 Good morning, everyone. Welcome back to theCUBE's coverage of Snowflake Summit 23. Live from Caesars Forum in Las Vegas, Lisa Martin with Dave Vellante on our third and final day. Dave, we're going to be talking about the future of generative AI with Snowflake and Neva. Yeah, Neva was a really interesting acquisition. It sort of closed really fast. And we saw it and said, okay, this fits into our model or our scenario where Snowflake is just bringing in different query types and different data types and all fitting together here this week. Yeah, it is coming into focus. We've got the co-founder of Neva and SVP at Snowflake, Shridhir Ramaswamy with us. Good morning, great to have you on the program. Good morning, thanks for having me, Lisa. Congratulations on the acquisition. Yeah, we appreciate you brightening early here. Thank you, excited to be here. Search is fundamental to how businesses interact with data and the search experience is changing so rapidly. Tell us about Neva, this acquisition closed last month, but give us the backstory. Yeah, so Neva was started four years ago. We make my co-founder and I started it. It started out trying to reimagine consumer search. You know, when we open up a new browser tab and type in, we get all of those results. That was the thing that we said we wanted to recreate from scratch. Ambitious mission for a small startup. You know, we were about 50 people, but we created a search engine from scratch. We crawled the web, we built an index of it, we served the data ourselves. And our big aha moment was last year when we saw the wave of gendered AI coming and we realized that all of a sudden, AI was almost like X-ray vision for web pages. You know, like when you tap on a link on your phone, you have no idea what's going to be there on the other side. Maybe it's a pop-up. Maybe they want you to email address. Maybe it's like flashing monkeys. But what AI lets you do is look at the content of those pages, be able to summarize it. We put all of these together and created the first generative AI search engine where you type in a natural language query and we actually give you an authoritative answer. ChatGPT can give you an answer even today, but you are left to decide for yourself whether you should believe it or not. But we would write out the answer. We would make sure that every sentence was referenced so you could see where that information came from. It was also real time because we ran a search engine underneath. It is really this combination of generative AI large language models and search. That was the magic unlock for us. That was the product that we released early this year. You know, fast forward to today. We are very excited about bringing retrieval, bringing search, bringing the power of language models to all of the things that happen within Snowflake. So what you just described in terms of the reference ability of the data, so very granular, not just a link at the end. That's right, that's right, that's right. Like Wikipedia page, go figure out where these sources came from. Like we want to do it at a sentence level. Okay, and so you started the company, I was going to ask you why you started the company. I think you said to sort of rethink, reimagine consumer search. Yeah. But then when you saw generative AI and ChatGPT, well, it was before ChatGPT, that's right. But you started to think about, wow, this enterprise use cases or was it? We were still on the consumer path. We used to, ironically, I used to run ads at Google. And I thought that ads had so dominated the experience both on obviously the commercial side, when you type in a commercial query into Google, you get piles of ads, that was my team, that was like my job. But then I also thought that organic web had become less and less about authentic content and more and more about SEO content, just like flooding its way up to the top. This is why we said we should reimagine the whole search experience. But as I said, generative AI was a big unlock for us because we said we can get away from giving you a set of links. You have to do the work to go figure out, maybe it's the 40th paragraph that has the answer to the question that you're looking for. What it let us do was figure out the right components of the page that supply the answer for you and give you this really pithy, understandable, but believable summary of what you needed to see. So, you well know, in the early days, people thought that ad-based search and keyword search and Google ads was a bad idea. VCs really weren't leaning in early on and it obviously became the greatest business known to humans. My kind has seen. It's crazy, yeah. So, you were, and of course, as well, at the time, like Alta Vista was the best search engine and then Google came out and they didn't try to make a portal and throw a bunch of info at you. They let you control the experience and it was brilliant. Were you thinking about, what were you thinking about the business model at the time? Were you thinking that there was a new way to use generative AI to drive new sources of revenue? So, the business model from early on for Neva very much was consumer paid subscriptions. And, you know, as it turns out, a fairly decent number of people that were users became subscribers, three, four percent. You don't need a humongous conversion rate. What Neva as a consumer search engine struggled with was getting lots of consumers to adopt it as the search engine. And the big tech players don't