 Everybody, welcome back to Las Vegas. We're here in the sugar cane at the Emerald Lounge. Mongo, he gave us an in-kind contribution. It's been fantastic. We've been talking to customers all day. We had Dave Ichichiria on earlier. John Furrier is upstairs in the third floor doing interviews with a bunch of AWS execs and customers and partners. Wall-to-wall content, live out of the Palo Alto studio. Check out SuperCloud 5. My name is Dave Vellante. I'm here with two good friends of theCUBE. Sarbjit Johal is the principal and founder of StackPane and a member of the CUBE Collective, and Andy Turai, my good friend from just down the street in Malboro, comes to the studio. Not enough. You should come to the studio more. Andy Turai, Vice President of Constellation Research, AI expert by the Turai, AI is his last name. So he's got that locked up. Jens, good to see you. Thanks so much. It's been a good week. Thanks to you. And we're not even on hump day. I know. We're not even halfway yet, right? But the information overload already, man. That's so much. Wait a second. Information overload already. It's always like that here, right? They just fire hose us. All right, let's start with the keynote. They hit us right away with storage. Yup. Yes. I was like, ugh. That's when I started. Storage, okay. I guess there's logic to that, right? 2006, started with. S3. With S3, simple storage. And then it's not simple anymore. So many tiers, so many policy-based consumption of storage. But this storage, actually, what was told to us during the analyst presentation was there are some caveats to the fastest storage. It's one single region. They're confined to an AC. Yeah, yeah, yeah. So we're talking about the new service, right? So, yeah. That's the thing about object storage. It's supposed to be global. Globally distributed, but this isn't. But it's fast. It's high-performance object. So that's cool. Fast object's been around for a while. I mean, Pure announced it years ago and others did too. Dell has it. But that's the way AWS rolls, you know? Their business model is the killer model, right? So, but I was like, come on. Give me the AI. Give me the juice. And it finally came. They got into it. It finally came, Andy. You must have been asleep for the first 20 minutes. Well, yeah, so I was a little surprised. But on the storage thing, right? They expressed one. What people don't realize is that, you know, they won the cheap storage war a while ago. Now they are making it faster, right? I mean, the claim, if it holds true, 10 times faster and about half cheaper. So if you already use that, particularly because storage is becoming a major issue for all the AI ML programs. And by providing this, I think they're going to win that. Plus on top of it, talking about AI ML, there are, look, the infrastructure war is over. Way long ago, cloud war is over. But the AI ML stack war is still on, all right? And that's where your open AIs and Googles are coming from a different side of things. And AWS, you and I talked about it multiple times. They go at it from a different angle. I'll give you all the building blocks that you want to build. And they established it one more time, saying that you know what? I will let you choose the model you want, the storage you want, the data lake you want, whatever you want. And I'll let you build the best possible model or find you in what you want. And they are staying to their core message, even now. And then of course, we had Graviton 4, that was a big shocker. We all knew it was coming. It was coming. You know, and- We knew that fight was coming. Six is coming. Right, right. They're finally on this kind of cadence. Yeah. You know, my question there is, I think you guys aren't, I mean, maybe you're not really semiconductor, deep semiconductor guys. I'm not really either, but I've talked to some. The question I have is, so Amazon was first with the custom silicon. When you and I have had some interesting conversations about ARM and the ARM standard, you Crawford and I did that one session together. The question I have is, the fact that it's ARM and ARM is so standardized and you can give the standard, you know, the custom design to TSM and they can manufacture it in very high volume with high confidence. Does that give Microsoft an ability to compress the cycle or is this a situation where there is no compression algorithm for experience? I don't know, I don't know if you guys have an opinion, but there really is an emphasis on the silicon. That was a big deal. First layer of the stack, the three layer stack pane. Yeah. Three layers. Yes. So it's infrastructure for AI. It's the LLMs and it's the apps. Apps, yeah. So they have different flavors of the three tiers depending on who you talk to at Amazon. Yes, I think there's a compression there of the experience because the whole industry is learning, so Microsoft has learned from it. As well as there's lack of experience we'll play into. So they will gain, so they're not like eight years behind or five years behind, let's say, let's say five years behind, they started like five years, but they may be two years behind. So that's where, that's how I see it. So Amazon's whatever, let's call it a five-year lead gets compressed to two years, still have a lead, great. So what about