Good morning, everybody. Good morning! Good morning. Did you guys catch it on stage last night with Jensen? Isn't that man incredible? What a force of nature. I love his vision. You know, you just put the AI factory on the data, flip it on, and poof, the insights start flowing. We're actually gonna show you that a little bit later on. It's a demo, of course. Don't take it too seriously. You know, I The two letters AI are being mentioned. It's almost like there's the only two letters left in the alphabet, right? You know, I hear there's a drinking game going on in the back, that every time I say AI, they have to do a shot, okay? I'm gonna be a little bit sparing with the AI references, just to kind of keep order here. What we want to talk about, what I'm gonna talk about the next 15 minutes, so it'll be relatively quick because I'm gonna, I'm gonna vacate the stage, and we're gonna really show you some incredible things. We always say, in order to have an AI strategy, we have to have a data strategy. You're gonna hear that, you know, more often in the next couple of days, because we can't just turn it on and hope for a miracle to appear. Data strategy is something that, you know, I can't help, but three times a day in a normal week, you know, I end up having the data strategy conversation, you know, with CIOs and all kinds of C-level people, because it's a fundamental choice that you have to make on your journey. Because if you're not having your data strategy wired in, you cannot just sort of continue on with the past. You know, we got to make very specific choices so that we enable these AI factories, you know, all the benefits that we're seeking from AI learning. There's real choices to be made as opposed to, you know, we'll just perpetuate what we've been doing in the past. We're just modernizing, we're going to the cloud, we're using new technologies. Now, there are fundamentally different choices to be made along the way. Our data strategy, no surprise, is the Data Cloud. You can't shut me up about Data Clouds. They're my favorite two words, in the dictionary these days. For that occasion, you know, we have, I'm descending Data Cloud on Las Vegas. Isn't that beautiful? Yeah. It'll be here, all week, and it'll move somewhere else, next week. The fundamental thing, you know, about the Data Cloud is, look at this thing. We call it the snow globe, affectionately, you know, within Snowflake. And this is not a visual, interpretation. This is an actual rendering. Every single white speck is a unique Snowflake account, and every line that you see in here is a data networking connection with another Snowflake account. This is very dynamic because the data is flowing. That's fundamentally what it is. Now, this is all like, okay, 10,000 feet, 50,000 feet, very conceptual. We'll drill down here just for the sake of argument, to see what's going on below the covers. This is a kind of a Google Earth-like kind of thing. We tried to make it look like that. We drilled into the Data Cloud of one of our great customers, Fiserv. Thank you, Fiserv, by the way, you know, for letting us talk, because for a lot of people, this feels very proprietary. Fiserv is more than, "Okay, we're running Snowflake." They have data networking relationships with other entities. I just want to use this by example to see, look, there's hundreds and thousands of these living in the Data Cloud. For example, they have a connection with a company called Clover. That's actually a wholly owned subsidiary, you know, of Fiserv. They, in turn, have a relationship with Heap, and they acquire that data to create customer 360 type of views. Fiserv is a very large financial services concern. You probably remember First Union, credit card processing things, a lot things of that sort. Creating data is very, very, very central to how they operate and what they do for a living. The data connections, they're like synapses. Not that I like to use that word, but, you know, they are creating synapses that are literally going all over the place, right? Here's a relationship with one of their banking clients, right? There is a fintech in there as well, that either uses their data or provides data back to Fiserv. It builds, and builds, and builds. We're just showing a few of them here. Carat is actually a data as a service company that Fiserv, you know, has built itself, and that's how they provide data to other entities, in this case, ExxonMobil. Everything is Snowflake, to Snowflake, to Snowflake, to Snowflake. The point here of a Data Cloud is, it's not defined by enterprise boundaries anymore. It's defined by your business relationships, defined by your ecosystem, and it's dynamic, right? You can build these connections, you can take them away, you can reconfigure them. That's really what's going on every single day of the week, all over the Data Cloud. When I zoom back out, I'm just zooming out, not completely, I'm just zooming out to the financial services Data Cloud. This is our industry Data Cloud. This is where all the entities that are in the financial services business come together. You know, we now have seven industry Data Clouds, so there's all these virtual different views, either as a, as an institution, as an enterprise, I have my Data Cloud. Industries have their Data Cloud, and then we have the big, big Snowflake thing across 40 different regions all over the world. I want to emphasize a couple of really key underpinning concepts here. This, what you see here is a data silo. 70 years ago, or according to Jensen, 60 years ago, when the first database, you know, was created, there was not a siloing problem because there was only one database. That was great. When there were two databases, now we have a siloing problem. Why is this table over here and not over there? Ever since, you know, we've been mushrooming like crazy, right? This is an incredibly hard problem, because silos are created just by running workloads. Silos are created by introducing new applications. I mean, you have to literally fight it. You have to have a very strong strategic posture towards a single data universe without boundaries. Obviously, this is hugely important to enable the AI revolution. Take another shot. Any other form of data science, analytics, it's incredibly inhibiting and impeding, you know, when these kind of boundaries say, "Okay, now we got to start pumping data across again, FTP, KPIs, all the three-letter acronyms." Unsiloing is really the core theme of what we try to do. This is like a matrix moment, okay? We're evaporating the silos. Isn't that beautiful? Keeping it going like that is the hard part. Secondly, the Data Cloud doesn't just consist of structured data. Obviously, in the world of data warehousing, it was all structured data. Snowflake actually got its start early on with semi-structured data, really good at processing JSON and so on. two years ago, we really started to bring petabytes and petabytes of unstructured data into the Data Cloud. Obviously, again, super important for AI and all these applications that are very much training on textual data. The problem with unstructured data is, how does it become a full participant in analytical processing if we know very little about it, right? It may have a name and a timestamp, and what's inside of it, very human readable, but not readable for software. Last year, we acquired a company called Applica. Later on in the session, we're going to refer to this as Document AI. This is the AI that we already acquired into the talent and the technology more than a year ago to help us derive structure from unstructured documents using language models and using AI techniques. I really love this rendering, because you get this formless mass turning into a, you know, a set of structures. This really helps unstructured objects to become full participants in analytical processing, right? That really amps up that opportunity. We're very excited about it. You're going to see more about it later this morning. It doesn't stop there. One of the things that we have invested most of our engineering resources in terms of enabling the Data Cloud, are workloads, right? I mean, the whole point of the Data Cloud is the work needs to come to the data. We want to stop the data from going to the work because that endlessly silos and resilos the world. If we can bring the workload capabilities, full spectrum, from analytical, to transactional, to search and everything in between, then the data can stay put, because the work can be executed effortlessly, you know, on that data. There's a lot here, okay? Starting off, I know most of you in this room are doing data warehousing. Data warehousing, once upon a time, it was a market, it was an industry. There were whole companies, that's all what they did. These days, we just view data warehousing as a particular type of workload, and we can kind of argue about exactly what it is and what it isn't. We sort of, kind of know what it is. Beyond that, you know, We have data lakes. Data lakes are maybe not the same as data warehousing, yeah. I was just talking behind stage with Mihir from Fidelity, they really view Snowflake as their data lake. Most of our customers view Snowflake as their data lake because it's very, very broad-based, we're going to make announcements this morning, you know, how we're even more expanding our data lake scope in terms of the type of data that we can address. Make no mistake about it, that is very much a workload that we execute on. Unistore, we talked about that last year. These are our transactional capability. This is about as different as you can get, you know, from large-scale analytical processing. We now have to update data objects in place. It really implies a very different underlying stack. We introduced hybrid tables. A hybrid table is an object that is both transactional and analytical, right? We now have a handful of customers already in production, even though the capabilities themselves are still in preview. There's tremendous demand for this, and we'll be pushing very hard, you know, to get bigger and stronger and more functional in this area. Stay tuned for that. It's very important for our ISV customers as well. It goes. These workloads are all orbiting around Data Cloud. Collaboration, data sharing. 70% of our large consumers, and we define large consumers as Snowflake customers that consume $1 million a year, are using data collaboration. On average, they'd have 6-7 of these data networking relationships. We call them edges. We have all different kinds of edges. 