 Welcome back everyone to theCUBE's coverage here at ReInvent. We're on location. I'm John Furrier, host Dave Vellante's in the analyst session with Adams Q&A right now. Wall-to-wall coverage four days. This is day three. Our next guest here, Andy Warfield. He was a distinguished engineer at Amazon S3. Cube alumni from 2014. Going way back, great to see you. Good to see you too, John. Thanks for having me. So we're just talking on camera about Andy's keynote. I mean, Adam's keynote around storage, being reinvented. So my land, walking around the hallways here. Storage is just, S3 is the granddaddy of them all in the Amazon parlance. It's true. Adam was reinventing storage. Big storage announcement. Let's get into it. Tell us about what the news was and how we'll unpack it. Sure, sure. Well, I mean, at this point, S3 is 17 years old. So it is, it's the first AWS service. The announcement this week is for S3 Express One Zone, which is a new S3 offering that is really geared toward low latency for S3 requests. It's about a 10x latency reduction over a regular regional S3 request. And that, and how does that work? How does, what does that impact the customer? I mean, what is that? The, the, I think the interesting thing, if you look broadly at S3 and kind of about S3's history, is when we first launched, it was largely around kind of archival use cases, right? Like sticking storage away, secure, it was cost effective and very durable. And as we move forward, over a lot of the last like five or 10 years, there's been a huge pull for data leaks and it's pushed us to build throughput out on S3, right? And so we see jobs that burst into hundreds of terabytes a second of data pull from S3 and customers keep pulling on the data closer to applications. And so the real impetus for the reinvention with Express One Zone is, customers want really rapid interactions. They don't want to have to copy data out of S3 into some intermediate storage. It's really responsive. They just want to work with it directly. So speed and cost impact. Speed and cost and simplicity around responsiveness of data, right? There's also some discussion around decoupling storage in the new JNA architecture from the compute. Right, right. That's been a big discussion. We heard that from a lot of the ISVs and partners. So storage is still has to, is a central part of JNAI. You can't have data without storage. Absolutely. So that's going to change the game. And we think on CUBE research, Dave Vellante is doing a lot of work on this. How's the storage equation change with data management? Because now you got to move data. Or are you moving data? Just can you relocate it? Moving data is faster. Also, JNAI needs access to data. And then you got policy around that. So the data pipelining, the data equation is going to be tied to storage. How do you see that evolving? Because the storage is more than S3. You got the one Express. What's going on here? Well, okay. Let's talk about it on two sides. There's a performance edge to this and like the data access side, we maybe talk about that separately. The thing that I think is really interesting in here is the thing that you're talking about, which is the customer experience of curating that data. They don't want to think about storage and they absolutely want to have good sound practices around the structure of their data and the governance. And so as customers are looking at generative AI, they don't want to be taking their data out of their data lake and shipping it to some external model, right? They really want to be bringing the model to the data. And that specific property, there's over 700,000 data lakes on S3, right? These customers that have built these data lakes, that have built a whole bunch of really effective tooling for analytics, being able to bridge and take advantage of all that curation and move into generative AI tooling is just a huge catalyst for being able to do new stuff. And what's that catalyst? We'll take us through that journey of moving to the J&I side. Is it hard or is it easy? What's the, is there blockers? What's the- This is a thing that I've learned about a ton over the past couple of years is I've started to work with our ML teams and customers that are doing generative AI. One thing, talking to the folks that are building enormous foundation models and some of the folks that are building smaller scale models is so much of the work that you hear folks talking about when they go to get actionable data, right? Whether it's models or inference is actually pre-training, right? They do so much work just taking the data and getting into a state that is labeled and usable and can be used by the tools. And the data lake customers, I mean you guys, you're showing me this data lake that you've built, like the data lake customers have already invested so much in that that they can just move straight to the tooling, right? It's a big accelerator. And that, Adam, so it's about zero ETL. Yeah, yeah. He's really high on the zero ETL thing. Is it really ETL, zero ETL? I think the, the, the, the, I mean it's an aspirational term for sure, but this ETL idea of having to change the format of the data and often have a human involved in that curation is it's costly. Yeah. It's complex work. I'm impressed with the Databricks event when they introduced Parquet and Iceberg open formats. You can start to see more open formats around the data. That should democratize at least some of the data lake aspects of it. And, you know, Snowflake has their own thing. Databricks has their thing. Everyone's got their data lakes. Is there going to be a challenge at some point to convert that into gen AI? I mean, how does that, how do you guys see that evolving? Because, I mean, it's conceptually clear. I got S3, I got a data lake. Is it magic? Do I just like turn it into gen AI? What are the some of the steps that people have to take to do that? Well, so we're really excited about that stuff both on the data lake side. So the open table formats that you're talking about