make this easy. On your iPhone, for example, you cannot change your default search on Safari to Neva. It's not even possible. And then Google makes it super hard on desktop on Chrome. This was part of the reason why we decided that we were better off applying our expertise in search, in AI, to enterprise use cases, rather than continue down the path of essentially hitting our heads against these giant mountains that the various incumbents on the web had become. Well, the other interesting data point is you saw what Microsoft did with open AI. They kind of, Microsoft wasn't really even in the conversation with generative AI. Then they leapfrogged everybody with the business model. But it hasn't really moved the needle on Bing because people are hooked on Google to your point. And so, now you've, okay, so now they're refocused the mindset on enterprise. Which, of course, we're enterprise people, so we love that. And then Snowflake sees the potential. So let's talk about how it fits into Snowflake because it's a really exciting, we've talked all week about Snowflake as many, many query options and they're extending those and increasing that flexibility. So this is one of the many. So let's talk about that. Yeah, so Snowflake clearly is known as the safe, trusted, secure, and efficient platform for all enterprise data. So there's a wealth of information that is sitting out there. But all of us know that this act of actually getting insights out of that data takes work. You have to write the right kind of SQL queries or maybe you will write it in Python with something like PySpark, but you have to do the work to make those things happen. What we are excited by is combining our expertise in retrieval, in search, with the query engine, with other tools like visualization that's already there in Snowflake to create, for example, conversational experiences. We demoed some of this, but the magic with language models is that you can talk to them, which means that the number of people that can ask questions is 10 times, 100 times larger than what it was before, but they all don't need to be schema and SQL experts. You can generate the SQL, we can learn from previous things that have been run, and then we can also interpret the data and say, hey, it looks like the Western region's revenue is going up while these other regions are not doing as well. It is bringing that fluid interpretation on top of the query engine that we think we can be a big asset to Snowflake with, and honestly, that's the reason they bought Niva, the company. What's the number one differentiator of Niva and why Snowflake? I presume there were others interested. Yeah, so from a Niva perspective, the number one differentiator is that we are technologists that have built systems at incredible scale. Even with the Niva, even with the 50% team on a very small budget, we ran a crawl that every day fetched hundreds of millions of pages. Our index was five, six petabytes, and then we serve six, seven billion pages on machines that cost like $150,000 a year. Doing a lot of stuff with very, very few dollars. Similarly, when it came to generative AI, we knew we would go bankrupt if we tried to do things like use the open AI APIs for doing generative AI. So we fine-tuned our own models, we deployed them within our clusters, so we bring unparalleled expertise in search the technology, as well as machine learning and language models. And this is what Snowflake found attractive. And for us, Snowflake was a great place to be, first of all, because there was immense opportunity to bring our technology expertise to bear, that's one. But I would say an even bigger thing is the quality of the team that we interacted with, the engineering leads, the founders Benwa and Thirik, Christian, Greg, Frank himself, I spoke to him several times before the transaction happened. They made us feel incredibly welcome and made sure that they were invested in the collaboration that would need to happen for us to be successful here. So they were number one by a mile. So Snowflake started with SQL, obviously. But when you talked to Benwa early on, he's like, no, it's not a SQL database. You know, it's much more than that. And then, I don't know if there's the right sequence, but Snowflake added data frames, which is kind of the way. That's on programmers. Talk to data. Okay, then they add search, right? And so now with Neva, obviously they get that. And my understanding is they'll take all these different ways of querying data and different data formats and then translate that into a consistent experience for customers that's governed and secure. And that's the magic. Can you explain that a little bit? Absolutely, you know, Snowflake originally started as this amazingly efficient platform for data manipulation, whether it's bringing and transforming and bringing in data or being able to query it. But things evolve. And having your code run close to the data has been this dream that, you know, everybody has had for a very long time. And so over the past two waves, actually over the last seven, eight years, Snowflake has been on this mission, first of all, to make sure that collaboration on data is much easier and then to have applications also run on top of Snowflake. This is why we call ourselves a data cloud rather than a data warehouse. And that's very real. You know, Neva technology is still being developed and integrated into Snowflake, but we announced what are called Snowpark containers. It's