70% of them are using it. It's literally growing in leaps and bounds. This is a very important part. Again, Data Cloud is defined by its networking relationships, not by any arbitrary workload, perimeter, enterprise perimeter, or anything of that sort. Interestingly, cybersecurity became a workload on Snowflake. Initially, it was really an economic play because people said, "Hey, I can save $0.75 on the dollar just by keeping most of that data in Snowflake instead of in a, in a SIEM." They figured out, you know, cybersecurity is also a big data integration problem, right? If I'm a security analyst, you know, it's much easier if I can query a series of IP ranges across all these different datasets. You can just imagine, if you let loose one of Jensen's AI factories, you know, on top of this data, you know what kind of questions you'll be able to answer. It's very exciting. Cybersecurity, we move on. Data engineering. We're all data engineers here in this room. You know, we sometimes estimate, it's very hard to know exactly, that roughly 40% of the compute that we consume in the Snowflake Data Cloud actually is, you know, data engineering function, so it is enormous. All our efforts, I wouldn't say all our efforts, a lot of our efforts in Snowpark over the last 6 months, ever since Python went GA, have been focused on really retiring, you know, legacy Spark jobs. This is the easiest pickup. If you're looking to save some money, I mean, you can save 2 to 4x on your spend on data engineering. That is because, you know, our interpretation of Spark, Snowpark, I mean, it will run faster, cheaper, operationally simplified, and much safer in terms of governance. We've seen enormous uptake on data engineering in Snowpark. If you haven't looked at it, encourage you to do it. There's just free money to be had there. All these plans. AI, another shot. We're going to talk about this a great deal. You know, 70% of our large consumers in Snowflake, obviously, you know, are running these type of jobs. I mean, ML has become very mainstream over the last couple of years. We see people running models all over the place, and this is just going to go into overdrive in a huge way, just because of all the new technology that's arriving. One of the big things we're gonna talk about this morning is how we are going to enable that inside of Snowflake, because that's really important for you to know exactly how we're gonna enact all of this. Applications. We've made a huge effort for Snowflake to become a place where people want to build applications, because essentially, you know, it's a layered cake, right? With the infrastructure, the public cloud, you know, we have the live data, we have the workload enablement, we have the marketplaces, we have the transactional capability for monetization. That's a very attractive place for software companies and software development to come, to build something, to sell something, to transact something. We now have hundreds and hundreds of ISVs, sometimes some of them quite large, actually, that are now building on Snowflake. We're really looking to set off a renaissance in software development on Snowflake. We really think that applications should be built on a Data Cloud, you know, not on a database. You know, I'm giving you all the reasons why we believe that to be the case. Now, there's lots of them out there. Some of them are built by large enterprises that are building customer-facing applications. You know, DTCC, great example. That's actually a regulatory function in the financial services market. They're really owned by their member banks, and they manage the liquidity levels that banks have to maintain. It's very, very cool. Because so many financial institutions are on Snowflake, it's all Snowflake to Snowflake type of functionality. You see Simon Data, Bank of New York Mellon, Okta, and it goes on and on and on. Now, I want one of them called out specifically. I was on a stage, actually, here in Vegas a couple of months ago with Blue Yonder. Blue Yonder is the former JDA. They are the largest supply chain management software company in our industry. I've always had a personal affinity with supply chain management because it's one of the largest, one of the last remaining areas in the enterprise that has never been platformed. There's actually reasons why it's never been platformed. It's very frustrating because it has been a spreadsheet, email business, very inefficient, and there's really an enormous opportunity for automate, for digital transformation and really, you know, bring this into the modern age. What are the problems with supply chain management? Supply chains consist of many different entities, right? Getting visibility across the data in different entities, guess what? You know, we need a single data universe. You know, we can't have these silos and all these boundaries stand in the way of that. It is fundamentally a data integration problem, we can solve that with Snowflake. The second thing about Blue Yonder is, as supply chain management is, you know, when they have events happening in the supply chain, they run these incredibly compute-intensive processes very, very fast to be able to figure out: What do we do? What are our options? What are the scenarios that we can enact here? Of course, you know, Snowflake lends itself really well, you know, because we can fire up instantaneously these workloads, massively provision them, and get the results. We finally have an opportunity to go after supply chain management. Of course, the network effect for retailers and consumer packaged goods companies is going to be enormous. Blue Yonder is a $1.5 billion software company. This is quite large for people to re-platform onto Snowflake. We're excited about that. Now, one of the things that's I remember five years ago in San Francisco, 2019, it was my first Summit conference. I remember doing a presentation. I don't remember what I said, but I walked off the stage and somebody said, "You didn't talk about governance." I'm like, "What?" I didn't know much. I was only in the job for 6 weeks at the time. We just stopped. Yeah, we didn't talk about governance, and we should have, because governance is so central to the Data Cloud, right? I mean, what this visual is trying to show you is that there is a shield, there's a protective shield around everything that we're doing. You may like, "Well, that's nice." The problem is, you know, when you come from an on-premise world, you know, you have your own security perimeter. Governance has sort of been infiltrating all our thinking. Every time we do something, we need to fully think through all the risks, you know, that are represented by new functionality, the risk of exfiltration, the risk of, you know, compliance violations, all these types of things. You know, when we brought Python out, I mean, you can do that in a week if it's just Python, right? The way we had to do it, to harden it, make it non-porous, and really eliminate all the risk, it took us two years to do that, to really make, you know, Python an enterprise-grade, you know, high trust capability. Every time we do something, this is the first time. This is also why it takes time for us to take things out of preview, because these things are in preview. Guess what? It's always security compliance type of issues that are slowing down the release of these features. Everything we've done in terms of workload enablement, it runs inside that governance perimeter. That's really the big change from Data Cloud from before, where it's just a data universe, a big one, and something that runs across enterprise boundaries and so on. We brought the workload platform into the governance perimeter as well, and we call that Snowpark. Snowpark is the programmability platform for Snowflake, and it is completely governed in terms of its implementation. Very important aspect of what we do. The user types, you know, historically, you know, when you come from where we come from, as database people, of course, you know, SQL people, you know, obviously the general population are very important audiences for us to serve. Over the last couple of years, we have massively focused on the left side of the spectrum, the much more technical user, you know, to give them, you know, the type of experiences and the type of interactions, you know, that they want to have. We really address the entire spectrum, and we view all these roles as equally prominent and important, and we will continue down that path. We'll make three announcements really quickly, and then in the next section, we're going to take you 15 feet underground, show you in excruciating detail what's going on there. Very, very excited this morning to share these things with you, because some of these things have been in the works for the last year. Some of them have been in the works for much longer than that. The first one is... You can probably interpret this picture, icebergs. We have talked about our support for Iceberg open table file formats for some time. We now have greatly expanded that. There is a whole discussion here. You know, when do I use, you know, Snowflake internal objects? When do I use Iceberg open table formats? Do I want to manage my own storage? Do I not want to? What functionality, if I'm giving up, not giving up? Is it a managed versus an unmanaged object? We're going to talk about that. This is a huge step forward. Obviously, the goal here is we want you to be able to reference any and all data. We're putting the infrastructure in the Data Cloud to make sure there are no limits to your ability to reference and address data, you know, especially in the face of, you know, all the new developments in data science. Super important. What we're also announcing, number two today, is what we call our Native Applications Framework. We've been working on this for some time. By the way, when we go to our partners pavilion, all the exhibition areas, you'll be able to see these applications. These are built in our Native Applications Framework. The highest level sort of analogy that I can give you for what this is sort of the Apple iPhone, their App Store, the Apple iPhone apps. There is now such a thing as a Snowflake application. We're now building applications as Snowflake applications. We're super excited about that because they're using all the common services, they're using all the common governance frameworks. Obviously, our database engine, which is really the centerpiece to the entire platform, that's all here. If you have an iPhone and you want to run on Android, different development environment, different deployment environment, you pretty much have to maintain separate platforms to address that. The great thing here in cloud, you build one application that can run unmodified, completely agnostic to the underlying cloud. If you're a software developer, you can address the entire market with a similar single implementation, and we think that's incredibly cool. These apps, there's going to be hundreds of them. If we're successful, there are going to be thousands of them. This is really a core part of our strategy. Apps are not just limited to application developers, software companies. Everybody these days is a software developer. Everybody is building customer-facing applications, internal applications, and so on. Finally, we're announcing Snowpark Container Services. You just saw an app disappear in that container. This is a huge expansion of Snowpark, and we're going to show that in dramatic fashion in the next segment. When we first introduced Snowpark, and we started to GA Python, a lot of people saying, "Yeah, this is great for a spark job and the Hadoop jobs, and, you know, we recompile them, you know, run them on