 Welcome to this episode of Intelligent Data Apps, a video podcast that explores the future of data. I'm Dave Vellante for George Gilbert, and we're pleased to welcome Soma-Soma Sagar to the program, back to the program. Soma joined Madrona Ventures in 2015 and focuses on B2B and enterprise-related startups. Before Madrona, he had a successful 27-year career at Microsoft and was most recently the Senior Vice President of the Developer Division. Soma, welcome, great to see you again. Thanks for spending some time. Good to be back here. Thank you guys for having me here. You bet. Now, when you were on our breaking analysis program, we talked about your Intelligent Application Summit that Madrona holds every year. So just to remind our audience, what is that all about? How do you think about intelligent apps? What are they and why are they important and different? Great. How much time do you have? No, the reason I say this is I'm just going to give you a little bit of background and context and then sort of come back to what's happening today kind of thing, right? In a today, everybody talks about AI, everybody talks about machine learning, intelligent applications, AI, generative AI. That's the name of the game today kind of thing, right? But at Madrona, we've been sort of investing in AI startups in one way, shape or form for over a decade now, okay? And initially, it started with what I call AI platform, AI tools, AI, sort of what do you call picks and shovels to help people start thinking about what can they do with AI and how do they think about enabling AI-driven applications, okay? So way back in 2015 now is when we first sort of coined this terminology, intelligent applications, okay? In fact, like, we made a sort of rather bold statement back in 2015, then we think that every application moving forward is going to be an intelligent application. And by extension or by default, if an application is no longer an intelligent application or doesn't think about how to become an intelligent application, we think it's on its way to death, okay? And we made this back in 2015 well before everybody got excited about AI and large language models and chat GPTs all over the world and whatever else we can think of, right? And we are glad to see that in the last eight years, it is sort of coming to fruition. You know, to remind people, chat GPT, which really took people's imagination by storm, is about like, you know, 15 months old or less than 15 months old, okay? So it isn't like, you know, hey, for as much as we are talking about AI and we think everything is AI today, it isn't like, you know, that like, you know, 18 months ago, the world did not know about chat GPT. The world had started thinking about large language. So the rate of progress has been like recent and amazing, but we've been talking about intelligent applications for a while now. And the reason we got excited about this is the following. First of all, let me tell you what I think is an intelligent application, okay? First, as we all know, there has been a tremendous amount of what I call data explosion that's happening. Everybody, every system, every application has got access to more data today than ever before. So you have all this data. The second thing is what can you do with the data? How do you build a learning system that an app can sort of, you know, take advantage of and know that it is not a one-time learning system. It's a continuous learning system. In other words, I have some data. I can use the data to train some models. I can do a better job for my customers because I'm interacting with my customers, I get more data, and then I go back and train my models more. It's sort of like continuous learning, continuous learning kind of thing. So to me, any application that takes advantage of the data that it has access to, that is able to build a continuous learning system and is truly getting better every day, is adaptive and getting better every day is what I think about as an intelligent application. And today with the platform advances, with the large language model advances, with the tooling advances, it's very easy for people to say like, you know, hey, I'm building a brand new application. How do I think about AI right from the get go so that I'm what we call a gen-native AI application? Or if I already have an application I should be quickly thinking about how do I bake AI into my application and we call it a gen-enhanced AI application. So we sort of think there are two categories here. People who are starting new and people who are trying to back fit AI into their existing application. You know, George, we were at GTC last week and it was a heavy picks and shovels event. As you well know, we actually attended the Broadcom Investor Day as well, more picks and shovels. And obviously a lot of action there. You know, the app's piece, that layer was not front and center at that event despite the importance of it. And I don't know, George, I wonder how you think about it and love to hear what Soma thinks I mean, you talked about the two different paths for infusing AI into apps, but I would think, George, that AI is really redefining what's possible with enterprise apps, at least the way we've traditionally thought about them. I think that's a fair assessment. And actually I was hoping we would ask, you know, Soma's perspective on how it changes, like if you can take us through a conventional enterprise app and then one that's AI enabled and then one that's AI first from the ground up. Great. So before we get into that, let me sort of share a perspective on why we might not have seen applications be front and center in the NVIDIA event. It's a NVIDIA event, okay? And NVIDIA is making chips, okay? First and foremost, Silicon, okay? So it's operating at the lowest level to NVIDIA's credit between CUDA and everything else they are doing. They make it easy for people to build applications, but they are not an application company, per se, okay? They're an infrastructure company. So I expect like, you know, most of their stuff is like, hey, what are we doing on the infrastructure side? Whether it is the Silicon layer or whether it is the infrastructure layer from a software perspective, how do I make it easy for you to do what you need to do kind of thing, right? So I'm not surprised that you don't see a lot of applications being front and center, okay? But having said that, we firmly believe that a lot of the value captured is gonna happen in the application layer. We are already starting to see that. I think in the coming year or two or three, it's gonna be all about how people decide to capture the value at the application level, and that's why we think the vast majority of the innovation is gonna happen. I wanna add one more thing here before I answer your question, George, okay? Go back to 2000, 2002, 2003, 2004, okay? Blackberry