a bit of an obscure name. But the point of it is that you as a developer, as a customer, can develop a piece of software and deploy it into a container that runs within the Snowflake security perimeter. So if you want to do a new computation, for example, if you wanted to use a language model to translate feedback text that you have in one of your tables into a different language, maybe it's in a different language and if you want to translate it to English, you can run a container to do this. You can download a model from Huggingface, push it into a container, have it be accessible for all of the code that you're running on Snowflake. And so this is us stepping up our game and saying, yes, there's the core strength of what we do. And our customers love that. You know, our biggest customers, they're so excited about what's coming up with Iceberg tables. On the other hand, they see the extensibility, they see the flexibility that we are offering and the increasing scope of what they can get done. And they're very excited. Things like third-party applications, native applications are enabled by essentially the same infrastructures, no-park containers. And that and the beginnings of companies like Alex's company, you know, where he talked about datasets that are now available for you, I think really represent the next frontier in data and applications really coming together. You know, we've been talking all week about how the Snowpark Container Service was the big announcement, right? And the power of that, you know, it's just coming into focus. I mean, NVIDIA, the announcement on Monday, Snowflake is essentially containerizing the NVIDIA stack and that brings it into the Snowflake environment. We've talked about the supply chain is a company called Blue Yonder, which is the old man logistics business. They're refactoring their business on top of Snowflake using relational AI, which is a knowledge database. And what that means is they can bring in all the legacy, man, logistics workloads, and now they can be first-class citizens in the Snowflake environment. So you start to see the power of all this. A lot of this is misunderstood. People, when they saw the Niva acquisition, a lot of people looked at it and said, well, let's consumer search, what's the point of that? Similarly, people I think are looking at the Mosaic ML, people, I had a VC text me saying it was Niva FOMO, but regardless, the truth is that, or the reality is that there's technology there that can be applied in a lot of different ways. And that's probably, it's certainly true for Niva and I'm sure it's true for Mosaic ML, but what are your thoughts on that? There are a very small number of groups on the planet now. This knowledge will definitely disseminate out but the ability to meaningfully train new language models, even if not the giant models, 30, 40 billion parameter models, but put them to good use is one of the most valuable skills that people can get their hands on. And we've done this at scale. This is part of what made Niva attractive for Snowflake. And there's definitely a realization on the parts of various companies that this is a really, really important skill, an important set of people to have within your company. What is exciting for us with things like Snowpark containers is whatever work we do, whatever work is done by the open source community, we can have like a simple script that one of our customers, you, can use to download the model, just run it and be able to use it. A lot of what powers cloud computing, modern computing, our phones actually came from things like open source came from Linux and we see a similar thing happening with large language models. So in that sense, I think this is an exciting time. It's also a positive time because this means that it's not the case that there are like, you know, just two players that know how to create language models, of course that promptly becomes trend seeking. And so I think this movement towards open source in language models is really useful for all of us. The money that you raised, I presume went into engineering. It wasn't like a lot of promotion and go to market. Is that true? At Niva, most of the money that you would, you know, you can think of Niva as roughly spending $10 million a year on Apex, 10 million on people and that's pretty much it. Yeah, and the reason I bring that up is because it's expensive to build these large language models. It takes talent and then you've got to get access to certain infrastructure as well. Is that true? At this point it is talent. It is some amount of infrastructure. The very largest, large language models, it is true, require like on the order of thousands of, you know, of GPUs. You can think of like, you know, an A100 GPU as costing about $8,000 a year. So if you want a fleet of 4,000 of these, that's like, that's a lot of money. That's $32 million a year. But for, you know, 50, 60 million, billion parameter models, you don't need those. And if all you want to do is fine tune one of the existing open source models and people are doing magic with this fine tuning, that's a thousand bucks. Like you and I can do this in one evening and that's the promise of what is to come. Well, that's what we're doing with the CUBE AI. I mean, you've seen it, it's amazing. We take all of our data from the last 10 plus years and we've created, yeah, and it was not expensive and it's extremely cool. It's the values in the data and what we do with it. It's not