Snowpark." A lot of people said, "Wait a second, you know, I have engines. I have applications. I don't even know how to, you know, what to do with these things. You know, I cannot just sort of, you know, rewrite them and recompile them and all of that." That's why we developed containerization, right? We can take whole applications, whole sets of services, and we can now run them inside the Snowflake governance parameter. If you've been on the, at the pavilion for partners, there's quite a few of them that are already running inside a container. It's actually quite easy to do. Though this is a very, very new thing, a lot of people are already running it very, very quickly. I mentioned Blue Yonder earlier, and they have a lot of legacy engines that they still need to be made available. They can't rewrite them, right? They just want to host them inside of Snowflake, use the database engine, and just make the services available like they did before. This really helps us swoop up any function, any application that already exists and say, "Hey, I want to run that in Snowflake." This is how you do it. It's not a coincidence- This is also going to be the strategy. This is a large language model that's now disappearing, you know, into the container. This is how we're going to host large language models, and we're going to show you that in the segment coming up. It's a great strategy because we don't have to port them, we don't have to hack them. You know, they're gonna go wholesale, you know, into the container, and then they can be addressed, you know, by the applications themselves. Very, very good way because we are anticipating very, very rapid development in these areas, and as a result, you know, we need to have a strategy to adopt that and then make that available to the world. With that, the messages, you know, of the announcement this morning, no limits. No limits means we're pushing the limits back on applications, pushing the limits back on the use of data, pushing the limits back on Snowflake. I mean, the whole strategy here is for the Data Cloud not to have any limitations, and you have our word for it, and we will continue to do that in any and all areas. With that, I will yield the stage, and thank you very much. You've been a great audience. Thank you. I'm Sheila Jordan. I'm the Chief Digital Technology Officer here at Honeywell. It's just so amazing to see all the different areas and businesses that Honeywell is involved in. Everything from aerospace, oil and gas, energy, healthy buildings, our safety and protection equipment, of course, Honeywell Connected Enterprise. I think the entire audience would agree that the last three years have been, I think, the most challenging in industry. Really what's been one of our saviors inside of Honeywell is the Snowflake Data Cloud. In January 2021, we started to see the impacts of inflation from the supply chain. In October, that number grew quite large. We actually had our supply chain teams build an inflation index across all those different products and pieces and parts and components and raw material. When you looked at the inflation index, and we could actually match that to our pricing application technology, it was like a match made in heaven. We could actually look at doing inflation-adjusted pricing, which had a significant impact on the company's PNL. When I think about the future of Honeywell, we're actually becoming a software company, we're putting together the products and services that we have traditionally delivered into applications. We're creating a whole stream for the connected worker. Almost every function in the company has seen a benefit by using the Snowflake Data Cloud. Snowflake has really helped us transition Honeywell into a data-driven enterprise, that we are allowed to have this information near real time at our fingertips. Our tagline is, "The future is what you make it." We're future shapers. Please welcome Snowflake Co-founder and President of Product, Benoit Dageville, Snowflake Senior Vice President and founder of Neeva, Sridhar Ramaswamy, and Snowflake Director of AI/ML Engineering, Mona Attariyan. Hello, everyone. It's so great to be here. My name is Mona Attariyan. I'm Snowflake's Director of AI and Machine Learning. This year has been absolutely amazing. I feel like the world is now as excited as I've always been about machine learning. Before we dive into all of our product announcements, we wanted to have a short conversation about our vision in this space. We believe that Data Cloud is a foundation that organizations need for AI, but what are we providing on top of that so that all of our customers can truly seize this opportunity? What better way to find out than to hear from our founder, Benoit, and our new SVP, Sridhar. Benoit, kick us off, please. Yeah. Thank you, Mona. Thank you, and hello, everyone. It's really great to be here at Summit. Just like you, Mona, I am very excited about this topic, and I want to talk about AI with, really, you all. I'm especially thrilled to have Sridhar here with us. Sridhar, you have been with us for how long already? Week four. Yeah, you are new to Snowflake, but Sridhar is a veteran in AI and machine learning. Tell us how you got here today? Thank you, Benoit, Mona. I'm thrilled to be here. I worked as the SVP of Ads and Commerce at Google and built the ads business there for over 15 years. We built some of the earliest and largest machine learning systems on the planet. More recently, I founded Neeva, the world's first AI-powered consumer search engine, which Snowflake acquired last month. Now we are all excited to be at Snowflake and to bring the power of AI and data to all of our customers. We are super excited to have you. Let's dive in. Benoit, what is our vision in AI? Yes, as you know, Mona, so over the years, we have made many, many investments in AI and ML, but what is really exciting about this moment is that generative AI, for the first time, is really going to democratize access to data, and that's critical. Before, for example, you had to know how to program in either SQL or Python, you know, to create data, and only a few people had this level of expertise. Now, all of a sudden, you know, So before, you know, these experts had to program, you know, dashboards or any other things for business users to be able to access the data. There was always this friction. Now, all of a sudden, you don't need to know how to program in SQL or Python. Anyone can directly ask a question. You can talk in natural language with your data, and the Generative AI layer can translate this natural language question into appropriate queries and even visualization. The other aspect is Frank, you know, just explained how, you know, the Data Cloud brings all workloads to your data, giving you the security, governance, performance, and ease of use, which is, you know, really critical. This applies also to AI. You really want to run all your AI workloads in the Data Cloud. Really our vision is three things: First, you know, users should have direct access to data through natural language. Second, you will be able to run any model inside the Snowflake Data Cloud and optionally embed these models inside data application, that then can be distributed through the entire Data Cloud. Third, all of this will run on your data without any security or governance trade-offs. Mona, your team is really at the center of this development effort, right? Yes. Yes, we are. There is a broad effort across many teams to make this vision a reality. From giving our customers a platform to develop and deploy models, to easily accessing models automatically built on your data from SQL, to our own LLMs that understand documents, and of course, LLM-powered products that make all of our customers more productive. There's a lot to be excited about. All right, Sridhar, you have built machine learning systems at incredible scale. What about the vision that Benoit just laid out made you believe that Snowflake is the place to build the future in this space? Absolutely. Let's double-click on what Benoit said just now, and what Jensen also talked about yesterday, which is language models as the new human-computer interface. Remember that for the past 50 years, we have had to live by rules that computers and programmers set for us. If you entered a number in the wrong format, well, that's your problem. All of a sudden, we can interact with computers, with applications in natural language, and actually have them understand what we are saying. This is a huge unlock, but there's a big but. Language models by themselves do not understand what is real, what is authoritative, what is believable, what is real-time. You need to combine them with the power of retrieval, with the power of search, to set the right context. This is what we did at Neeva to launch an AI-powered search engine. It's really the combination. Now, of course, Snowflake is the trusted platform that is safe, secure, efficient for data and applications. It's the combination of these that's going to be the magic unlock for all of us. Whether it's much better catalog search, or an assistant that can help you write SQL faster, or an assistant that you can just talk to and get insights about, you know, your data, your customers, you can expect a lot from us. We're very excited. Yeah. Yeah, I agree. Thank you, Sridhar. Indeed, you should expect a lot from us, and we have already built a lot, and you are going to be really amazed by what Christian is going to walk you through, and especially my favorite part of Summit, the Summit keynotes, which are live demos, and you are going to see many of them. Let's do it. Thank you so much. Thank you, Benoit, Sridhar. Thank you. Thank you, everybody. Thank you, everyone. Thank you, folks. Please welcome Snowflake's Senior Vice President of Product, Christian Kleinerman. Good morning. Good morning, Snowflake Summit. Good morning. Good. Awesome to see you all here. Thank you for being here with us today, and I can assure you, all of us at Snowflake are committed to giving you an amazing experience at the conference. Hopefully, all of you go back to your organizations excited and inspired about what is possible with our new innovations. You heard from Frank, the broad framing of the Data Cloud. You just heard from Sridhar, Benoit, and Mona about AI and how we're thinking about it. Now we're ready to go one level deeper on actual innovations, actual new launches. You wanna see some demos? Yeah! Not really? How many demos? How many you want? 10? You're crazy. We will have some announcements and demos, and we love showcasing the actual technology. With no further ado, let's get into it. Today's announcements, we've included three different chapters on the talk. First one is around single platform, second one is around distributed, deploying, and monetizing applications, and the last one is, how do we help you program data, get value out of your data without trade-off? With that, let's get started on the single platform. If you have been with us for a long time, in the very beginning, our founders, they wanted to create internal tables where you ingest