of mobile devices where it's sort of, you know, the name of the game, those days, right? In fact, they had some pretty good penetration in the enterprises. Not necessarily with consumers, but with enterprises, okay? 2007 is when the world woke up and said, oh my God, I want that device to win my pocket. And that happened when the iPhone showed up for the first time. If you remember, iPhone, when it came out, it didn't even have the app store. It wasn't ready for applications. It said, like, I got a platform, I got a built-in set of applications, but here is what is possible, and that captivated the world's imagination. The first set of applications that came out on the mobile device where, like, you know, you can decide what your favorite is, but like, you know, there are a lot of flashlight applications. There are a lot of weather applications. There are a lot of, let me, you know, get the stock ticker, my stock price, you know, application, right? So, and my favorite games I want on the phone. Those were the first set of applications that sort of ruled the day, year one, year two, year three, kind of thing. It wasn't until five years later, with the advent of two application, I sort of called them canonical, you know, definers of the mobile world. One is Umber, and one is Airbnb. One ended up disrupting the transportation industry, and one ended up disrupting the home rental home space, okay? But those applications really showed people what is possible with the mobile platform, okay? On the one hand, you got access to the device no matter where you are. On the other hand, the capabilities, whether it is GPS or whether it is like, you know, any other capability that you care about, you location information, right? You bring all of those together to reimagine or to imagine what previously people they couldn't even imagine or didn't think was possible, right? In my mind, as far as an intelligent applications go, we are in the pre-Uber and pre-Arabian B space. As much as there is some tremendous amount of work that's happening in the application space, both for enterprise and for consumers kind of thing, I feel like, you know, sometimes I wonder whether we really know what is possible with AI as we look ahead. So I'm like, you know, eagerly waiting for what is going to show up in the next year or next two years or next three years or next five years that is going to redefine and tell us what is possible and what is not possible, okay? So as a capital allocator, Soma, you know, you can't control the timing of these markets, right? You'd have to make bets and you have to leave enough, you know, capital so that you can invest, you know, in your bets and give it some kind of patient capital, but what strikes me is what you're describing in mobile apps with Airbnb and Uber is in a similar way, maybe playing out here and here's what I mean and I'd love your thoughts on this. The AI for consumer use cases is very obvious if you're a meta and you're doing social media or if you're a Google and you're doing search, the better AI you have, the more ads, you get better targeting, you're going to do TikTok, you know, the better AI means better targeting means more ads and consumers tend to lean the innovation and it sounds like you're saying, I don't want to put words in your mouth, but a similar trend is playing out here, you don't know when it's going to happen, but eventually you're going to get a broadening of the consumer apps piece, maybe beyond the metas and the TikToks and the Googles, into another sort of maybe consumer based AI app and that will eventually trickle into the enterprise. Is that how you see it? I actually think that there is innovation that from first of all, nobody can predict the future but my view is like the innovation on applications is happening and it is happening today, okay? When do I see like a completely defining moment for AI platform? I can't tell you whether it is like this year or whether it is three years from now, but people are starting to use AI deeply in applications today and are starting to capture the value. That's happening both for in the enterprise use case and it's happening in the consumer use case, okay? And I can give you some examples if that's of interest to you all, okay? You know, take a company like Typeface, okay? That's a startup that's been around now for two years or two and a half years or whatever it is, okay? It's sort of taking a step back and saying like, you know, hey, how do I redefine the modern content stack? Not the data stack, but the content stack, okay? And how do I use AI and more importantly, generative AI to be able to redefine what is possible in terms of content generation for enterprises, okay? So the core premise and the promise that they have is the following. Let's say you are an enterprise, you have your brand design, you have your brand assets, you have your proprietary data, and you have your voice in terms of how you would like to talk to customers and how you engage with customers. How about I take all of that and take any of the existing large language models and let me fine-tune that and train those models with your proprietary information. And because you're an enterprise, you really care about that, not leaking to any other enterprise, okay? How do I keep that protected for you and just you so that any content that you generate, whether it is a sales content or marketing content or HR content or what have you to be able to be automatically done in a multi-modular way using your brand sensibilities and your voice so that it feels like, hey, it is from your enterprise kind of thing, right? That's what this company is doing, okay? That's happening today. In fact, the first dozen customers, enterprise customers that have started using them are sort of going gaga in terms of how it's enabling them to do things at a much faster pace, being able to do it at a much lower cost, being able to see some phenomenal uplift in terms of customer engagement. And this is all from like, you know, hey, how do I build content much more effectively for you to engage with your customers? That's a hardcore enterprise use case that is resonating really, really well with enterprise customers. So that's one example, okay? The other example I can give you is a company called Glean, okay? That's a company that says like, you know, hey, there's a lot of data in an enterprise, okay? And historically, it's been tough to figure out, how do I search for information? How do I sort of vary that information? How do I interact with that information to be able to get what I need to get done in a much faster way? Glean says like, hey, I'll help you do that, okay? This is something in some sense, like you know, you could argue that enterprise search and enterprise information sort of efficacy, okay? Has been a topic that people have talked about for 20, 