. It's a big worry, and we're very committed to carbon zero and water positive impacts on the environment. And you know, I think look, a lot of the places where we operate, we bring a lot of really high-paying jobs. And so in a world where people are excited about high-paying jobs, where they want some of these skilled labor, where they want high-paying, they and communities that are looking for more employment, they actually quite welcome us and like to have us there. So, you know, are we going to put a data center inside of the city limits of Seattle? Probably not, but that's not really where you put them anyway. And so, I think that's okay, and I think you have to be thoughtful about where you put them and how you impact the communities. And not everyone is as thoughtful as we are, by the way. And so, there's another problem where everybody kind of gets swept up if there's a couple of bad actors who maybe don't follow some of the environmental regulations or don't work with local communities, but we spend a lot of time making sure that we do that.

You know, look, I think there will be advocacy groups that will say things all of the time. That's true for now and will be true forever. Our view is, look, I do think there's a risk of not everyone necessarily acting in that same way and thinking about the community, and I think that could cause problems for everyone. And so I think it's important that everybody actually takes these paths and goes towards carbon zero and goes towards water positive and thinks about their impact to the community and makes sure they pay great wages. So, I think that's a super important thing, and we would advocate for everybody doing that. Even if it adds a little bit of cost to your operating, I think it's important. You know, look, I think there's, I haven't yet talked to the consumer who wants to turn off their Netflix and doesn't want to use Airbnb and doesn't want to call an Uber and refuses to use Cloud or ChatGPT or doesn't want to search on Google. At some point, you can't have your cake and eat it too, and those things are all powered by data centers. Most of them use AI to accomplish what they're doing. And so, you know, I think there is kind of a reality there where there's no, I don't want any data centers, but I still want all of the things that come out of data centers. That's not really a reasonable trade-off, and at some point people have to make that call and realize that.

Well, here's the thing that I think that there's subtle differences in there that I think are important. I do think that half of white-collar jobs may change, but it doesn't mean wipe out, and change are different, right? And so, you know, Excel wiped out all of the jobs of people who are hand calculating things, but those people then learned how to use a computer, and there you go. They had a job again. And so, I think that the key thing is not to look at a still picture of the world and say that job's not going to exist, so I guess those people won't have jobs. New jobs will be created, and I firmly believe this fact because I will say that, you know, if you believe that half of jobs get wiped out, the whole economy collapses on itself and everything goes away, and then you're not going to have AI, and then you have to go back to those other jobs at some point. The math doesn't work out. But more than that, these jobs, the AI is a new technology that is inventing new things, and we're already seeing new jobs get created. They're different jobs. And so, what I tell people at Amazon is, look, there are going to be lots of jobs. And you talked about entry-level jobs. Number one, they're your cheapest employees. They haven't learned bad habits. You can teach them the culture. They're willing to learn the new tools. They're some of the very best employees you can possibly have. But I tell all of our employees, if you look at what your job was 2 years ago, and you look at what your job is going to be in 2 years, it's going to be vastly different. You're going to have a job. You're going to have probably a more exciting job and an interesting job, but you're going to have to be willing to learn. And I actually think it's one of the things we start to look for in employees is not what are the skill sets you have, but do you have the ability to learn? Do you have the willingness to dive in and learn new things and the agility to reason about problems, to think about how do we solve customer problems, and how do I apply that to the capabilities that we have? And I think when you have employees like that that have that mentality that I want to learn, I want to lean in and do new things, there's a crazy amount of opportunity in front of us to run really fast, build new businesses, deliver value to our customers, reduce costs. All of those things are huge opportunities. And I think I find, and there's a reason we're hiring 11,000 interns and new college grads this year at Amazon. Because they come in with an energy and excitement, a new view on things, and if you just have the exact same people that you've had for the last 15 years, you don't get that energy and excitement and new ideas. And so, I think there's a number of reasons why personally I find that logic to be faulty. But it is a little bit of a view. It is, you know, if you look at the exact snapshot of the jobs that existed 10 years ago, some of those may go away. But they're going to be replaced by hopefully better and more interesting and exciting jobs.