data into Snowflake, and you are able to get value out of it. Over the years, many of you gave us feedback, and you said, "I want to be able to interact with data that is in external storage." Here at Snowflake Summit, five years ago, we introduced external tables. Later on, we introduced external tables for Iceberg tables. Last year, we introduced Iceberg tables. Many of you tried it, loved it, but you also said, "I want sometimes to have Snowflake control the reads and writes and the transactional consistency of the tables. Sometimes I do not want that." Today, there are trade-offs in performance. If it's an external table, you don't get the best performance. If it's a Iceberg table, you do. Today, we have an exciting announcement for all of you, which is we do not like to make you have these trade-offs. We're announcing unified Iceberg tables as a single mode to interact with external data. The good news is it's gonna have two modes. One of it is in an unmanaged mode, where coordination of changes and writes happens by a different system, your choice. You can have a managed mode, where we take control of the data, and we coordinate writes. The most important thing is we do not wanna give you any trade-offs in terms of performance. What you see here, first blue bar, it's an unmanaged table. I'll let you figure out who wrote those Parquet files. They're not very optimized, but it's still more than two times faster than external tables. If you use the managed tables, where Snowflake is doing the writes, the performance is on par with internal. Unstructured data is another core pillar of our platform. We added support 2 years ago, and one of the things that we heard very consistently from many of you is: How do I more easily get value out of my unstructured data? Frank already alluded to, before any of us were talking about LLMs. I assume you heard about LLMs? Yeah, me too. Before they were in vogue, we acquired Applica, and what we're very, very excited to announce today is the private preview of Document AI, which what it lets you do is ask questions in natural language from documents that you have stored in Snowflake. Most important, I don't know that anyone has operationalized an end-to-end pipeline where you can give feedback to the answers you're getting from the AI and be able to fine-tune and retrain the model. What you will see is the opportunity to take fields and structure out of documents, and then you can put in a table, use it for another pipeline, use it for AI, for ML, anything else you want. The model that powers this is a text and image model, and we wanna show you how cool this works. With this, we're gonna jump into the first demo of the morning, and I wanna first introduce. Dash, he's like our demo master. Come here, show us. Look at this outfit. When I tell you that we are committed, this is commitment. Look at the shoes. Now I want to introduce Polita Paulus, engineering leader at Snowflake, and she's gonna do the demo for us. Polina, welcome. Thanks, Christian, and good morning. At Snowflake, we believe it should be really easy to use AI to get more out of your data. That's why I am really excited to be with you here today to talk to you about how Snowflake is using large language models to put your unstructured data to work. Let's check it out with a demo. For this demo, Dash and I, Dash, you're looking great, are gonna play the role of a data engineer at Ski Gear Co, a company that produces ski goggles. We want to make sure that we're able to manufacture those goggles on time with our expected volumes. Recently, we've been having some problems with our injection molding machine. Let's check it out. To understand these issues, I'm gonna use Snowflake to analyze inspection forms, and we're gonna be able to see a full history of problems to the injection molder, when they occurred, and why. To start with, I've created a project called Machine Inspections and uploaded about a year of our PDF inspection documents. This is built directly into the native Snowflake UI. I can upload some more if I need to, but I think I've got enough to get started. These documents contain a mix of fields and free text, and analyzing them is either going to be error-prone and time-consuming, or it's going to require ML expertise, which I don't have. With Snowflake's new Document AI, currently in preview, I can do this with no ML expertise required. By default, Snowflake's Document AI uses a zero-shot model, which just means I can get good results without having to fine-tune or train the model. If I need to, I can always fine-tune it to improve my results. On the right, you can see a preview of one of the documents that I've uploaded. To start extracting information, I can just type questions in plain English. You can see I've already created a few of them, like, "When did the inspection happen? Who performed the inspection?" You can notice here that Document AI is actually reading that signature. What was the inspection grade? Dash, let's add one more question. I'd like to know what part was defective. When I ask this question, Snowflake's Document AI is going to do an analysis of this document and give me back an answer along with a confidence score. In this case, it told me that the injection molder has the problem. Okay, let's flip through a few more of these documents, Dash, to see how it's performing. What you see right here in real time is Snowflake analyzing this document with its large language model, and we got an answer: the mold clamping unit. This is interesting because you can see that the inspection actually passed with no issues found. This is no problem. We're just gonna update this and tell it, "None." What this does is it provides feedback to Snowflake about how we're going to fine-tune the model. Now I can click Start training, and publish the model to my account. In the interest of time, I've already published this to my account. Now that it's in my account, I can actually share this with other teams that can use my fine-tuned model as well. With a simple SQL query, I can run this new model on all of my inspection documents in one go. Now we've got some results, and we can see that every three months, the injection molding machine has failed inspection. This is really interesting. I can share this insight with my quality engineering team, and they can use this to run maintenance every two to three months, so we can stop failing these inspections. Other data engineers, analysts, and developers at the company can use my model directly as well. That's really cool. Now I want to be able to use this to process new document inspections as they come in. This is really easy because it's fully integrated into the Snowflake platform. I can create a pipeline using streams and tasks to process these new documents, and every time a new document comes in, it will run, and I've actually set up an alert to send me an email anytime a new document fails. We can even use Document AI on text-heavy documents, like warranties and contracts. We can see here that the injection molding machine's warranty says that this part is actually good until November 2023. We might actually get some money back. To recap what we just did, we analyzed PDF documents using Snowflake's first-party large language model, then we were able to extract information with regular English questions and a few clicks rather than writing code. I was able to fine-tune the model to improve my results, then I could publish that model for anybody in my organization to use. I created a pipeline using streams and tasks to process new documents as they come in and send me an email when something fails. Join me for the What's New, Document AI, and Unstructured Data with Snowflake session to learn more. Thank you, Dash. Christian, back to you. Thank you, Polina. Is it cool? Yeah. Yeah. Yeah. From the early days of Snowflake, governance was a priority, and you saw it from Frank. Governance is a big reason to have data inside of Snowflake. Security was where our founders started on day one. Over the last few years, we've been investing heavily on privacy, which is another aspect of governance. Not only because there's a lot of regulation around privacy, but in this age of AI and GenAI, you feed some PII to one of these models, and God help you when that PII is gonna show up, who knows in which context? From that perspective, we wanna give you the most comprehensive platform to secure and protect your data, all the way from classification of the data, what is sensitive, what is quasi-sensitive, masking of the data. We have ability to do private data products, private machine learning, and of course, be able to audit the entire process. We have 3 exciting announcements for you today. First one, we're introducing in private preview, what we call query constraints, which is a policy that you can set on a data set and say, "What types of queries are allowed to run?" Maybe I don't want some columns to be selectable, which is a projection constraint, or maybe I don't want some columns to be queried without an aggregate. Very exciting. We're also introducing and integrating to Snowflake differential privacy, so you can introduce noise and track a privacy budget, again, to protect reidentification of sensitive data. Integration will start with Python, and later on, we'll do it into the core SQL engine. Last but not least, we continue to invest in our clean room capability, not only making the platform better, there is some UI, and we're working with amazing partners to deliver industry-leading privacy multiparty computation. Now I'm gonna shift to the core of the engine. To do so, to share some of our innovations, I wanna invite one of our founding engineers, Allison Lee. Please join me in welcoming Allison. Allison? Good to see you all. It's great to see all the excitement for Snowflake. One of the things that I love about being an engineer at Snowflake is how focused we are on making our customers' lives easier. This has been true from day one, and it's a key reason why Snowflake is a single product with one core engine. Whether you're using SQL or Snowpark, it's all powered by the same engine, so every enhancement that we make is applied across the board. Of course, this makes it critical that the engineering team focuses on the most high-impact areas of development, and this is why we have a data-driven approach to engineering. Since Snowflake is a single product, we can easily analyze how Snowflake is being used and figure out the best things that we can work on for you guys, without you having to tell us what we should be focused on. Once our engineering team makes an improvement, it just shows up. There's no need to enable a new feature or tweak some parameter to get the best out of Snowflake. We want it to feel like magic to you and just work. Simple as that. Of course, our performance work is never done, and this data-driven approach to engineering works particularly well for performance work. We're constantly looking at your experience with performance in Snowflake and figuring out how we can improve it and assessing the resulting impact. I'll be talking more