J.P. Morgan Presents ACQUIRED at AWS re:Invent (Matt Garman interview)

How you going, Matt? Good to see you. Thanks for being here. Nice to see you.

No, just been hanging out. I was out by the pool.

No. How many of these media things have you done so far this week? I've done a bunch. But it's awesome. It's such a great opportunity. It's so awesome to have all of our customers here. And as I think I just said in the video, and hopefully you guys have experienced the energy and excitement that we have when you get this many people together, not just to learn about AWS but they learn about our partners. They learn from each other. It's a cool opportunity. It's been fun to build, and this is our 14th year. I remember the first year here, I think we had two or three thousand people, and I thought it was crazy that two or three thousand people would get together to talk about our technology. And we've grown since then.

I distinctly remember when we launched AWS, we went to some conference, I can't even remember which one, but we had one of those booths, saying hey, we're starting a new thing called Amazon Web Services. People were like, why would I buy computers from a book seller? They were very confused for a while, and we eventually got them over that.

Yeah, it probably wasn't going to be AI and agents necessarily. But I will say it may have been AI, because I think we were at the time, even in 2022, if you go back another 10 years maybe not, but in 2022 we were very excited about the potential of AI and saw where it was going to go. Do I think it was going to be as it is right now? Of course not. I don't think anyone would have projected exactly that. But it wouldn't have surprised me that it was a main focus of what we were doing. We already had SageMaker. We already had a bunch of the things that we were building. I probably would have guessed it was bigger than it would have been in 2022, that it was a bigger focus for people today, because we saw some of the early machine learning benefits of the technology, but nothing like it. I think there are some other areas that I would have expected that have actually grown faster that haven't. They'll get there, but I would have thought robotics would have grown faster than they have today. It's gone a little bit slower than I would have thought. In fact, we launched a service early called RoboMaker that was just ahead of its time and didn't take off because the robotics weren't ready for it. But I think that's some of it. I will say that the challenges our customers are facing today are remarkably consistent. We often say today less than 15 to 20% of workloads have actually moved to the cloud from on-prem. If you looked at Monday, we blew up a bunch of servers, if you saw that video. That's all about getting rid of tech debt and helping people migrate off mainframes and Windows and VMware. People back in 2022 definitely wanted to get off those, and they have to some extent, but there's still a bunch there.

Yes, they are. And years from now, some people were saying, 'Do you think next year when you come here, everybody will have moved to AI?' I was like, 'No, no.' Only 20% of people have cloud after 20 years. For those whose data is still locked in a mainframe, they're definitely not jumping in and getting value out of AI yet.

It's not going to be 2026. I'll tell you that. There is a long runway ahead of us for sure.

I actually think that question is getting harder and harder to answer over time. I can tell you that AI and Bedrock is a multi-billion dollar business. We sell GPUs. But I think that undersells the impact. If you have an application like Workday that has AI assistants and agent tech workflows, is that an AI application? Does all of Workday count? Do you just want the GPUs they're running? What counts as AI? If everybody's storing stuff in S3, some of it is vector embeddings, but some is referencing their objects, driving them to get more objects into S3. Is that an AI workload or not? I introduced this concept of inference as a building block last year, and I think it's more true than ever. There are AI applications like Perplexity, but increasingly every application will just have AI in it. There's not a database application or a storage application; everything is just a database with a wrapper on top. When you think about Chase Payments, is that a database application? No, it's an interesting payments application. Is it going to have AI in it? Yes. Is it an AI application? I don't know. Today it's a multi-billion dollar business for us. It's a huge tailwind, not just for model running, but all those pieces. People are driving to the cloud faster to get value from AI. Their applications are getting more usage, so they need more CPUs. We're using more Graviton. All of those pieces.

Yeah, we'll workshop it live.

I don't know. I haven't thought about that. I would say there are companies that are all-in on AWS, or on the cloud generally. They have their town, and then ones moving the majority of their workloads. There's a continuum. I don't know that an AI-native company's relationship with us is materially different than that of Netflix. They're 100% in on AWS. They have been for many years. When they need features, they talk to us. They're great partners. We learn a ton from them. That's true for AI companies and Netflix. Could Netflix go build data centers? I guess so probably, but we have a great partnership and they haven't needed to. Even when you look at foundational model companies, they need cloud. They're not building their own. But I don't know if they need it. Some large companies like OpenAI have talked about building Stargate. They can build data centers. Are they going to run them efficiently? Is it a good use of their capital? Those are open questions. It's possible, but I don't know why they would. They get a lot of benefits from the cloud, especially from customers also being in the cloud. That's really powerful, particularly when you move from consumer applications to enterprise applications. You get benefits of scale and technology, and your customers are in that same cloud, so you can deploy models into their VPC. There's a flywheel where enterprises trust AWS with data security. We've built that trust over 20 years, so a new startup with an exciting model can put it on Bedrock and access that trust level.

The good news is not many people are going to do that going forward. That's true.

They're not deployed. Okay. We'll talk afterwards and we'll get you an account manager.

I would say yes and no. There are a couple of pieces. I would challenge you to build an interesting AI app without using compute, storage, and databases. It turns out all those things are still needed. When you have the front end of a chatbot, CPUs are the most effective way of rendering that web page. You need storage. All those components are the same. I remember when mobile first started getting big, people asked if there are different servers needed. The answer was no. There were some front-end development tools that were different, but you still needed compute, storage, and databases. AI is different in that I think inference is a new developer building block that never existed before. The world invented a new Lego. You had all your Legos and could build whatever, and now you have a cool new Lego. You can build incredible things. You still need all the other Legos, but now you can build super interesting things. Agents are an extension of that. You're still going to need Lambda as an execution engine for agent workflows. How you go to market might change, how you think about user experience might change. I think org structures are going to change. If you're starting a company, I don't know that the org structure we've used for 20 years is the right one. You can deliver with much smaller teams. A huge effort that used to take hundreds of people may take ten or five people, delivering just as much and faster. We were in an event with the CEO of Cursor a couple months ago. It's a tiny team, maybe 100 people now. You just need tiny teams. The roles may change. You may not have a traditional dev manager, but a dev lead. The whole org structure may change. I don't know the answers, but I think those things will change.