the integration? They made a big deal about this, Andy. And the integration with Anthropic is one example and the silicon, of course they have their own models. You know, Titan, the hallway talk on Titan is, like it's really not there yet, right? So that's part of the reason why I said, no, you're going to have optionality, so you can't fix it, feature it. What's your take on the LLMs strategy generally and then specifically what you heard today? The funny thing is, both Matt Wood and as well as Salipsky, they made it common saying that you can't predict which model is going to win, all right? You'll never know. And then, and they also are of the opinion, or at least trying to play the game, that you know, for each of that, the model need would be different. You may not land up using one LLM suits all, which is a story from OpenAI and a story from even Google to an extent, right? But they are taking a different hour saying that, you know what, one model, there's no one size fits all model, I want you to train your own model. And that's why I asked them a specific question about, are you going to turn your Titan into the mega LLM that is going to solve everything? They're like, no, we don't want to do that. We want to custom train to one extent, that it'll solve certain problems. And then we also want to be, I mean, if you look at the collection, I looked at the Bidrock models, I also had the demo AI lab that I talked to the guys who built it. The collection what they have between Anthropic and their own models and then even Lamas there, and then they have the AI 21 labs, their hugging phase models, pretty much they want to become the model playground for everything. And then they have this party rock that they let you build the models, build the apps using that and figure out what fits you and then we will help you fine tune that model. And remember he was also talking about the fact that will help you fine tune it fairly easily, there's a low code, no code way to fine tune it and then you can have your own instance. That's totally opposite story of what the other guys are telling. We will give you the entire model, you use it as you see fit and Amazon is more like, no, no, we'll give you all the models, you choose what works for you and then we'll help you fine tune it and we'll help you rag it and then we will run it for you. So how important is it that you have your own large language model? I mean, Google obviously with Vertex AI, Microsoft and OpenAI, you know, notwithstanding the meltdown the other day, Amazon, IBM, do you have to be best of breed if you're a big cloud company or is open source ultimately going to solve, open source and other proprietary models going to take care of it? I think for training the large language models, like it takes a lot of money, not only money, a lot of like, you know, leadership and then a lot of skills, right? So not everybody can do it, it's a scale business number one, number two, it's like a money business, right? So only few companies will do that and there's no point in training another, you know, another sort of LLM which does the same thing of what we already have. So you're saying Amazon doesn't need to be like a best of breed LLM, does Titan need to be best of breed? No, I don't think so. Amazon is not in that game. Google is, you know, Google is B2C company. Facebook is, right? And why a proxy open AI is for Microsoft, but Microsoft, even Microsoft didn't need the, like, okay, how do I write a point for me? You know, like, you don't do that at work. So they can still make money? Yeah, selling other people's LLMs because they're selling the infrastructure underneath. I think infrastructure is a key for Amazon and not only infrastructure, but platform. There was a lot of Q&A around the chips and like, I was in that session where I asked the question, like, hey, how much of your own managed services that have been moved to Graviton and your own chips, including the non-Graviton chips, right? They say, we are almost there. There are some chips which are sort of abstracted from the customer because everything is like managed services, API call, but some are not. So they explained it pretty well. So I think the differentiation in a nutshell is that from chips, the economics of running big, huge number of data centers, I think that's where the economics comes from. On top of that, it's the platform. It's like the software stack, if you will, in between the consumers and the infrastructure. I think Amazon made a good, strong case that they're the leading infrastructure player, to your point, Andy. But what about the AI? What about Q? What about the LLMs? Well, so, again, they are on point with their message, right? Because I'm not going to get into the LLM war, let others figure it out. And they're going the other way, like, you know, Databricks is doing the same thing by acquiring the Mosaic ML, and there are companies that are allowing you to build the LLMs. So there is more money involved in building LLMs. Imagine this. If your company want to either build a smaller distilled LLMs or the large LLMs that satisfy your needs with your own fine instance, and where are you going to go? Amazon has become a one-stop shop for that. And the funny, interesting thing was, Salipsky, when he was