30, 40 years now. We really haven't been able to solve that in a meaningful way until, yeah, I showed up. Now, even now, like, you know, hey, you know, Glean is taking a run at that. I'm sure Microsoft will tell you that all the Microsoft co-pilot, particularly the M365 co-pilot and the Windows co-pilot is all geared towards that. And probably there are three other companies that are doing that. So people are thinking about, like, you know, enterprise use cases. Now, if you come to consumers, for example, right? There's a company called Perplexity, okay? Which is sort of taking a step back and saying, like, hey, can I reimagine Google? Okay, that's really essentially what they are saying. You know, previously you came in, you wanted to search for information, but let me give you an experience where you not only, like, you know, want to find information, but I'll make it easy for you to interact with the system and get the kind of information that you truly want and then take it to a task and the completion of the task as opposed to come and search and then, like, I'll give you a bunch of information and links and then you go around, you have to figure it out yourself kind of thing, right? So you are starting to see people reimagining some of the things that were not possible, but is that the maximum amount of reimagination? Are there more? I think there is a lot more that's waiting to happen in the coming years. Well, Soma, let me just pick up on something you said. Picking on typeface and glean as examples, that these are starting to reimagine enterprise functions, but they also have their own stacks. Is that because the modern data stack is not mature enough to handle the unstructured data or do you see essentially a new parallel stack emerging to support unstructured data? So I think about, like, the AI stack as having, like, literally four layers, George, okay? Answering all of this, the first layer, the bottom layer in my mind is data, okay? You need to, there is a bunch of work that you need to do to get into access data, to curate data, to clean up the data, to get your data ready for it to be used effectively in the AI world, okay? So there is a data infrastructure layer, which is what your modern data stack is all about. That needs to continue evolving. One thing you just talked about is like, you know, I think historically we've done a reminning the world has been a fantastic job when it comes to structured data. In the last many years or last bunch of time, they've included semi-structured data, but now it is becoming clear that for the kinds of data that we have and the kinds of work that we could do with the AI and the kind of progress that we could make, you know, just focusing on structured and semi-structured is not gonna be good enough. I've got data in documents. I've got data in, like, you know, mail messages, right? I've got data in my applications. I've got data everywhere. And all of this data is unstructured data. And unless I get a handle on how I'm gonna deal with unstructured data and get it ready to be able to use to train my large language models, I'm not sort of really taking advantage of all the data that I have, right? So the notion of getting your unstructured data, you know, cleaning it up, getting it ready, being able to use it for training models is an important part of what needs to happen with the modern data stack. There is a company, for example, called Unstructured.io that's squarely focused on this problem. They would actually tell you that, like, nah, hey, we are all about getting unstructured data ready for LLMs. That's all they do. But that is an important part of the modern data stack that needs to evolve and needs to get better at so that you have the underlying infrastructure. But that's only the first layer, okay? If you move up to the second layer, then you'll have to think about models, okay? And you've got large language models. You've got, like, you know, small language models that are starting to show up. You've got open source models. You've got, like, proprietary models. And for a while there, there was this active debate about, like, hey, am I gonna use only one model? Is one model, does the world require need only one model or is there room for multiple models? We've always believed in, it's gotta be multiple models. You know, open source is gonna have as much of a game here as proprietary. And application developers are gonna pick and choose which model they want to use for when. No application developer today is telling you that, like, you know, hey, I need only one model. They all want what we call a model cocktail, a collection of models so that they can optimize for price, performance, latency, and even accuracy, depending on the kind of use cases so that they have the best experience, right? So everybody is looking at a collection of models and how can they be smart about which model to use. And then you've got a collection of platform services because you need to think about orchestration among all these models. You need to think about security. You need to think about privacy. You need to think about, like, you know, a whole bunch of, you know, what I call historically like middleware or platform services that you need to provide. And then sitting on top of this are gonna be the applications horizontal, vertical, domain specific, industry specific, and I think, right? So that's really, like, how the AI stack that I think about is emerging for intelligent applications. So just to recap, what I heard is that there are certain capabilities in today's modern stack that are not mature like the data engineering layer, which is the sort of access curation cleaning that we could have a parallel data engineering sort of pipeline that models are a set of services within the platform orchestrated. Here's where the platform orchestration, the security, the privacy, that's the developing governance, the governance and the workflow model. And then, so what I'm hearing is that the modern data stack has to extend itself, but it's not being replaced by something that's growing up in parallel, something just to support, say, unstructured data. You know, I agree with you, I think the modern data stack is what it is, but it needs to evolve and add new capabilities, okay? Like, I'll give you an example, like, you know, three years ago, you know, Victor Database wasn't like, you know, on top of people's minds, right? As people started to think about fine tuning and rag or, you know, retrieval augmented, you know, generation kind of thing, right? To be able to continue to fine tune your model with new things or whatever it is kind of thing. You know, people felt like, hey, I need a vector database that's going to be helpful to me, right? So as and when you see a need, you think about like, how do you extend the modern data stack or how do you evolve the