This is an interview about AI. My fiancé works at Anthropic. See my full ethics disclosure here.

Few people have a better vantage point on the AI economy than Matt Garman. He joined Amazon as an MBA intern in 2005, before Amazon Web Services even launched, and went on to become one of the first product managers for EC2 — the service that kicked off the cloud computing era. In June 2024 he became CEO of AWS, now a roughly $130 billion-a-year business that sits underneath much of the AI boom. It is the primary training partner for Anthropic, has signed compute deals with OpenAI worth up to $100 billion, and this year plans to spend some $200 billion on capital expenditures, the bulk of it on AI infrastructure.

Amazon is also increasingly in the business of selling software that does white-collar work: a developer agent, a security agent, an "agentic teammate" suite called Q Business, and an AI recruiter, Amazon Connect Talent, that schedules and conducts voice interviews with no human in the loop.

Which makes it something of a surprise that Garman doesn't worry about AI destroying huge numbers of jobs. In fact, though, he is among the louder voices arguing that jobs might change, but they're not going away.

Last year Garman called replacing junior employees with AI "one of the dumbest things I've ever heard," and told Wired that never hiring junior people is a "non-starter for anyone trying to build a long-term company." When I pressed him on Anthropic CEO Dario Amodei's prediction that AI could wipe out half of all entry-level white-collar jobs within five years, he was unmoved. "Wipe out" and "change" are different things, he told me. Excel may have eliminated the jobs of people who calculated by hand, but those people learned to use computers — and the labor force expanded overall.

Amazon, he noted, is hiring 11,000 interns and new grads this year.

But what struck me most during our conversation was the tension Garman embodies. Where our previous guest, Replika and Wabi founder Eugenia Kuyda, told us the fear of job loss is "super justified," Garman might be the most full-throated optimist we've featured in the whole series.

It's an optimism that doesn't always seem to square with Amazon's own moves. The company has cut roughly 30,000 corporate jobs since last October. Its CEO, Andy Jassy, has written that AI will "reduce our total corporate workforce" in the years ahead. Among other strategies, the company plans to replace half a million jobs with robots. Garman is talking up the value of labor even as he sells the technology that may displace them.

In our conversation, he attempted to square that circle with a familiar argument: that technology destroys individual jobs but creates new and better ones, and that the most durable skill is a simple willingness to learn. "If you look at what your job was two years ago, and what your job is going to be in two years, it's going to be vastly different," he told me.

This marks the end of our first miniseries for the Platformer podcast. Thanks to everyone who read, listened, or watched, and wrote in with your thoughts. After a short break, we'll be back with a second series that attempts to take some of the questions raised here in new directions. And then you can expect Platformer to get back to more text-based news and analysis, as I shift my podcasting energy from Hard Fork to a new show with Kevin Roose.

In the meantime, though, I can't think of a better concluding episode than this one.

Highlights of our conversation are below, edited for clarity and length. Listen to the entire conversation wherever you get your podcasts — just search for Platformer — or watch it on YouTube at youtube.com/caseynewton.

And let us know what you think — we're new to podcast production, and welcome your feedback at casey@platformer.news.

Casey Newton: You joined Amazon as an MBA intern in 2005, before AWS even launched, when maybe some saw it as a strange side project for a bookstore. Twenty years later, it's a $130 billion business. So it strikes me that you've already lived through a cycle of people initially underestimating something that turned out to be really big, but also maybe people at various times during that transition overestimating how quickly it would happen. As you're looking at the AI landscape, does any of that history rhyme for you, or does this just feel like a totally different moment?

Matt Garman: It's 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. We had to explain this new way of thinking about building applications, where you did it differently. You didn't think about how you have one big SAN and a couple of servers that you took really good care of and then slowly scaled. Rather, you had this disposable infrastructure that was serverless, where you didn't really worry about the underlying pieces of it and focused on software and scaling.