on stage, he was talking about the fact that he can build the 20,000 cluster GPUs. That's actually, imagine the power of that, that's super computer level. And then he brings in Jensen on stage, and he says, we have a 19,000 cluster on your own BGX cloud. I'm like, you know, I'm like, that's a funny thing that he brought in that they were talking about against each other. What did you think about the Jensen interaction? He stole the show, man. Yeah, yeah. He always does. Well, but we've seen Jensen, somebody joke, I saw Jensen more than I've seen my kids this year. Right, because he's at every show. He was at Dell, he was at Snowflake, he was at HPE. It serves them, it serves them well, right? Yeah, yeah. I think on the queue, I think it's a clever approach. We were talking for the last few weeks, like, hey, we, only Microsoft, just one vendor, has so many, you know, agents for developers, agents for like, they call it co-pilot, right? Yeah, yeah. Co-piles everywhere. Co-piles everywhere, right? So how many you need? So you need one, where on the back end, it will sprinkle, it will get the right agent for you. So you have one co-pilot, but it will have the agents on the back, plugged into the back. Not only from Amazon in this case, but also from third parties. So, they showed the examples. Okay, so what's the large language model behind Q? Do we know? No, they didn't tell us that. It's got to be Titan. It's not Titan on this head. They said it's not Titan? No, no, they say that. Titan is one of the models which it can go to. It is a facility. But what they build it on? It's a facility, it's not, okay. It's okay. You're going to tell me all the above? It's not, it's not like. There's got to be one that they started with though. Well, that's the point, right? You don't need to know. What's the primary one? We don't need to know? That's the messaging. Is there not a primary? Well, I know that's their messaging. I'm asking, you know, let's unpack that. I don't know. They didn't discuss that. Is there such thing as a primary? I think so, right? There is, right? Is it possible it's Titan? I think primary is Titan. And if it is Titan, then it's maybe got some work to do. Yes. Oh, good. I mean, on that note, I think. It's a good demo. I think AWS has bought time from the market by promising a lot. They have promised a lot actually today. So an open AI firing Sam Altman and that whole drama bought him some time too, don't you think? I was going to ask a funny question, but the session ended. I said, open AI's all workforce, like Satya was saying like, I can hire all of you. Will you hire anthropic people if something happens there? My question was about talent. Actually, it's a very legit question. Like, where is the talent coming from? Where is the research coming from? Because the research is the key part. The best data scientists, the best people who are doing gen AI, they're paid, you know, five to 10 million a year, right? And Microsoft, no, not Microsoft. AWS is known for being frugal. So my question was that, and which I asked for like walk and talk, and they say, yeah, we have the best talent, they said, but another thing I challenge AWS on is that you always say, like we work backwards from what customers want. I said in AI, you can't work backwards. You have to lead, right? So you've made that point too, but you're also bringing up another interesting point on anthropic. Does AWS have a board seat for anthropic? Not, I don't think so, but now they will after seeing what happened. They're going to go through a full, I would think they would go through a full body scan of anthropic from a governance standpoint. So that they don't get caught off guard like Microsoft did. Yeah, but again, they don't, look, Microsoft made all their bet with open AI, and that's why they want to be very careful that they don't want to make a bet in one company. They are making multiple bets, if you look at it. They even bet on AI 21 as well, and they're bringing hugging face into play with them. Microsoft's got, they're positioning that Microsoft is locked into open AI, which they are, I understand it, but how hard is it to do other LLMs? They did that. They're doing Lama too. They're going to do other LLMs. I don't think to me, that is not the right, like it's sustainable, like messaging. There's messaging that is sustainable is where it took them an hour to get there, but it's like privacy, trust. If you're using our queue, our co-pilot, it will be trusted. We build that in, he finally made that point very strongly, but it took an hour and 15 minutes. I think, I mean, I would have led with that. I would have led right there, smack them right in the face. I think Adam- The Godrail is the program that they release on that behalf as well. Yeah, I mean. And then there's a data discovery and then the data governance program, they have that as well. So there are elements in there, but it's- I mean, I get the emotional 2006 S3, but to me, they just buried the lead. No, they should have started with that. I think they should have started. But it was good, by the way. It was Adam's by far his best keynote. Yeah, he tried, but he paused a couple of times with it because the slide was not up and he- Yeah, but he was good. I mean, I thought that Jensen interaction