modern data stack? Or like we talked about the unstructured data and how to handle it in the most optimal and effective way. So there are things that I think we need to continue working on the modern data stack, but I don't think we need to fundamentally redo the modern data stack. So let me just, so let me follow up on that. So if I understand that correctly, so again, thinking as a company that who invests in these companies, are you, is your premise that the incumbents that have the data, and they obviously have customers are in a good position to leverage AI and fuse that into their platforms versus what we saw with the cloud where cloud native gave significant disruptive advantage. Do you not see that same advantage here? Do you feel like the incumbents are in a good place to evolve their platform? I think the closest example I can give you is what happened with the data platform, right? You could argue that like, between Google, Microsoft, and Amazon, they were the incumbents, okay? Snowflake and Databricks did not exist. And today the world has five data platforms at least that are of some consequence. And that's because like, you know, Snowflake and Databricks executed really well and they found a niche to start off in one place and then they grew and grew and grew and they sort of, you know, are competing with the incumbents. So that is entirely possible. So a incumbent who's got an application, who's got all the data, can, you know, do what needs to be done to make AI a core part of the application. That doesn't mean that like, you know, hey, only incumbents are gonna win. There's always room for it. It's almost like, you know, hey, who's got the best execution capability for a particular problem and how well they are able to do that is what is gonna decide, in my opinion, whether an incumbent is gonna be successful or whether there is room for a startup or for a new player to be successful and who gets to become the leader in the new world, right? So to me like, hey, this has always been the case and it is no different in the AI world. Thank you. So let me follow up on a comment you made in, just in terms of defining the modern data stack and you mentioned owning the data. And we were talking about this earlier, where, you know, with Snowflake, they literally own the data because even though compute and storage are separate, you have to go through their compute to get to the data. And then in the data bricks world, because of Delta tables, you don't go through their compute. I mean, you often, it's mediated I guess by the Spark execution engine, but the source of truth is owned by the catalog, in this case, Unity. Do you see that all five of the major data platforms evolving under customer pressure, competitor pressure, to that sort of architecture? And in that case, who owns the data? That's a great question, okay? And the interesting thing about all these five data platforms is they all started from different places, but they're all converging to a similar place. I can't tell you that they are exactly the same, but to a similar place, because that's gonna be largely defined or dictated by customers and by the market, okay? Everybody would love for data to be with them, okay? The reason I think, you know, Snowflake did a marvelous job of saying like, hey, we are gonna separate out storage from compute, is because like, hey, particularly when they start off as the data warehouse, compute is how like, you know, you can sort of, you know, add truly add value and you can say that like, you know, hey, I'm gonna like, you know, my business model is gonna work well, but I don't want it saddled by just storage. So let me separate those out, storage is storage and compute is compute and like, you know, I need the more compute I can write, the more, you know, successful my business is gonna be kind of thing, right? I think that paradigm shift has been fantastic, you know, in a lot of ways kind of thing. Snowflake would tell you and so will data breaks and so will, you know, fabric and so will like, you know, Amazon and so will like, you know, the Google BigQuery guys, everybody would tell you that we would love all of the world's data to be with us. The reality is it's not gonna happen. The reality is people are gonna be like, you know, in a very rarely you're gonna see enterprises come and say I'm gonna standardize all of my storage in one place. For a variety of reasons, including the fact that these businesses are dynamic and ever-changing, storage systems are gonna be multiple, okay? So what you need to do is you need to think about like, you know, hey, is there a meta data strategy that I can think about, right? You know, what is my data lake strategy? How do I think about like, you know, seamlessly integrating with all these multiple sources of create so that I have like, you know, one overarching way that I think about things and can that reside with me, right? That's I think where I think the world is headed and where everybody is thinking about because as much as Snowflake would want to say like I want the world's data on my platform, I know unclear that it's gonna happen in my lifetime. So it would be fair to say that if there is one source of truth in the future, it's the meta data and whoever defines that meta data. Okay, Dave, let me switch topics just a little bit. I know in either in your portfolio or in the Intelligent Application Summit, there's been a lot of focus on next generation development environment that's sort of powered by Gen AI. And I'm thinking, you know, for some examples, prominent examples like cool side and replant. And we were talking earlier, I had heard from AWS in particular about their next generation development environment, they're thinking, they said the mindset is don't think about one model that's generating code. Think about a system of models that are exposing, you know, capabilities. And in their case, there's like a topology or graph of Amazon services. They help you build a knowledge graph which is in bottom of your metadata and all the meaning about it. In other words, you express your applications as a knowledge graph. And then there's the execution graph which is essentially the DevOps view and that they're gonna bring all those three together. How do the Gen AI development companies that you're working with step into that world view? Yeah, I think there is a good world view. I would say like all the big guys are thinking about similarly kind of thing. But reality like always charges for developers. It's about like, hey, do I have the flexibility to be able to pick and choose what I want to use to create my own development environment, right? Very rarely developers are gonna be excited about saying like, you know, hey, let me, I'm excited about the Amazon ecosystem. So let me just buy lock stock and barrel into their development environment. That's