Iceberg, Houdi and Delta Lake, those are coming up in almost every conversation that I have with customers on data lakes. And a lot of that is built around S3 as the object store. And so there's loads and loads of connectivity on top of that. A lot of the work that we're doing with things like bedrock and bedrock agents is removing that extra last mile of preparing vector embeddings, right? And doing the tight work to really engage your data with the model. I saw the vector embeddings announcement on stage today. I didn't catch it was shipping or not or is it in preview mode? I'm not sure on whether it's on a shift or preview. I'm going to check on that. I was doing an interview right when they announced it so I had to kind of check on that. But that brings up a good point. The whole vector database thing is a hot trend but it's kind of been around for a while. But it's a feature, it's not a company, you know? I mean, people are launching their own vector databases. But the premise is that that should be near the data store. Yes, yes. Because they go hand in hand. Yeah. Is that an architectural lock in? Or is that, or how to, if I have say, Mongo or Amazon or data stacks or everyone's going to have a vector database. Can embeddings work together? Is there interoperability with those embeddings? That's a challenge. I think it's totally a challenge but I think it's a case where, you know, our approach is always around choice. And so with the vector side, whether it's open search with support for vectors or pine cone or any of the other databases that are starting to surface vector, we want to make sure that you can put the vectors where you want them to be and use whichever you want. And I think a lot of those things are turning out to meet inside the catalog. And so you'll see those things meeting in the catalog as an interaction layer and then integrating with the AI tools. Yeah, I was talking to someone, they're like, well, just move the data and just re-vector them. It sounds easier, but if it's a big data set, I mean, how hard is vectoring to do a vector index? I think it depends a lot on the size and scope of the data, right? Yeah. But I think especially with the bedrock integrations around the agents, I think it's pretty fast to get going with your data. All right, so yesterday's keynote, what was the big thing that jumped out at you on the Asma on the storage piece? In terms of the value of the customer, take us through your thoughts on it. You know, the thing that really stood up for me on the storage piece, I mean, I'm tempted to say it's express, but I'm really invested in that one, so I'm so excited that the teams launched it. The story that I loved the most was Pfizer, right? These examples of customers that leaned in heavily on changing their data practice to invest in a data lake and then grew it and are now realizing this agility to go and experiment with stuff like Generative AI, right? To experiment with new things. I think that really speaks for itself in terms of what people are doing. And he's a totally great point. In fact, we had a couple of interviews here from some ecosystem partners, Estrometers One, and they're doing a lot of data engineering. The theme is data engineering, not data science or database administrators, data engineering, where people are actually re-engineering their environments and they're using open source like Airflow is what Stromer uses. So like, these are engineers. They're architecting. Absolutely. So this is a, I mean, it's not a new thing that is happening more frequently now than ever before because engineering is like, okay, what's my object store? I got to figure out S3. How is that set up? How does that work in conjunction to the system I'm building over here? Absolutely. I mean, it's a systems architecture view. It's a systems mindset. It's absolutely a systems mindset and it's neat because it's so tied to the workload and the data, like you say, right? It's on a query level or on a training level. You need to understand the consequences of the structure of the data on the workload. And there's a really interesting aspect on the S3 side of this, which is that the express announcement was an Adam's keynote, obviously like a big deal. There's a quieter thing that's happening on the S3 side, which is that in conjunction with Express over the past year, we've been investing like crazy on the open source and the client side. And so you see this evolution with the S3 folks to really start to think about our ownership of data engineering and the data path to also include the application. And so you see this work investing in a PyTorch plugin. We've got an S3 plugin directly for PyTorch that accelerates check pointing and data loading, right? S3A, right? The amount point stuff for file, like all of this stuff. It's funny, I'm laughing because I'm excited by the fact that we're almost full circle to web services again. The decomposing of the services. Now you're building around S3 is its own entity. Absolutely. Before it was connected in the system before, but this highlights kind of what we've been saying in theCUBE and you're kind of illuminating here is that we've been speculating on this new infrastructure. Obviously Adam laid out the three level stack, I was a model specific, but we've been looking at say neural networks, AI systems. Is there an AI system emerging? Like an operating system like Linux has a function. Kubernetes has a function. Is there a new connective tissue system that is going to emerge for AI? And we're kind of getting at it now, you're starting to see the beginnings of, of what's around the chips, what's around the storage. And that was a big theme in Amazon's keynote. It's not about the chips, it's the interconnect. It's totally true. Yesterday I did the storage talk and it was really fun. As part of it I looked at the history of EBS and we started with the early days of there's a really fun example of EBS Velcroing in SSDs in the early days to add SSDs to their servers to where they are today