about how we're assessing the impact a little bit later, but first, I wanted to tell you more about some of the new advanced analytic capabilities that we've been working on for all of you. Since you last heard from me, we're continuing to expand Snowflake's robust support for geospatial data. As part of that, geometry support is now generally available. This means you can ingest any type of spatial vector object into Snowflake and do your analysis on it, whether you're operating on a spherical or flat surface. Additionally, now in public preview, you can easily switch between different spatial systems. For instance, if you're switching between a state-level system and a global system. Now generally available, we support use cases like intersecting coordinates and shapes and invalid shapes. When you take all of these things together, it means that it's much easier to migrate your spatial data into Snowflake, whatever that data looks like. Another area that we're focused on is giving you the ability to work with ML models directly from SQL. Now, in public preview, we have a set of ML-powered functions that allow you to build more reliable time series forecasts, quickly identify what's contributing to a change in a metric, and detect anomalies and trigger alerts. The coolest part of all of this is that you can do that without any machine learning expertise. In addition to expanding our advanced analytic use cases, the team's been hard at work on something, which, for me, as an engineer, is the most exciting thing, and that's making Snowflake faster for all of you. We deliver performance-related enhancements with nearly every release. You might not be familiar with all of these on screen, and you shouldn't have to be. That's part of the simplicity of Snowflake. A lot of these are pretty geeky, but each of these had an impact on the customer's performance experience. However, it can be challenging to understand the overall performance impact when you take all of these things together. This is why we've developed the Snowflake Performance Index, or SPI. This allows us to assess the impact of all of the great performance enhancements that we make across the entire year. Since we first started tracking the index back last August through the end of April, so that's about 8 months, we found that query duration for stable workloads in Snowflake has improved by 15%. By stable, I mean recurring workloads that are consistent and can be compared over time. What's most important is that this is all based on actual customer usage, so this is based on your workloads. Some of you are probably familiar with industry benchmarks, such as TPC-DS. These are commonly used to analyze performance for certain types of use cases, and they definitely have their uses. We use them as part of our development process. When it comes to assessing the impact of the work that we've done and asking the question, "Have we made our customers' lives easier?" These benchmarks aren't tied to any actual customer usage, and so they really don't cut it. This is why we've developed the Snowflake Performance Index. What's most interesting to me and our engineering team is analyzing how you, our customers, are using Snowflake in the real world and making your workloads and your queries faster. I've talked a bunch about performance and query duration, but don't forget that when performance improves in Snowflake, that's closely tied to cost. That means that when we make performance optimizations in the system, your costs can go down. It's free money, as Frank would say. All right, well, with that in mind, I'm going to hand it back to Christian, and he's going to talk a bit about the work that we've been doing to help you with cost predictability and control. Thanks. Thank you. Lots of enhancements, and we work very hard to make it transparent for all of you, so things get better without effort. If you look at back at the old days, planning for an upgrade, that was crazy talk in this day and age. Allison rightly said that we are focused very much on helping you govern and manage your spend on Snowflake. None of us at Snowflake want you to overspend. We want you to be aligned with the value you get from how you're consuming resources in Snowflake, and the framework by which we enable this, we want to give you visibility into how resources are being consumed. We want to give you control and policies to manage that resource utilization, and we want to give you optimizations, and Allison just covered many of those. In reverse order, let's talk about controls. Today, I'm very, very excited to announce that the core budgeting capability is now going into public preview, which will let you specify for a subset of resources of your choosing, a budget that you wanna track against, and not only get an alert when you've exceeded the budget, 'cause that's sometimes or oftentimes too late, but also be able to know when you're on track to exceed that threshold. This is in public preview now. In the topic of visibility, I think there's only one item that we've heard very, very consistent feedback from you, and it is the ability to have a warehouse utilization metric. What this lets you do is specify how utilized is a warehouse, a cluster, in a given point in time, which will help you optimize. I hear questions from you all the time: "Should I have a larger warehouse? Should I have a smaller one? Can I consolidate?" This is the building block that lets you see all of this. Very exciting. You just heard many new capabilities and enhancements that we've done with the core platform. For us, the real reward, the real benefit, is when we hear from our customers getting that value from the innovations that we do. We want to invite one of our partners and customers onto the stage. Please join me in welcoming Mihir Shah. He is from Fidelity Investments, and you're gonna hear from him a little more. Mihir, welcome. Good morning, everyone. Good morning. This is a great time to be a data engineer, isn't it? You know, fundamental shifts in technology and how it impacts the business, comes in maybe every decade or every two decades. I believe that right now we are at the beginning of a new cycle. This cycle is led by data, so all of you are at the tip of the spear. You know, we always said, you know, when I was running architecture at Fidelity, one of the core principles of architecture was: get your data strategy right, and everything else will fall into place. This is so true now, more than any other time, that to get your data strategy right. We have always heard the phrase that data is an asset, I'm coming from a financial service industry, and the first question you ask is: What kind of asset is data? It is a liquid asset. What does that mean? MIT CISR wrote a paper that term, they coined the term data liquidity, and to describe how quickly or easily one can create value from your data assets. In your organization, can you quickly convert data into insights, or launch new data-driven products, or improve your understanding of the customer or your operations? It's not very easy. The problem is that data is trapped within silos of legacy systems and databases. The fragmentation, the lack of an enterprise data model, common taxonomy, or definitions of data, is preventing us from fully utilizing or realizing the value of the data that we all have. What I'm gonna do is briefly walk you through our journey, and how we are trying to create value from all the data that we have, and, and you'll see, you know, what it takes to do that. Fidelity is a very large organization, provides financial products and services to more than about 44 million individuals, thousands of institutions, and financial intermediaries. We have many different businesses. There's a significant synergy between them, but there's a need for each of the businesses to be independent and yet integrated, okay? Many large companies are like that. Fidelity has always been a sophisticated user of data for decades. Okay? We had multiple data warehouses to support investment decisions, finance, fraud detection, customer analytics, et cetera. At the beginning of our transformation, which started about, maybe about three years ago, we had about 170 legacy databases that served our analytics needs. Some were very, very large, but most of them were just really small databases, local databases that supported dashboards, reports, BI, power tools, et cetera. Our strategy is to build a single integrated data platform for all of Fidelity and eliminate all of our 170 legacy databases in this space. We call it the Enterprise Analytics Platform, and it's built on the Data Cloud. Just to kind of, we are kind of in year 3 of our implementation. We are in production. We have about, you know, 800 data sets or data products in production in this, on this platform. We have daily about 7,500 users using it. We have 200 applications running on this platform today. We are not done yet, we still have ways to go, maybe another 18 months or so, but we are already deriving tremendous amount of value. Here's how we derive some of the values, okay? Look at what we used to do in the old days. Take a basic data set for any financial institutions, right? Customer accounts, balances, and transaction. Just that data set. Everybody needs that data set. Finance needs it, so they have it in their own data warehouse. Marketing needs it, they put it in their own data warehouse. Risk and fraud detection needs it, they put it in their own data warehouse. What happens now, today, is that we ingest that data set once, and then we share it. Finance still gets its view, what they're used to seeing, so does marketing, so does, you know, compliance, but everybody's looking at that one single data set, and we have eliminated all the ETLs going into multiple databases, and you get a single version of truth. Okay? Now this is just one data set. We have 700 more. You know, we go through the same process. Now, the whole process of getting a single version of truth across the company, you get consistent results, consistent, you know, queries. People talk about AI and ML, but the first value that we get is we are getting basic facts about our business that we never knew or they were never accurate, right? How many customers do we have? How many employees? How much money came in today or the last week, right? These are basic facts, and you should actually know those basic facts. We'd never had it, and now we do. Because all the data sets are already in the platform, new use cases are taking us 70% to 80% less time to actually activate, okay? These are analytics use cases, data science use cases. Our data scientists are ecstatic that the data is all there, and they can spend more time on tuning their models and building new models, versus actually marshaling data and worrying about the quality. We already decommissioned probably tens of millions of dollars of on-prem hardware and appliances. As far as LLM strategy goes, we're still working on it, okay? One thing we're not worried is that all our data is already in one place, organized by an enterprise taxonomy, and whatever that strategy might