only thing I'm gonna use. No, that never has worked with the developer community, right? The same reason why earlier when we talked about models, no developer would want to say like, hey, I'm gonna rely only on one model for the rest of my life and life is good. Kind of thing, right? They want choice. They want to know what is possible so that they can pick and choose what works really well for them in their application and even that within their application, depending on the use case, it may vary from one use case to the next use case kind of thing, okay? I think from a Amazon perspective or a Microsoft perspective, any of these guys perspective, I can understand why they say like, particularly if you're looking at enterprises, enterprises have a ton of data that I think has been very underserved so far. So them coming back and saying like, hey, we've got a knowledge graph that sort of really looks at all of your data in a holistic way and we are able to like, you know, help you make sense of it is good. The second thing is like, you know, hey, we want to enable more people to be able to make as opposed to only a few people to make. So the more I can think about like, you know, frameworks or sort of development environments, right? That enable you to seamlessly pick and choose what services you want and how do you integrate them or how do you connect them with the data that you care about to be able to get the result that you want. That's really sort of the core sort of actions that people are excited about and everybody's got to try to offer up a platform and an environment for you to be able to do that. But the thing to remember is even within that developers would want the choice, would want the optionality to pick and choose things from various varying different places, the kind of thing that are really important to them. So the whole world is keeping that flexibility in mind as they are delivering an end to end development environment has a better chance of success than otherwise. So let me see if there's a synthesis here. Like perhaps like a poolside or a replete becomes the development environment, the personal development environment, but it plugs into that broader system of capabilities that potentially multi-cloud. But that the back end, the cloud becomes the back end that's trying to present the services, present the data and its meaning and present the DevOps view. And then the developer can bring their own environment to that ultimately. Agreed, agreed because you really wanted to go back one step further, George. GitHub co-pilot environment really sort of started the trend for like, how can AI come into play in a development environment kind of thing? I was recently joined to one of the largest in our financial services firms in the US. I was talking to their CTO and asking him like, you know, you are like sort of using co-pilot or GitHub co-pilot, how is it working for you? And he said something that was phenomenal for me. He said, and the number he quoted actually was, I thought it's a little smaller than what I expected, but he said like, hey, we've deployed it across all of our engineers and we've got thousands and thousands of engineers. And we've seen across the board, at least a 20% improvement in productivity. Okay? So I said like, hey, does that mean you don't need 20% of your workforce anymore or if you're engineering workforce anymore? He said, no, no, no. For the very first time, I'm able to have a conversation with one of my leaders and say like, hey, what about this capability? What about that? And they are coming back and saying like, hey, I want to do this, I want to do that. And oh, by the way, in that conversation, there is no mention of I need additional account. And this is the first time it's happening in my career. Because of the productivity, we are now thinking about like, hey, what can we do more for our customers? Because we've got this additional bandwidth, additional capability, okay? And that is just one part of like, you know, what is possible for development environment. Take companies like Poonside, they are paying like, hey, you know, fine, you know, we'll sort of generate food for you automatically, but there is so much more to software development than just writing code. So how can I systematically think about using AI in every part of software development? And so the opportunities are humongous there. So I'm looking at some survey data from our ETR partner and the percentage of 365 customers that say they're going to roll out co-pilots is pretty substantial. I mean, yeah, a lot are going to be sort of in the 10% range, but it's really striking to see how many organizations are a quarter of the staff, 20% of the staff. I mean, it's very, very large adoption. So that would suggest some of the conversations we've heard about pushback on price. If a customer is getting that kind of productivity improvement, then price isn't really going to be that big of an issue with whatever, $30 per user per month. That's kind of a no-brainer, isn't it? I also think that you're bringing a good point over on pricing. I think because of how AI is coming in, people are talking about like, oh, you want to use a co-pilot pro or whatever it is, you need to pay 20 bucks or 30 bucks per month per user kind of thing. But remember at the end of the day, sooner than later, I think, everybody is going to want AI. This notion of pricing AI separately, I think as a starting point, maybe it's okay for all the right reasons, but over a period of people are just going to gravitate to like, hey, I need Microsoft 365 and guess what, you know, don't even talk to me without co-pilot by definition I want with co-pilot. So let's talk about pricing. So I think there's arbitrary notion of more price for AI. I think it's going to just be baked into like, you know, whatever the base pricing is. Do you think that will result in a net contract value increase? In other words, will the value of that, will SaaS players be able to, you know, incrementally charge for that value? In my mind, absolutely. The question is like, you know, hey, today I paid, I don't know how much I pay, but let's say I pay 30 bucks for a Microsoft office and I need to pay 20 bucks for co-pilot kind of thing. Suddenly I have to pay 50 bucks. Does it need to be 50 bucks for the rest of my life? No. Can it be something different? Can, does it, will it be more than 30 probably? Right? So I do think like because of the cost incurred in building AI today, and these costs will only go down over a period of time, but today there is a real cost in company that have to figure out how to monetize that so that they can continue doing whatever they need to do kind of thing, right? But, but, but, but will it be more than 30 in my mind? Yes. Can it be less than 50 probably? Yeah. It's kind of like our Amazon Prime and our Netflix. With