Close. I think we probably didn't exactly hit it, but I think we're directionally on that path. You know, I think, 24 months ago, and this is a good example of this. We have more software developers today than we had two years ago. Two years ago, one of the best skills that you could have is authoring a really great line of Java code, right? I know the syntax, I know how it works, I know how to write Java code that will compile. When it doesn't compile, I know how to go find the issues and rewrite it, etc. Most developers and engineers at Amazon are not doing that anymore. There is a quickly diminishing number of people are still doing that. And there are still some types of code where it's still mostly done by human, but it's more of the type of job as opposed to the person. And increasingly, the AI models are getting much better at some of these other types of coding, too. And so, we're not all the way there yet, but I think we're well down that path, and developers are more productive than ever, but you still have to understand how software works. You still have to understand how systems are put together. It's not just, go write S3 and the AI goes and writes S3. We're really far away from that. I can't imagine when that happens, right? But every one of our software developers is coding by agents, and they're directing agents, and they're understanding kind of what the agents are great at, and they're starting to build AI-native code that's all run by agents, the tests are done by agents, the unit tests are run there, the patching is done that way, and you're thinking of more, how do I have multiple agents that I'm directing, when they get stuck, when they want to do something else, when they want to guide on the user experience, how do they do that? And actually more of the work is being done at the early stages on what's the right thing to build. How do I architect this to perform well? How do I, are the customers really going to like this experience? And you spend more time on that and less time on troubleshooting authoring code. And yet we're not 100% there yet, but well on the path.

Not unrecognizable, but I think look, I think these jobs, I think all of our jobs are going to keep morphing over the next couple of years. And hopefully for the good. I mean look, today for example, you know, there's still a number of things in the software development life cycle that are not fully automated and we're working through how do you make those more efficient? In fact, if you go talk to a lot of AI native development teams in Amazon and other places, they'll tell you they're starting to spend the smallest amount of their time coding because the AI tools make that great. They're spending more of their time thinking about the product. They're also now spending more of their time like babysitting deployments because the deployment systems haven't necessarily caught up. And so we're actually thinking about how do you make sure that you go fix those and make those even smoother, and now you think about how you think about security patching? How do you think about a number of these other parts of that whole life cycle? And once you can get, you know, you start to make sure that all of those bottlenecks get through the pipeline, because you want to spend frankly as much time as possible on that early stage, which is thinking about the product, thinking about what's the right thing to build, thinking about how it interacts with the rest of what you do. And then getting feedback from customers. And actually one of the nice changes is that move. And I don't know if it's unrecognizable because it's the same end process, but it's just focusing on more different parts of that, spending more time with your end customers and understanding that they like your product. Are they using the thing that you built? And getting that feedback back so that you can feed it back in that first part of the stage. And I think that middle part does kind of shrink up and be, you know, sometimes the smallest part.

Well, sure, but except for that, you can create that narrative without the details. The details were like these were roles that were removed where either their businesses that we weren't interested in being in anymore. It's not that AI got rid of that business. It was just an area that we had grown up and it didn't take off and that was a space that we didn't need. A lot of what we removed was we really tried to flatten the organization, and this goes to the culture organization where we didn't want managers, managers, managers, but if you could flatten those organizations so that you can push decision-making down to the edges, you can move faster, you can be closer to customers, and it does remove some headcount when you do that. That's not because of AI. AI doesn't do any of those things. I mean, it allows you to be effective in those things, but it's not that all of a sudden AI is being that manager role. Instead, we're saying from an organizational perspective, we want our engineering teams and our product teams to be closer and we want there to be more span and less layers. And that's what we did. So, you know, there was no AI doing that efficiency. That was just an organizational and frankly we said, 'Look, we may do less, but we actually don't think so. We actually think with smaller teams we'll move faster. We actually think with faster decision-making we'll move faster, and we'll actually end up doing more.' And I think that's what we've seen. We said we'd take the risk that we would do less to have the right culture, but it hasn't proven out. We've actually seen that our productivity and outputs have improved because of the speed of decision-making, not because of any magic of AI. I think similarly, though, AI is now making those teams that are close, there's a nice flywheel effect where when those teams are closer to customers, they can iterate much more quickly because as I talked about, that middle part is much less part of the work now. It's define the product, get feedback from customers, iterate on that product, and you have to be close to those customers at the end to really understand how they're using the product to make that work. And so there's a nice flywheel benefit there that those teams being closer to customers allows you to actually iterate faster.