I view a lot of what's happening right now similarly. There are some differences too, which I'll call out, but in similar ways I do think AI is going to completely change how people build applications. It's not just doing the exact same things you were doing before but a little bit faster or a little bit cheaper. It's fundamentally changing business processes, it's changing workflows, it's changing business outcomes and customer outcomes and customer experiences.

And it's moving so much faster than that last change. Through the first five, 10 years, we grew what we thought was really, really rapidly. But we looked up five years, 10 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-premises environments.

I don't think AI is going to take that long to transform many businesses. When I talk to customers, and really everyone out there, that's probably the biggest difference — the speed at which this transition is happening. It's obviously a totally different technology, and it has really wide-ranging impacts across a lot of different industries. Even slow-moving industries are thinking, this is not something that I have 20 years to get ready for. This is something I have 20 months to get ready for, maybe. Maybe not that long, but it's something that's going to happen much faster.

Newton: Why do you think it's happening faster? I imagine some of it is just businesses feeling like there's a lot of competitive pressure to go fast, but is there something else in there I should think about?

Garman: 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 was still running in their own data centers, and you had to install the new update to the next model, and then you had to scale your own GPU capacity in order to run it, and you had to deal with network bottlenecks — that just was never going to happen. But now, because of the cloud, and because enough of your workloads and enough of your data is in a cloud world, and these models are available in a cloud world, that cycle is that much faster. Those things compound upon each other and have led to it happening at a much faster rate than a previous shift.

Newton: I was excited to talk to you because of this unique vantage point you have running AWS. Basically every company experimenting with AI is running through you. You have millions of customers, every major model is on Bedrock, and you can see what the enterprises are actually doing. Sometimes we hear enterprises have run a bunch of early experiments and they aren't seeing a huge return on that investment. Other companies seem like they're starting to figure it out. Give us a sense of what you're seeing in that range.

Garman: Almost everyone ran a huge number of proof-of-concepts. That was the common thing everyone did two or three years ago — they just told everybody, "Go see what this thing does. It's some magical new technology, we don't really know what it does, go experiment." So 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, across the board, 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 — an interesting chatbot they built, or an interesting content generation thing, or an interesting workflow — was cool and it worked, but it wasn't actually driving real business value, because appropriately people had to learn about the technology first.

Now what we're seeing is companies saying: We understand directionally where this technology could go. Now, where can I get that return, and how do I roll that into production? It turns out those aren't the same thing. Even if the proof of concept was directionally where they want to go, now you have to think about data security. How do I think about governance of who has access to this data? How do I make sure about security of these agents — where they can go and where they can't go? How do I think about compliance? There are a bunch of actual regulations and rules and operating pieces you have to think about. And how do I think about real business value? What is the cost of this thing? What's the cost savings or revenue increase because I rolled it out?

I was talking to a room full of CIOs just a couple of months ago, and I raised [a question]. Show of hands — there were probably about a hundred people in the room. "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?" Ninety percent 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. And by the way, it's the first couple of percentage points of that transition. But people are starting to see that return to the business.

There are plenty of other proof-of-concepts where they're shutting them down because they don't actually see that. So you have just as many — and probably way more — where people are shutting them down and not seeing those returns. But they're really doubling down on the ones where they see that benefit.

Newton: The obvious place where people are seeing returns is in coding, software engineering. It seems pretty clear there's just a ton of value there, and the labs are selling tokens as fast as they can make them. But because you're looking across the entire industry, I'm curious if there are other sectors where you're starting to think, oh, these folks have really figured it out, they cracked something.

Garman: Coding is a clear and obvious one, and I actually still think we're in the early stages of where that's going to evolve to. The whole software development lifecycle is seeing massive improvements in efficiency and gains.

But if you take that half a step further, those gains in software development are enabling agents. That reasoning — the ability to write code and to get work done — is then enabling agents to autonomously go do business-process things. We're starting to see people roll this out in telcos thinking about network optimization, or financial services companies thinking about loan processing, and it goes on and on in industry-specific pieces.