be, the data is in one place. How did we do this? you know, in a large company like ours, it's difficult to execute an initiative like this. The key to success is to create an operating model that works for your situation, okay? For us, what we did was we actually debated whether it should be centrally engineered, completely distributed, how do we budget it, how do we actually work around chargebacks and security, and all the different processes that were already there in our, you know, in our organization. What we did was we created a centrally managed platform, which is managed as a product, okay? The platform team, the product team, manages the engineering, the data taxonomy, the data model, the architecture, the engine, and the enterprise standards. We also have a strict and transparent rules in place for data owners, data sharing, and all the agreements in place, okay? This allows us to operate this platform as a marketplace, okay? The marketplace, there is the owner of the marketplace platform, but there are strict rules for the data providers and the data consumers. What we are doing is, now that we have all our data in one place, and we have these processes in place to govern our data and share our data, we're extending the same concepts with our suppliers and our customers. Market data vendors provide us data, not as a feed, but as a share. SaaS vendors are now using data with zero ETL and giving data back to our warehouse as a data share, eliminating fees, okay? Very similarly, we will be extending the same capabilities to our institutional customers. Earlier you heard Frank say that before you have an AI strategy, you need a data strategy. To execute on a data strategy, you need to think about a new operating model. You need to think differently about the funding, the investments, the budgets, the organization structure. This is one of the biggest hurdles in large organizations. Everyone is aligned on the need for a data architecture, but nobody wants to change the organization structure and processes. I just showed you what we did. That may not work for you. You have to figure out what changes you want to make in your operating model. There's a clichéd phrase, saying, "Data is the new oil." Okay? It's used to describe the value of data, but I like that analogy because the investment and the effort it takes to convert crude oil into consumable product, drilling, building platforms in the roughest parts of the world, creating thousands of miles of pipelines, building. You know, building and operating refineries, that's the kind of effort you need to convert data into a consumer product and create value. Thanks to Snowflake and the Data Cloud, the technology exists, but I'll leave you with one question: Are you willing to change your operating model to enable the power of data? Thank you very much. Thank you so much to Mihir. That's warms our hearts to see the impact that we can have in our organization, tearing down silos and consolidating data. This is the end of chapter 1. Now we can move to part 2 of our conversation today, and it's all about how do we deploy, distribute and monetize data products. Can be a data set, can be a native app, as you heard. The face of all of this is our Snowflake Marketplace, and the momentum that we have is amazing, and we continue to innovate. Lots of launches, and you'll hear more throughout the conference, the ability to have public listings and private listings, the ability to automatically fulfill products across regions and across clouds. That's going generally available at the conference. At the end of the day, what really makes a marketplace like this interesting is the content itself, and we have some exciting news to share with you today. To do that, I'm going to invoke the help of AI, in person. By AI, I, of course, I mean Alex Izydorczyk. Alex, welcome to the stage. You know I'm gonna start calling you AI all the time. I was born for this. You are well known to some people, but I don't think everyone here knows you. Can you give us a little bit more about your background? Absolutely. First, let me just say that I'm thrilled to be here. The energy is absolutely palpable. I spent the first six and a half years of my career, and the last six and a half years of my career, at a hedge fund called Coatue, focused on using external data to make real-time predictions about the economy. What is inflation doing, what is consumer spending, and so on. Now I've started a new company with Snowflake's help, called Cybersyn, which is a Data-as-a-Service company that's native to Snowflake and provides content for the marketplace. That's super cool. Why Snowflake? You made a choice. Why? It's a great question. Besides just you and your charming personality, there's two reasons. At the big level, we have a shared mission. Cybersyn's mission is to make the world's economic data available and usable, to make it mobilize the world's data. That aligns with Snowflake's main mission as well. I would point you back to Frank and Jensen's note, to thinking back what the most important table in your organization that lives in Snowflake. Well, what about using everybody else's most important table, too? That's Snowflake data sharing, and that's what Cybersyn is helping to enable by providing this content on the Snowflake Marketplace. At the tactical level, Snowflake gives us the distribution. It lets us connect to all of you and lets all of you access our content in one click. That's awesome. Where are you? Is this just, like, a good idea, or do you have customers or any customer stories? We're rolling. So far, we've released a series of public domain datasets, gaps in the Snowflake Marketplace, where we've provided original content to fill those gaps. Things like inflation data, population data, so on. We've had more than 500 Snowflake accounts, Snowflake customers sign up, and some organizations, I'll call out Blackstone as an example, have used some of our datasets, such as our SEC feed data or inflation data, for their own data science use cases. I'll point out, these sophisticated organizations, it's not as if they cannot get this data on their own. They can, but with the power of Snowflake data sharing, we save you that ETL work, and that will allow you to sort of focus on the downstream value add. I'll say, I think that's not true of all your data, they can just get it. I think you're leveraging Snowflake Summit to announce some new data products? Exactly. As Frank said yesterday, not everything can be free. We are launching two proprietary products. We call these products foundations, because they're something you can build on, and the keyword is product. They're not just datasets, they're datasets and native applications in the form of Streamlit. We're launching a consumer spending dataset and an e-commerce dataset that we think will be useful to retail CPG and financial services clients. That's awesome. The question that I think all of you may be thinking of, and I'll ask on behalf of everyone here: Where else can I find your Cybersyn products? Nowhere. We're exclusive to Snowflake. Yeah! You have a booth at the conference. You and your team are here. You're going to have a session. If you want to learn more about Cybersyn, find him, learn. He has amazing data sets. Alex, thank you for being here. Thank you. As Frank alluded to in his opening remarks, the marketplace is not just data. Last year, we shared our broader vision for a marketplace for native apps. Today we're very excited that the Native App Framework is going in public preview. In the last few months, that we've been only in private preview, the momentum and the excitement is through the roof. You see some of the logos. Everything you see here is an app that is actually published already in the marketplace for more than 25 providers, over 40 apps. I was chatting with our marketplace ops team, I think there's another 80 or 90 in the queue ready to get approved. There's a lot of momentum. Some of the names here, like Bloomberg and others, it's completely amazing. We're very excited about the progress. In the last couple of years, we've been on a mission to simplify monetizing data products. Everything we did originally was around usage-based business models: billing by number of queries, billing based on time. Of course, all of you and many of our partners are so excited that you came up with 50 different business models, and we cannot just implement each one of them. Today, I'm completely excited to share the introduction of what we call Custom Event Billing, which is that for any app or any data product, you will be able to bill in whatever the right units are for your business. If you wanna do a one-time bill, per time, per user, whatever you want, usage-based or not, we will simplify the billing of whatever your model is. The other thing that we hear in the concept of billing, more on the other side, on the purchasing side, the way I hear the question or some of you express it, is, "I wanna buy data or apps with Snowflake credits." We quickly say, "Well, I don't think that's what you mean, because Snowflake credits is sort of like a unit of compute." We got you. We do understand what you mean. We are also very excited to announce the introduction of the ability to buy from our marketplace by drawing down from capacity commitments to Snowflake. This is now generally available for all of our customers in the U.S., and we'll continue to expand later on. What you can do is if you've made a capacity co- commitment to Snowflake, you're gonna be able to deduct some for apps, some for datasets, and of course, your traditional consumption of resources. Are we ready for another demo? You wanna see Native Apps? Okay. I'll say, our engineering teams did not all come here, and many of them are watching on the live cast, and they wanna hear you. Are you excited for another demo? I wanna welcome Unmesh Jagtap, product leader for Native Apps. Unmesh, come on in. Thank you, Christian. We are super excited to bring the Native App Framework to developers around the world, with the public preview launch in AWS. Today, I'm gonna show you how you can build your apps with Snowflake's highly reliable and global multi-cloud infrastructure, how you can build your businesses with Snowflake's global marketplace and flexible monetization models, and how you can deploy your apps close to the customer data while retaining full control over your intellectual property. Let's see all of this in action by building an app that predicts lead times for manufacturing orders. Dash, are you ready? All right. We want you to bring your favorite developer tools and best practices to Snowflake. I personally love VS Code, so we are going to write our application with Snowflake's VS Code extension. As you can see in this project, I have multiple Python files for Streamlit code, for Snowpark code. All I need now is a manifest file for the app config and the setup script that installs the application in the consumer account. Now, we are ready to package these code files that I've already uploaded to my Snowflake account. Let's head over to Snowsight and create an