those, those services that we get value out of and they keep charging us more, we say, okay, fine. You don't even notice it, right? So, because the value is there. Right. Dave, let me jump in and ask one other question. You know me, I'm, I want to go down that some technical rabbit hole, but this one's important about someone, we were touching on this earlier before we started, which is right now these environments that we're talking about are helping developers generate code, but it's still essentially human, written and readable code. Now, one of the things that we saw at Amazon was that over the last 10 years, their forecasting and supply chain planning has migrated almost entirely into a system of cooperating forecasts and agent planning models. Now, they might work on legacy operational applications, but the, the idea of saying, of programming what products are related to what products for substitution or product families or how to account for certain periodicity and events, the model learns all that. That's not written by humans. All of it is in the weights of a set of models. How should we think about what type of software it lives in the model? This is really software 2.0 from Andre Carpathi. How should we think about what belongs in the model? What belongs in symbolic human written or machine assisted human written code? Where is that line and how does that shift over time? Have you heard about Devon? Yes, but I haven't seen the details. Tell us. So basically Devon is coming out and saying that our cognition labs, which is working on Devon is coming out and saying they're like, hey, today you have a software programmer. Tomorrow, why don't I give you the next software programmer? Oh, by the way, is an AI agent. Not a physical human being, but an agent. That's the vision that's been dreamed and everything. And if you talk to the GitHub co-pilot guys or if you talk to the poolside guys that have talked to any of these guys who are working on this, they'll tell you that someday it's not just about code generation or code completion, but it's really like, hey, let me give you a spec. And I expect the code to be automatically written and everything. In some sense, this is sort of a funny anecdote. I was speaking to a group of Microsoft product people last week, okay? And somebody was asking me this question about like, hey, for product managers in this AI world, what skillset should I be thinking about? Again, I sort of, in a tongue-in-cheek way said, like, hey, the first thing you need to do is go back to basics, learn how to write down things really, really well. Because sooner than later, we are going to be in a world where, you know, you're going to write down what you want your software to do and then an AI agent of some kind is going to generate the code automatically for based on your intent. So you better be super clear about what is it that you want, okay? So to me, like, hey, the world is moving there. Now, will we get there like, you know, six months from now or three years from now, we can debate on that, but that's the direction that we are going in, right? So to me, like, you know, software development could be, so today, for example, right? If you're a Python developer, okay? You can say that, like, you know, hey, let me go to chat GPT or let me go to one of these, you know, already out there agents kind of thing. And I'll tell them like, you know, hey, here's what I want, and out comes a bunch of code. Now you let it then go and see, like, you know, is this code really correct? Is this code good enough? What, you know, are there any issues or the security issues or the vulnerabilities or the this or the that? You'll have to go think about all this kind of thing, but it is going to get better, right? And the reason they're able to do that is because they've gone and trained with the vast amounts of code that exist out in the world today, whether they're open source, a lot of it is open source, but some proprietary as well. So to me, like, you know, hey, the more the models can be trained with, you know, code, and the more you can sort of, you know, teach the model or the model learns how to, like, you know, react to sort of new definitions and new requirements. I think we have an opportunity where, like, you know, the vast majority of software, of code development, you know, there is a potential for AI to be immensely helpful. I think for the foreseeable future, it can't be all AI, you still need human beings. So to me, it's all about, like, you know, can AI be a companion to a human being where all the heavy lifting can be done more and more with AI? And the human being is focused on really the highest value tasks. You know, it's interesting, I mean, certainly what Satya would lay out in Georgia. I just wanted to ask about governance, but if you have, I don't wanna disrupt the path that you're on. Well, let me ask one more question on this vein, or in this vein, which is, in this world, we've seen traditional enterprise applications build up, you know, these large, they were focused on process automation, and they spent decades building up data models, process models, screens, a whole stack, and those silos trapped a lot of data and made flexibility, even within those stacks, difficult. How should we think about applications in terms of how they're built, how ISVs build them, or how corporate developers customize them, how that mix changes, and what the application stack looks like? How does that evolve? That's a good question, George. The way to think about it is, you know, what is an application, at the end of the day, it is sort of imagining a set of workflows, okay, automating those workflows to the extent possible, and delivering that in a user experience that works well for the user. What AI is gonna help you to hopefully is, is be that much more better in terms of both identifying and reimagining what workflows you can sort of enable, and more importantly, the level of automation you can do on those workflows, okay? And to me, like, every application developer has to be thinking about that in today's AI context. The other thing they have to be thinking about is like, you know, what is the user experience that dovetails with the reimagination of the workflows to be able to provide a seamless experience for people to be able to do what they need to do with the application, all right? So I think there is a fair amount of reimagination and rejigging that needs to happen at those two levels. Process logic itself, I think there is a bunch of process, there is a bunch of process logic that needs to come into place, and I think you can decide where AI could be helpful for automation that can be part of the core process logic. Okay. So, Dave, you go ahead. Yeah, you guys were talking about metadata before, and George, you and I have talked a lot about unifying metadata. I