where, you know, we're built on top of Graviton and Nitro SSD and our own transport protocols. I think the way that you put that, that it's early days, but by looking across the entire stack, the amount that we're going to be able to do in terms of supporting these workloads is going to be really proud. The Jensen thing surprised me. I don't think he was going to be on stage, but it's nice to see him up there. You know, he talked about MV link, okay? Connecting the processors together to have one system. That solves a lot of energy problems, right? So again, throwing a bunch of CPUs on a rack, you're limited by power. So okay, that's your constraint. A motherboard was limited by size, you know? When you put these systems together. So I found that to be interesting. Also he used the term AI factory, okay? Which I found was, he says indecode internally and the DGX to AWS relationship is the AI factory that they use internally. But interesting, they look at it as a factory. I mean, it's a system. Yeah, absolutely. Well, and like you're saying about the components in the rack, I think one thing that we're certainly seeing is that's the expensive bit of the TCO. And so customers really, really want to keep those GPUs busy all the time. And that's an example of the full systems view. Whether it's like getting data onto the box, getting data into GPU memory, right? There's innovation opportunities across that whole space that we need. It's almost funny, the psychology of making those busy could be, are they busy because they got bought and they had to be busy? Or they, it's like, remember the old days, the lights need to be blinking on the routers and like, it's working, see packets are moving. You know, so there's a GPU frenzy, but a lot of people just hoarding the GPUs to get their hands on them because it's a scarce resource. So I see people looking at the GPUs and like, okay, I'm going to provision something. And they go, oh, I got to do more. Right, certainly, I mean, the storage customers and the sort of training heavy customers that we're working with are absolutely focused on GPU efficiency, right? Really not just keeping them spinning but keeping them spinning efficiently and focused on the output metrics. Yeah, my takeaway from the keynote was really clear that cost and performance, a huge driver, huge price performance, Rebecca speeds and feeds, which is great because I think that's what everyone is focused on because the frenzy is going to be, the feeding frenzy is going to be the developers on the LLMs. So I got to ask you on that point, I asked Adam this on the Silicon with Anthropic, what are you learning on an S3 about how models are being built? Because what came out of the Anthropic announcement yesterday is that there's an innovation that the chip guys are learning on the custom side from Anthropic, how they build models. So the chips are learning from the model developers, models are getting benefits from the chips. Nice relationship. Oh, that's what you're saying. Yeah. What are you learning on the S3 side? No, this is an awesome question. So over the past, I want to say year, we've spent so much time on the storage teams talking, it's not just S3, the FSX Luster folks as well, with the Ultra Cluster and all that, talking to big foundation model training customers, talking to the ML teams inside Amazon and actually talking to the framework builders like the PyTorch folks and so on. And for the most part, the message to storage teams is just, don't be visible, right? Like make sure, Be secure, and don't be visible. Be secure and durable, make sure there's paths to curate and make sure the data is available, but do not slow down these workloads. And so a lot of the work that we've been doing is really around making sure that GPU is getting data as it needs it, right? And just looking at that end-to-end path from storage over the network to the GPU and making sure that it's there. And so PyTorch had a load of opportunities for checkpoint optimization, right? You were seeing these models that were like pausing their work while they like shipped a checkpoint so that they could survive, you know, crashes in the software, things like that that happened. Yeah, I think the key thing that you mentioned is the theme is coming up here and the key was end-to-end. Whether it's chip or this, it's an end-to-end system. There's a life cycle, you got to look at the whole thing and totality workload performance. Whether it's token, context window or whatever, it's got to be end-to-end. It's got to be end-to-end. All right, so I've got a final question for you. What are you most excited about going forward, coming out of this event? Obviously everyone takes a break after the event because it's brutal, but it's a lot here. A lot to get intoxicated around with all the tech and the new announcements. What are you going to look forward to after the event's over? I think probably for me, I don't know. Like you say, every year there's so much focus going into reinventing, you're laying on stuff. I think the bridge between the Datalake side and the Genai side has really, interestingly, like had the side effect with the storage teams, especially S3, where we're really, really focused on the developers using S3 for primary storage. And so I think as we move forward, the next year we're just going to keep pulling closer to the builder and the applications. And I'm excited about how that's going to happen. And be invisible, secure, adaptable for them. I was happy. Yeah, that's the theme. Andy, thanks for coming on, I really appreciate it. Good to see you. All right, keep alumni from 2014, breaking down the storage. Key aspect of the Genive AI is the system, it's the end-to-end, is what's around the models, what's around the storage, getting those developers productive, feeding the Genive AI systems with data. They got to be stored somewhere and Amazon's got some great announcements on again. We're looking forward to it, more coverage after the short break. Stay with us back to the studio.