our ability to create a large language model. We're using open source to do that. Exactly, exactly. And a vector database and you know, our guys figured it out pretty quickly. And we want to implement things like the vector database, like like really good retrieval and search natively into snowflakes. So you as a customer again can say, I want to talk to this table. Can I bring up a Streamlet app to do it? And off you go, the work is done transparently underneath. So you, I'm inferring that you didn't have a huge, it was more talent, not necessarily the infrastructure cost or did you have to incur a lot of infrastructure cost as well to, how did you do it through the cloud or did you do your own? We ran on cloud, we ran pretty much on bare metal. As I said, our in-front people costs were roughly on par. But let's face it, $20 million a year for a startup is a, that's a lot of money, it's a lot of money. And what works in 0% interest 2019 does not like work the same way in a very 5% 2023. And so you see a lot of startups reset and think about how they should navigate the future. And that's what we did. And we are very happy, honestly, this is a much bigger impact on a much larger group of customers and people. And so it's pretty exciting. Great timing. I mean, it's essentially the organizations like yours that got into it in the last three or four years are now in a position to help the next wave, which is developers like us to just low cost develop in a month and for 100 grand versus what would normally take six to nine months and a lot more money to get to just an MVP. That's the magic of scale. That's the magic of democratizing this kind of access. I think it is amazing that, you know, three kids like renting a database can create an application more or less for the entire planet. We want people to feel that excited about writing an application on top of Snowflake inexpensively, but it's about, you know, like human power and human creativity. And that's a fun place to be. And it's governed and secure and all the promise of Snowflake. It's incredibly important. So you talked about how NIVA is going to be really kind of a fuel inside Snowflake. What enterprises are going to be able to do from it? When you look out at the future horizon in terms of like future predictions, what do you see? What excites you about the power this is going to give enterprises in all industries? So, you know, it's the simple stuff. Make marketplace search, for example, much, much, much better. So you can just type in a query, you say like, hey, I want CPG data in the United States and we're like, here are some great data sets for you. Maybe you want it in Denmark. You know, we make that possible. It's simple things like that. We are very excited about powering much better catalog search. It sounds like, huh, that's like obscure. No, people really want to know what data they have. Some of our bigger customers have, believe it or not, 100,000 tables and to just make sense of it all so that one group can know what is available in a different group to make that stuff discoverable, easy, but do that really well. So it's things like that that are very immediate. There's also like a language model AI track that we are pursuing. You saw some demos. The idea is that any complex cognitive function like writing a SQL query, we want there to be an assistant that will make it easy for you. That understands your context. That understands the database that you're operating in previous queries that have been run on the database. But do it with language models that are within your perimeter where you're not sending data off to some cloud provider with unknown policies on what happens with those queries. And then make visualization easier. The Holy Grail, of course, is like this chat box where pretty much anyone in a company can type a query, a question, not a query, into and get the data and the visualization that they need. Honestly, I think we are a little bit away from that, but make it much easier to write SQL queries, easier to get visualizations. These are all very nice steps in the process. And we also want to enable our customers to be able to integrate their own APIs. So for example, if you are a FedEx, maybe you want to write an application that can call out to an API that can figure out where a package is. And you combine that with the data in Snowflake. So we call it a 2P, a second-party framework, instead of a first-party framework, which is basically how do you make language models usable by our customers as well? I've now described work that is going to take us more than a year. So if you're sitting here next year saying you can build an application with the data on Snowflake and your own APIs and do it in 30 minutes, that would have been an amazing year for Neva. Yeah, well, thank you so much. We, congratulations on the acquisition. We thank you for sharing the future of generative AI with Neva inside of Snowflake. And we look forward to watching this progress over the next year. So you got to come back. Thank you, Lisa. Thank you, Dave. Great to have you, Shradar. We want to thank you for watching theCUBE. Our day three coverage continues. Up next, Snowflake is working with the New York Stock Exchange to take the Snowflake startup challenge to the next level. We're going to reveal the winner next. All of our content that we filmed this week on theCUBE.net, editorial on siliconangle.com. You're watching theCUBE, the leader in live tech coverage.