was just on a call with a company that is actually doing just that company called Hammerspace, I don't know if you're an investor or not. David Flynn. No, we are not. David Flynn, Fusion IO, doing some really interesting stuff, much like he did with SSD or Flash up the stack, but nonetheless, seems to me that unlike in the Hadoop world where there was a lot of shadow data, anybody could kind of do anything, and then after four, five, or six, or seven years, organizations said, hey, we got to get control over this because it's not governed. It seems to be the opposite this time around, that governance is out of the box, that the legal constraints, the other organizational edicts on privacy are really have to be lined up before you can have a successful AI rollout. So I wonder, what are you seeing there? Seems like a lot of opportunities for maybe some of these AI native platforms to disrupt some of the governance platforms that are out there. Are you investing in that layer of the stack? And what are you seeing, Silva? Absolutely. It's like, I think the way in my mind, it works is they go hand in hand. You know, usually you see a tremendous amount of technology innovation, and then compliance and security and privacy, all those things try to catch up, okay? And in some sense, I feel like we are in a similar situation where people have access to AI, people are using AI, and they are sort of being, there's more and more conversation about ethical AI and responsible AI, and AI compliance and AI governance, kind of thing, right? In fact, like EU just talked about the EU AI Act, right? Which is sort of like their way of saying, or EU's way of saying like, hey, here is how I should, our people should start thinking about AI from a governance and compliance perspective, kind of thing. I would say in the last year or two, there's been a lot of startups, new startups, okay? That are coming up in the AI governance and AI compliance space. There are still companies that are sort of growing up in the data compliance and data governance space, and sort of trying to figure out like, can they cross the catheum and also be AI governance and AI compliance, or are they two different things, kind of thing, right? So whether it is the large compliance companies or the incumbents that you call, or the startups, you know, there is a fair amount of energy and excitement around AI compliance and governance. Like anything else, one of the things that I worry about is like, is our enterprises truly, truly, truly ready for that today, or are they still in early phases of getting excited about AI and trying to kick the tires and trying to learn what it means? That to them like, yep, someday I need to wake up and think about governance and compliance, but let me wait till tomorrow as opposed to, oh my God, it's a burning problem today. So I think timing-wise, I wonder whether, like, hey, it's sort of a little more later in the coming, but I think it is important, it is happening. Startups are coming out every day. In fact, like, we have seen probably half a dozen startups in the last few months on specifically on AI compliance. In fact, I invested in a data compliance company just in the last month or so. So we are absolutely looking at compliance and governance. We think it is important. I just wonder whether, like from a timing perspective, is it today or is it in the next 12, 24, 36 months kind of thing? So given that, Soma, I mean, how aggressively are you? I mean, I know you're investing in AI, but with all the talk around LLMs are going to get commoditized and I'm sure you obviously understand that well. And as you say, the up-the-stack, particularly the applications piece, the intelligent applications piece is going to take some time to evolve. Very hard to predict when. Could be very vertically market-focused. I don't know, sorry, let me stop there because I want to make sure that one thing I say is clear. I think intelligent applications are happening today. Yeah, you made that point. You did make that point. Yeah, what I don't know in terms of timing from an application perspective is like, hey, when do we see the Uber and the Airbnb of the AI world? Is that tomorrow or is it three months from now or three years from now? I don't know that kind of thing. But applications are being built today that are intelligent applications and they're going from strength to strength. So from an investment perspective, I think today is the day, if not yesterday, not tomorrow. Okay, so following up on that then. If I understand it correctly, you're saying it's happening today, but largely we're talking about infusing AI into existing apps, although you gave a couple of other examples, which I guess you would consider a perplexity, obviously an AI native and... Yeah, perplexity is the native AI company type Right, but we haven't had that Uber Airbnb moment, is what you're saying? Yeah. And that's coming. So how do you think about that from an investment standpoint? Do you just invest heavily and hope that day comes? Are you step back and wait for signals? What's the strategy there? I think like anything else right now, if you're an early stage investor, you look at the concept, the idea, you know, squint and see what might be in the future and then take a bet on the team, right? That's what you can do, right? And then you wait for some early signals, like, you know, when typeface tells me that they've got like, you know, dozen or more enterprise customers that are Gaga, that's a good signal that like, hey, 12 enterprise customers are saying that, maybe there's another 120 enterprise customers that they could say that and someday hopefully 1,200 enterprise customers, right? So you look for some early signals, but it's really taking a bet on the secular trend of how AI is going to fundamentally change how application builders and application developers can rethink and reimagine what they could do in the application layer. One of the things that was striking at GTC, I'm sure you saw Jensen's, you know, production was, the number of robots that he had on stage in a private meeting with analysts, it was like a two hour meeting, he said within eight to 10 years, we'll have, you know, generally functional robots. He was very confident of that and that just kind of blows your mind in terms of the application layer that can occur there. I'm not sure that's enterprise or consumer, it's just new. I think it is both. It is both. I actually think it's got to be both. I don't think it is going to be either or kind of thing. Sometimes you may see some, you know, breakthrough innovation happening in one that leaks into the other, but I think, you know, both sides are going to be moving the ball forward as fast as they can. Yeah, and then the other really striking thing was the whole notion of simulation or content creation versus animation. And you