Yeah. And I would say, look, this is something that Amazon has been doing for a really long time. We automate work so that our employees can start to do higher value work and we reduce costs for customers. But this is not a new thing. This is something that we've literally been doing for probably 25 years since we started the company. And so, in fact, it's something that we push our employees to think about. How do you automate parts of your work so that you can have broader scale, so you can get better leverage on what you're doing, and do more. And so, you know, we don't pack boxes in the same exact way that we did it 20 years ago. There's a lot of efficiencies, and we think about how do you organize differently, how do you think about supply chain differently, how do you think about robotics differently. We think about the software development life cycle, we think about how do you automate parts of this job so that you don't have to do it. This has always been part of what we do. Now, AI is a great new tool that's allowing us to do more here, but this has always been true. And so, you know, if you think about those things that you just said from a recruiting perspective, our recruiters would prefer to really handhold candidates once they're through that pipeline, answer some of those specific questions that they have, go out and really source great talent, or think about that once they're like inputting all your details and things like that is not what makes a recruiter's day. It's not what they get excited about. It's not why they got into that job. And if you can automate that away, now we can get more talent into the pipeline and get better leverage, and we'll find other things for those folks to do. So, it absolutely is automating away jobs and replacing them with new things to do.

Our recruiters are some of my most important partners. We're recruiting every single day, and I would much prefer our recruiters to focus on going and finding great talent than doing the blocking and tackling of inputting details and making sure that they get the email of what time their interview is going to be. Those are things they don't need to spend time on.

Yeah. I don't know the answer to that other than every time I've actually dug into this, I see us hiring more, right? The current one, the biggest people are worried about is maybe there won't be software development jobs. There are more software developers being hired this year than ever before in history. So the evidence doesn't point to it. And you know, you can go through every single job that exists in the world and that'll take us a long time, probably longer than we have today. But I'm not going to know those answers other than history has shown us that this is the case. And if we're really creating value, then that is creating more opportunity to go do more.

Yeah, I think the most important thing is to make sure that you actually measure the thing that you want to measure. And so we never intended to have those leaderboards that way. It really was a way of pushing people to say I want you to do this AI native piece. You measure in the wrong way and people figure out a way around the measurement as opposed to the goal. So it was never the goal for us. I think that's actually a good rule for everybody that you get what you measure. And so you just got to make sure that you measure. And for us, it's not that we, you know, I've seen people who are like pulling back and saying, 'Ooh, I'm not sure AI native coding is worth it.' To me that is bonkers. It is the biggest efficiency gain that I've ever seen. I'm more than happy for our competitors to make those choices, but if they want to. But for us, that's not it. But the thing you actually want to measure is not can you use the most tokens? The only person that wins there is the person selling tokens. You really want to measure who's getting great output, right? Are you having more code check-ins? Are you pushing more features to customers? Are you closing bugs faster? All of those kind of things. I think those are the right measures, and so we, it wasn't like we were like, 'Ooh, we don't want to do this measuring of tokens anymore.' We just realized we had the wrong measure in place. And just quickly changed to measuring some other outputs. And I think others had this kind of at a much more extreme than we did, because ours was never quite, I think others actually did have that as the output. Ours was never really that. It was really more just kind of measuring to make sure that people were using the AI tools in a good way. But I think that's right. And it's a good rule in general, measure what you want to get.