application package, which is an independent and self-contained unit of code and data that you can share with your customers without exposing your intellectual property. Before committing our code changes and creating the first version, we should test our application. With Native App Framework, it's super easy to do. All you need to do is to install the application package that we just created in the same account. Once the installation is done, an application instance is created. While this app is installing, Dash already has one ready to go. This is what our application built with Snowpark and Streamlit looks like. I think it looks amazing. Dash, what do you think? Let's commit our code changes by creating the first version. With versioning built into the framework, you can incrementally release new features or bug fixes alike. You can target releases to your customers and confidently and safely deploy changes. While this version is created, I want to show you something super important. Today, as application builders, you have to spend a significant amount of time and money getting your apps ready for security compliance. Snowflake automatically reviews every single version of publicly shared apps for security threats and abuse. We believe that this is going to accelerate the sales cycle for you and time to value for your application customers. As you can see in this example, this version was submitted twice. It failed the first time because of a security issue, which was then corrected, and now we have an application that's ready to go. Let's publish our application on the Snowflake Marketplace. For that, we're gonna head over to the Provider Studio, where we already have a listing for the application. You can monetize your applications right here in the Data Cloud without having to set up or manage complex billing infrastructure. Snowflake gives you a range of flexible monetization models to choose from. As you can see in this example, we are creating a completely custom billing model, where we can charge the customer per lead time prediction, and we'll also include a 30-day trial. Now we are ready to publish our application. That's it. In just a few simple steps, we've published our application to Snowflake Marketplace, where more than 8,000 customers can instantly discover your applications.... Let's look at what the customer experience for using and discovering these applications looks like. For that, we are going to head over to the Snowflake Marketplace. As Christian mentioned, we have more than 30 applications live on Snowflake Marketplace today. If you are an AWS, you can start using these applications now. Now I'm going to show you something very exciting. This is Snowflake Marketplace, powered by conversational search, using large language models behind the scenes. Let's see if there are any products to reduce the supply chain risk. The search returned not only data sets, but also the applications, including the one that we just created. Let's click on the application listing. As you can see, the security-related requirements of the application are called out clearly so that the customer of this application knows the security posture of this application even before it is installed in the account. I can see that there is free trial, I understand the pricing model, I can pay for it using my existing Snowflake capacity commitment. We are ready to install and use this application. At the same time, Dash already has one ready to go. It seems like this application is requesting access to some data in the consumer account. Let's grant the application the data it needs. With Snowflake Native App Framework, I can bring apps like this close to my data without leaving Snowflake, without sharing any of my data with anyone else. As application providers, you can be completely rest assured that your intellectual property is secure from app, from your customers. While this application is crunching some serious data, I would like to recap our demo. First, we built an application with Streamlit Snowpark and with the tools that developers love. We built a completely custom billing model and monetized and distributed our application through Snowflake Marketplace. We saw how the customers are going to bring their apps close to the data from Snowflake and run it against their data within minutes. If you are developers, start building your applications today. If you are an AWS, check out the applications that are live on the Snowflake Marketplace. If you'd like to learn more, join us for the What's New session. Thank you, Dash, and back to you, Christian. It's pretty cool, right? This brings us to chapter three, where we're going to help all of you get more value out of your data, program your data, without compromises on what type of programmability you can do, or security, or privacy. To start that, we hear lots of questions on, "Oh, do you understand developers? Do you have enough tools for developers?" We have a long list of announcements. I'm going to tease quickly four out of them. Throughout the conference, there's more. Tomorrow, there's a builder keynote. In no particular order, we're committed to deliver a native Python and REST API for all core operations in Snowflake. You'll see here at the conference, the beginning of preview for scheduling tasks and doing DAG operations, all natively from Python. Enhancement one. Number two, we're excited to announce the introduction of a new Snowflake CLI, completely open-source tool, focused on, excuse me, developer-centric use cases. You'll see a number of demos and instances where we can use it. Very cool for those of you that prefer CLI program models. We're introducing brand new logging and tracing APIs, both of them in public preview as of today in the conference, and you'll be able to log data and get into a common event table within an account, so you can debug stored procedures or functions, or do whatever you'd need to do to understand how your code and your application is performing. Of course, you can hook it up with alerts. Last but not least, we're very, very excited to announce an automatic synchronization with Git repositories, where code will be synced between a branch that you choose and a stage in Snowflake, so you can maintain code in sync from within Snowflake. Cool, right? Okay. Frank talked about Snowpark, and as a reminder, it's the secure hosting of the Python runtime, the Java runtime, and a number of libraries that help you program data. You heard it earlier, Snowpark only went generally available in November of last year, and with just six months or so, the adoption and the excitement is through the roof. More than 30% of our customers have used it on a weekly basis. Thank you. We run over 10 million queries to Snowpark every single day. Of course, we continue to innovate on Snowpark all the time. We have a number of exciting announcements for you. Granular control of packages. You can decide allowed list and block list of packages within your account. Maybe you separate development from production, support for the newer runtimes, and I'll let you read the rest of the list. The innovation in Snowpark does not stop. For us, one of the most exciting pieces out of Snowpark is not what we build technology, but what our customers do with it. We have a video for you to hear from our customers. Please roll the video. With our use of Snowpark at PowerSchool, we see several benefits. It allows us to bring new solutions to market much more quickly. As we build new AI ML models, and we test those, we iterate those, we refine those, I think the full framework that Snowpark provides is game-changing. It's fast, it's easy to use, and we are seeing that we can benefit our business by adopting it much more broadly. Before Snowflake, our data engineering workloads were very time-consuming. We did not have easy access to data because we had to move our data from Snowflake into another location in order to access Spark clusters to perform our data science workloads. The first use case where we've implemented Snowpark is in our real-world data space, which is fundamentally looking at how we can provide better insights and evidence associated to our patients themselves. We utilize Snowpark to analyze and engineer features for model development. Snowflake provides a unified platform for managing and analyzing data with a single governance model, enabling seamless integration between Python and SQL. Snowpark allows data scientists to use Python, which is very important for us because that is the tool that most data scientists enjoy using. We also no longer have to move data, so our data access, whereas previously it would have taken days to get a hold of the data that we need to start a project, and that makes it much more enjoyable, because rather than having to put a lot of work into trying to get a hold of data scientists can immediately start working on solving customer problems. What we're looking at doing is actually scaling out Snowpark into a variety of areas in commercial, corporate, and manufacturing and supply chain use cases. What Snowpark now allows us to do is tap into the scalable compute and storage architecture that Snowflake offers, and allow developers to do that in Python, language that they prefer. One of the main benefits that Snowpark brings to EDF to help us to achieve our mission to help Britain achieve net zero is a huge amount of scalability and speed to create new products. It's meant that we can create 12 products a year, whereas before we could only create three. This approach helps to improve data governance, scalability, and collaboration among teams working with data. Thank you to all of you that have adopted Snowpark, as Frank said, started saving money. More important, getting great solutions, great applications, great use cases. The other part of programming data is the ability to have pipelines. In particular, streaming pipelines. We've shared our direction in last year's Summit keynote. Today we're very excited about the milestones that we're hitting on these. In particular, for Snowpipe Streaming, we started the process to roll out a generally available, so we can be able to land data in Snowflake with very low latency. Once you have data in Snowflake, what you want to be able to do is transform that data. We are very excited today to announce the public preview of Dynamic Tables, so you can declaratively go and transform data regardless of the type of query that is needed for that transformation. The best way to understand this is with a demo, so please welcome Saras Nowak to the stage. Saras! Hello, everyone. As you all heard, it's important to track when these injection molders in the factory are down, because it causes our ski goggle output to drop off significantly. Let's build a continuous data pipeline that collects streaming sensor data from these injection molders and analyzes against maintenance data. This is in real time, that our quality engineers can use these new insights to make sure our uptime and throughput is improving all the time. To start off, let's use Snowflake's Kafka connector, which uses Snowpipe Streaming, generally