was a little awkward, like shake hug, you know? Because you know- It's hard to match him. Well, it's funny, I know, he's such a big personality, but it's also, Jensen has such strong relationships with a lot of the technical people, the technical CEOs, over the years, and he's known them for decades. Whereas, you know, I don't think he and Adam have that history. So, but Jensen, John called him a chameleon. He was at Dell talking about laptop AI. I mean, you know, he's just wherever you need. I'm an arms dealer, here we go. Hand it out. All right, final thoughts here. I don't think he was looking at a prompter or anything. He was just doing it from here. He knows his stuff inside out. He's so strong. Any other key takeaways? I mean, we hit Q, the AI stack, some of the other stuff I think is still NDA, so we can't talk about it. I thought the Pfizer gal was really strong. I thought that was a really good customer testimonial. And the other thing is, you know, he did start with industries. It was ironic to me because he started with financial services, the industry that was the last to go to the cloud. Financial services, healthcare. He led with those two automotive. He brought in telco later. He had tech, obviously, with Salesforce and SAP. I mean, go ahead, please. Sorry, there are two things that came out which we didn't talk about. One is that they were talking about this called the easy to capacity overflow. So one of the issues a lot of these companies have, when you're trying to train the models, when you're trying to do things, you're always worried about the availability of GPUs because they're in such a demand that they're not able to get it on demand or even reserve. So this particular option, what it will do for them is that you could schedule for days, months, or weeks ahead saying that, okay, that's the capacity I'm going to need because I'm going to do something big. I reserve it, guaranteed it's available to me. So that's one I thought was pretty impressive. What was the other one? The other one is obviously they're going out to the data lakes with their Express One, right? So basically they are trying to help you with the zero ETL and the Express One combination, cheaper data. So if I'm data bricks or snowflakes on stuff, I'd be a little worried. I don't know. I mean, yes, they're copying data bricks and snowflakes yet again. But they, to me, still have a metadata problem. The metadata is locked into different, the two pizza teams have created, metadata has its own data store and glue has its own metadata data store. Data zones really has, it only has business metadata. It doesn't have operational and technical metadata in there. So they've got these, the stove pipes that they got to deal with. They have to make data zones into some kind of abstraction so that the co-pilots can interact with the knowledge graph, if you will, and take action with confidence. And they're, I think they're quite a ways away from that, to be honest. Microsoft is too. I think there's a lot of slideware in there, but that's the North Star. And I think both data bricks and snowflake, they're a little bit ahead. What they showed at their events was also ahead of the time. You have to do that these days. But I do think those two companies are more focused in their ahead of the game. I think they have threatened some of the existing players, like which are built on top of them a little bit, but not, like they won't topple them. That's, as you said. Another thing is that, the collaboration they did was, on technical side, was that when you are doing SQL for analytics, you can send the metadata about economics. Like I want, I can wait for this body to run later, give me smaller GPUs, CPUs. So that was very collaborative, I think. That's the industry first. So that's one thing. Like the storage tiering, which again, I mean, it's not an uncommon industry trend. It's just AWS adopts it, and it goes into their awesome estate and their business model, so it becomes so attractive because they got half a million customers. And the last observation I have is that, I think they needed to do a lot of myth-busting. And they have a lot of what? Myth-busting, like oh, they are weak in GPU, they don't have a good relationship with Nvidia. I think that they did that myth-busting. Oh wait, didn't they initially shun DGX Cloud? Didn't they initially say no, no? And now they're embracing it after Microsoft announced it? I mean that, that's to me. Yes, out of necessity. Right, so I think that was a missed opportunity. They should have been first to embrace DGX Cloud, because that was inevitable. And I think they were initially like, nah, we want to own that. And then when Microsoft announced, they had to have Jensen here, and they had to not. Then Jensen came close to them, because Microsoft was doing their chip. Oh, I will go hug the bigger bottom line is Jensen wins, again. Anyways. All right guys, we got to go. Thank you so much for spending some time with us on theCUBE. Hey, keep it right there with SuperCloud 5. We're pumping in content live from our Palo Alto studio. John and I are here in Las Vegas with Shelly Kramer and George Gilbert's also here. We got Rob and Rebecca, they're asleep by now in Barcelona. Keep it right there, more action from Las Vegas.