I think it's like at some point it's just a resource and it's a cost, and I think people want to, like you with anything that costs money and is a resource, you want people to be thoughtful owners. And actually one of the leadership principles we have at Amazon is act like an owner. And so we want all of our employees to think like an owner. And if they think the investment is good, then we give people some guardrails and things like that. But generally, if they're thinking about this is a good investment, we don't want them wasting money. And in the same way that we think about running EC2 instances, it's not really that materially different. People didn't launch 10,000 EC2 instances just to see if they could, because it cost them money and they would see the bill. But we give employees the ability to go launch EC2 instances if they need them. And so in the same way, we want our employees to be using AI to make themselves more efficient and effective and deliver more value to our customers. And so we encourage that. And we're building tools to help them with that. Like I think one of the things that's driven up a bunch of cost is that people have kind of naively just said, 'Great, I'm going to use the best model for every single thing.' And it turns out that you don't need Opus 4.8 to generate code. The Sonnet model or even the Haiku model does a really good job at that. Now, when you're going to plan and do some of those higher reasoning things, using Opus 4.8 is a really good model for that. And so actually in Kiro, which is our Amazon coding tool, we do a lot of this for customers where we pick the right model and then we appropriately use our amount of your token budget and things like that. And so we use cheaper models when we can that are fast and they're sometimes faster, they're less expensive, but they accomplish the results. And so that's a good way of balancing some of those things where if we can give tools to customers where they don't have to build a lot of this complicated logic, but they can just get the benefits of those. I think that's one of the things that helps. But for us, we also kind of want our employees to think like owners and to know, and the more you can expose those costs, I think it enables people to do that.

Yeah. I think it's actually a similar story, and I've been pleasantly surprised with how many places that I've been able to get leverage to myself. And as you might expect, one of the most rare resources for me is time. And so if I can save time from preparing for a customer conversation or looking through some updates from teams or thinking about business metrics and can get great summaries, that allows me more time to have more customer discussions, right? And when I have a customer discussion, that period is very highly leveraged time for me where I can sit down with 30 minutes, talk to a CEO who's interested in the cloud and learning about our services. It's really leveraged time. If I have to spend 30 minutes preparing for that meeting, that costs an hour. If I can spend 5 minutes preparing for that meeting, I can actually fit more of those in a day. And for me that's really highly leveraged. Amazon Quick is a product that I've been using really extensively recently to do this. And part of the win that I've seen is what Quick does, it actually builds context across all of your communications and all of your data sources inside of an enterprise. So it can read my emails and it's on my desktop. And so it can read my email, it can read my Slack, it can look into Salesforce, it can look into our sales numbers, it can look into our business metrics, it can look into our wikis and everything on our internet. And when I ask it questions, it can go do deep research, it can understand all of those things, it can look into my email and actually get, instead of me scanning email forever finding where those five messages are that I want to pull together so that I could actually put this coherent thought together, you just ask Quick to do it. Come back 5 minutes later, you have a great summary. This is the five emails that I pulled it from. Here are the attachments. It really is game-changing because all of that work I consider to be non-differentiated because it's how do I get the information? And then I can think about it, right? Now I can think about okay, these are business trade-offs. I can think about this is how I can have a robust conversation with a customer because I know what they care about. I know what services they're interested in. I know what their business challenges are. I didn't have to spend an hour preparing for that discussion. I've found it to be really impactful for me.

Yeah, I think the path is, there's a couple of things I think. Really long reasoning tasks are things that AI is getting better at, but the more steps you put in a workflow, the more deviations there are in the workflow, the worse they often perform. And so that's one of the things that I am anxious for. I think there's a lot of opportunity and a big area that we're investing in is technologies that can layer on top of LLMs and AI to actually provide assurance for workflows that have to be right. And I think that's an enormous opportunity for many of our customers where, you know, it's an interesting trade-off where the non-deterministic nature of LLMs is what makes them so powerful. And it means they're like, but how do I know it's going to be right? And so we've actually we have a bunch of researchers in this space that I think are leading the industry in this where if you can layer on some actually provably correct algorithms on top of that, we actually think we can get the best of both worlds there. Those are some areas that are early, but I'm quite excited about.

That's exactly right. I mean, that's a great example of where I think we can get.

Yeah, thank you for having me. I appreciate it. It's fun.