think about all the challenges that content companies are having now. You look at what's happening with Disney and Pixar, et cetera, and you see how AI is generating this phenomenal content that it's all simulation. How do you think about that in terms of its disruption potential? Yeah, that is a company called Runway, okay? That's really saying like, how can I use AI to be able to help you generate video content? And I'm sure their dream one day is, like, hey, can the next Disney movie just be like somebody sitting in front of Runway and being able to bring it together, right? So, you know, it's still probably find a little while away before they could do that, but that's their vision and that's their goal and technology is advancing fast enough that it's going to be possible someday to do that. Yes, sir, George, the writers were right to be concerned, but anyway, I know I'm getting off the topic of your podcast, but it's great. Yes, if you know this is... Okay, so let me just end then, Soma, on the topic of like AI, ERP, like how might we imagine ERP? Is it something that, you know, Microsoft's sort of talking about this where you might mind the logs of a bunch of operational applications and you sort of create a new process on top of or across the different existing legacy applications, but this process is more adaptable and agile because it's not deterministic, it can evolve and it learns not just from the logs, but from the expert users who when they, you know, have to disengage the autopilot, they put their hands on the wheel, that's a signal to the model that it didn't get it quite right. Do you see a new generation of ERP sort of growing up almost on top of a sedimentary layer of the legacy operational applications or do you see something being written from scratch with, you know, with this new AI enabled stack? I actually, ERP is a fantastic example because it's probably one of the most complex, complicated, you know, enterprise applications, okay? But it's fundamentally composed of two things, data and processes, okay? The problem is that data from so many different sources that have been built up over the years and integrated with over the years and the processes that have been built up over the years kind of thing, so it's sort of, you know, is a whole bunch of spaghetti that sort of, you know, somebody hopefully in the world understands how these things work together, it's very, very complex, okay? That's a classic example of a space or an application where how easy is it gonna be to truly bake AI as a core part of the application? I think the jury is out on that, okay? At the same time, the amount of work that needs to happen to build a ERP application is humongous, okay? So it's one of those things where, you know, you'll be damned if you think this or if you'll be damned if you think that kind of thing, right? Having said all that, I know of at least like, you know, a couple of efforts where people are starting to take a piece of, you know, fresh paper and a pen and saying like, you know, hey, let me reimagine what a modern ERP solution could look like, where AI is at the center of the application right from day one kind of thing, okay? But because it is a very complicated thing, it's gonna take like, you know, multiple years before the thing shows up in some meaningful way that the rest of the world can say like, you know, hey, this is the future, this is what I need to do kind of thing. So there are efforts, particularly in startup plan, to reimagine what a AI-driven ERP could be or what a AI-driven CRM could be, right? You've got Salesforce, you've got like, you know, SAP, you've got a whole bunch of other people who are what do you call incumbents in the ERP space and the CRM space. And there is a sort of, you know, early thinking in the startup world of what like, you know, hey, is it time for us to start reimagining what the next generation could be? Now the incumbents are also I'm thinking, you know, I'm sure I'm thinking about how do I bake AI into my existing workflow, my existing application, existing processes with their existing data. And now when one of them, you know, make enough progress that they could be the future probably, but I think there is room for innovation in and from a startup perspective to see what is going to be possible. But there's definitely tension. You know, when we talk about the Dell Data Platform, we, in our, oops, I lost my microphone here, sorry. In our ideal world, we would have, right today, I think we can all agree, there's data and metadata that are locked inside of application silos. And we would love to have a world where data is accessible by all, metadata is unified, but to the extent, for instance, some of that Salesforce does a really good job with AI or Workday or Oracle or SAP or whoever it is, that's going to create affinity with those siloed platforms. And that is sort of antithetical to this vision of a data platform. Now, maybe new companies will emerge that are AI native and develop new processes and really rethink how businesses are done. And perhaps that's where the disruption comes from, but it seems like there's a lot of inertia there and that could take some time. It is a lot of inertia, but also even the new application has to think about how do I integrate with the existing systems and existing data? Because without that, you would be dead on arrival kind of thing, right? So as much as you can reimagine and build an application, you still have to have enough connectivity with the existing world to really be able to move people from the old world to the new world. And so it's going to be a, it's going to be a, what should I say, it's going to be a time consuming process either way. But I think there may be enough of a platform shift here and enough of a value that I think people may, people first of all are excited about thinking through what it could be. And I think there is room for it to be successful because people have been sort of using the same ERP solution for 30 years and all the frustration that exists if there is light at the end of the tunnel in terms of why their lives could be better, they may say like, hey, fine, and let me sort of go through the pain of transforming myself from the old world to the new world. Well, Soma, I really appreciate it. George, are we good? Yeah, very good. I think we've tapped Soma out. Well, I mean, Soma, you have such a really wide observation space and depth of knowledge in so many spaces. I really appreciate you spending some time and helping us serve our community in this Intelligent Data Apps podcast. Thank you so much for coming back up. Absolutely, and thank you again for having me here. It's always fun and pleasure to be here and chatting with you all. Take care. All right, good deal. All right, for George, Gilbert, this is Dave Vellante, and we'll see you next time.