You're starting to see workflows where it was successful 20% of the time, now 80% of the time, now high-90s percent of the time. And now it's: Okay, what can I extend it to? It used to be a five-step process; now it's a 30-step process I can ask it to do. You're starting to see those real line-of-business pieces. I was just talking to a large insurer the other day, and they were saying, "I used to have this backlog of 60 to 90 days of insurance claims processing, where people would just wait. Now I can process them so much more quickly, and people get their answer back and get paid faster. And now people are coming to us and we're winning new business, because they're seeing we're actually great across customer support, able to answer questions quickly when people have issues."

It's a great example where agents are all about both driving efficiencies, where there's an inefficiency in the process, but actually driving top-line revenue because of that, where people are able to dive in where you need them to. Agents are able to make things more efficient, and customers benefit.

Newton: A big topic this year as AI has spread has been the amount companies are investing in capital expenditures. Amazon has committed $200 billion this year, about a 50% increase from last year. Explain to me how you all think about how much to invest year by year. What's giving you the confidence that the demand is going to be waiting for you once all that gets built?

Garman: We spend a lot of time as a leadership team thinking about this, so it's not a random number we pick out of a hat, as you might expect for numbers that large. And it's driven by a lot of data. We get a ton of data from customers. What are they looking at? What are they looking to spend? Which workloads are they getting to move? So we have an enormous amount of data, both about what customers are doing now and what their pipeline looks like. At this point we're working deeply with customers at a strategic level to think about what they want for the next one, two, three, five years. So we have a pretty good view as to what customers are looking to do on us — and I think it's probably an undersell of what they're going to want to do.

Amazon is inherently an operations business. You order something, it shows up at your door the next day. That is a very complicated operations process: How do you make sure you have the right inventory? How do you think about inventory turns? How do you make sure it's loaded on the truck at the right time and gets to your house? We took that ethos at AWS. In many parts of AWS, it is an operations problem. How do we think about supply chain? How do we think about long-term having enough land and data centers and power, which are oftentimes three-year-out, five-year-out investments?

Those are investments where we have a lot less certainty as to what demand is going to be that far out. But they're also two-way-door investments in some ways. I can buy a bunch of land, and maybe the land goes down 10%, but if in five years I don't need it, I can resell it. It's a durable asset. I view power in somewhat similar ways. The chances of the world needing less electricity in five years, or 10, is so small. Maybe we invested in the wrong way, where we spent X dollars per kilowatt-hour and it should have been X minus Y, but it's not a wasted investment. So those long-term investments we feel pretty good about. And even in the worst case that we don't need them, there are ways we can offload them to others that do.

Then, as we get closer in — think about a server. We have a pretty good idea of what our demand is going to be three to six months out, or whenever we have to commit to server parts or chips or memory. And a lot of times customers will give us three- or five-year commitments to use that capacity, which also minimizes that risk. So we take a portfolio approach. Our business model is that we're going to have this upfront capital that we invest, and it's been true for the whole 20 years of AWS: the faster we grow, the more the upfront capital investment is. But we really like the returns on that investment. Someone was telling me recently, if you really like the return on invested capital of a business, you want the "C" to be as high as possible. And for us, that's a little bit of where we're at right now. It's not speculative — we have a lot of mitigations in there, and we really think intentionally about how we can reduce risk. But we love the ROI of the business, and if you like that, you want the C to be big.

Newton: Let's talk about jobs. You called replacing junior employees with AI "one of the dumbest things I've ever heard," and told Wired that never hiring junior people is a "non-starter for anyone trying to build a long-term company." There are other views in the industry. Dario Amodei, the CEO of Anthropic, has said AI could wipe out half of entry-level white-collar jobs in five years. What's he missing?

Garman: Here's the thing. There are subtle differences in there that I think are important. I do think half of white-collar jobs may change, but it doesn't mean "wipe out." "Wipe out" and "change" are different. Excel wiped out all of the jobs of people who were hand-calculating things — but those people then learned how to use a computer, and there you go, now they had a job again. 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