available soon, to ingest the sensor data from these injection molders. Streaming data is ingested as rows directly into a table in Snowflake without having to land it in a separate object store first. Sensor data has started streaming into Snowflake. Dash, let's switch over to Snowsight to see the row count going up. Great. Next, we're going to use Dynamic Tables, which are now in public preview, to create a pipeline to process and transform this data to get us a clear picture of machine maintenance over time. Since Dynamic Tables are declarative, you simply define the output of the transformation as a SQL query, and you set the target for data freshness as one minute in this example, all using SQL. This particular Dynamic Table is calculating the latest number of outages of each machine using a window function, and the quality engineer assigned to maintain it by continuously joining the streaming sensor data with the maintenance logs. Dash, let's run a manual refresh to populate this Dynamic Table. This will tell us which machine is the best candidate for replacement and which tech to talk to. The refresh is now completed, so we can query the Dynamic Table to see the results. Seems like machine 3 seems to go down most often. Next, you'll create a second Dynamic Table that reads from the first and joins against machine location information. This will tell us which factory line has the most outages and requires better maintenance schedules. Looks like factory line 73 has the most effective maintenance techniques, and it should be the model for other factory lines to improve their maintenance schedules and increasing our overall throughput. With that, my pipeline's ready. three things to note here, the results of my pipeline will be automatically and continuously refreshed as new data arrives. Because the data is materialized, it is always fast to query. Second, with built-in incremental refresh support, my pipeline will only process data that has changed, helping keep my costs low. Third, I don't need to handle any complex pipeline orchestration or manage any dependencies. It just works. These streaming pipelines are now critical for our business, observability is key. We can track and monitor Dynamic Tables in Snowsight and quickly diagnose and resolve any issues. We can look at the refresh history for any Dynamic Table to see the current data freshness metrics over the last 24 hours, and the status of each refresh, and the data processed. We can switch over to the Graph tab to view the graph, to see the dependencies, and use that to troubleshoot any pipeline issues and make it super easy. Let's say we want to track all our factory efficiencies. I could write the queries to do this myself, but what if I didn't have to? Instead of writing SQL, we'll use Snowflake's new text-to-code capabilities currently in development. I want to create a Dynamic Table to show me the efficiency of our machine output with respect to energy consumption. I can simply use a comment in a worksheet in Snowsight to ask the question, and Snowflake will use large language models, or LLMs, to automatically generate the Dynamic Table SQL for me. Just like that, the text-to-code capabilities, it delivers the DDL for a Dynamic Table that will answer my question, all without me having to write any SQL. With that, let's recap. In this demo, we first ingested streaming data using Snowflake's Kafka connector, using Snowpipe Streaming. We used SQL and Dynamic Tables to continuously join that streaming data with maintenance logs, with results refreshed as soon as new data streamed in. Third, we showcased our full observability of these pipelines on Snowsight. Finally, we used Snowflake's new LLM-powered conversational experience to create Dynamic Tables without having to write any SQL. If you want to learn more about Snowpipe Streaming and Dynamic Tables, please join us at the What's New session for streaming in Snowflake. Thanks, Dash. With that, back to you, Christian. Thank you, Saras. An LLM that does Dynamic Tables, that's cool. Now let's talk about AI and ML, 'cause we've not talked enough about AI/ML. By the way, a lot of what you've seen so far is applied AI and ML, but also we're committed to being a platform for all of you as a customer or as a partner, to be able to build solutions. We want to support the entire life cycle, whether it's feature engineering, training, scoring, measurement, we want to support it all. We have more exciting announcements for you today, specifically around feature engineering, feature training. We heard we want simpler libraries, simpler program models. The same thing that we did with the DataFrame API for Snowpark, how about we do that for AI and ML? That's exactly what we have for you today. Announcing in public preview two new libraries. One is around pre-processing, so you can do one-hot encoding and data preparation at scale with the same engine that you heard Allison talk about. Number two, the ability to do training with distributed algorithms inside Snowflake, with some of the most common algorithms, things that you would find on scikit-learn or XGBoost. Also, we've heard feedback from all of you. "I got a model. What I do with my model?" We're very excited to announce the introduction of the Snowpark Model Registry, where you're gonna be able to store models, of course, be able to publish them, discover them, and probably most important, deploy them, whether it's in a number of runtime environments, but we'll make it easy for you to manage the end-to-end life cycle. Cool? Yeah. How many of you have trained models that never go into production because it's too hard to put in front of your users? This is where Streamlit has been shining for years now. The types of experiences that show up in Streamlit's community cloud are amazing, and it's the fastest way to productionize a model. By the way, the innovation on Streamlit also does not stop. We're very excited to announce the availability of editable data frames. Now you can do both input and presentation of data in Streamlit apps. A beautiful column configuration, where you can specify, do you want a trend line, a chart, a checkbox? How do you present data on a editable data frame? We're also working on LangChain integration, so you can now be able to see or show the thoughts and steps of LLM agents. Exciting. The one that I'm most excited about is the introduction of a brand-new Streamlit chat component. The team raced against the clock to make it available to all of you today. I don't know if any of you have seen, but conversational AIs are coming. Conversational experiences and AI apps are coming, and if you're reinventing the wheel, building a chat interface, check out Streamlit's new component. There's a brand-new wheel working very well for you, open source, for you to build amazing experiences. We have more than 6,000 LLM-powered Streamlit apps in the community cloud. It's amazing. It is the fastest way to go and show amazing results to your organization. I would encourage all of you to go try it. Last year, we teased, it was just, like a early demo, how am I going to host a Streamlit app securely inside Snowflake so I don't have to do a separate hosting? We're very, very close to public preview. What you're going to be able to do is create an app, a Streamlit app, and host it, and have that be made available to your business users, dramatically shrinking the time to productization of your AI and ML efforts. The public preview, as I mentioned, is starting very soon, and even with the private preview, we've seen over 2,000 Streamlit apps being created and published inside of Snowflake. This is quite amazing. When we did Snowpark, we said, "Java and Python, and we should be good." Boy, did we learn about different programming languages in the world. What about hosting Rust and Kotlin, and C++ showed up? Actually, there's lots of interesting apps in Superfast. I wanna host that inside Snowflake. What we wanted to do is accelerate time to value. We wanted to be able to support more runtimes, more languages, more libraries, and the way, the fastest way to do it was with the introduction of Snowpark Container Services. This is now in private preview, and what it lets us is host a Docker container securely inside the governance perimeter that Frank spoke about. With Snowpark Container Services, you'll be able to export jobs. Imagine a procedure you can run or a function that you can call from your SQL statements or Snowpark applications, or for the first time ever, you're going to be able to have long-running service inside Snowflake. The other thing that all of you said when we talked about containers is, "I'm going to need more instance flexibility." As part of Snowpark Container Services, we have a much broader list of what type of hardware you can run, how much memory you can use, et cetera. Of course, in this day and age, there was a very specific type of instance that was needed to be supported, and that is the support for GPUs. Yes! You want to see Container Services? Yes. Okay, Jeff Hollan, product lead for this effort, come on into the stage. Thank you so much, Christian. I am so excited to be here with you all today. We're going to show some incredible things with you now here. As we're building these ski goggles, any machine failures can be very costly and disruptive. I'm going to use Snowpark Container Services, now in private preview, to train a model to help us predict and prevent those type of machine failures. Now, with Snowpark Container Services, I can easily run any code or any container entirely in my Snowflake account, but I can also run third-party containers as part of a Snowflake Native App that I installed from the Marketplace. Let me show you what I mean. When I build models, I love to use notebooks, and Snowflake partners like Hex, have some beautiful notebook experiences. For me to use some of my Snowflake data and have it go to any third party, often requires I have to go through a number of different approvals, but not anymore. With Snowpark Container Services as part of a Snowflake Native App, I can now run full multi-container apps like Hex entirely in my Snowflake account. There is no additional infrastructure that I have to manage, and all of my data and processing stay entirely within Snowflake's secure and governed boundary. Dash here has installed Hex from the Snowflake Marketplace. This might look just like the Hex you know and love, because it is, but check out the URL up here, Dash. This entire experience is being hosted from my Snowflake account. Incredible. For fun, why don't we go ahead and ask Hex Magic, which is Hex's LLM-powered assistant, where it's running. You can see even Hex knows how awesome it is that it can run and be powered by Snowpark. We've got the experience. Let's now get our data ready for training. I'm going to use Snowpark Python DataFrames to create a set of rolling window aggregations on machine temperature for these machines. You can see here, Dash has a cell that's going to give us aggregates for week, month, and year. Once I've run this Snowpark code, I'm going to join that with the historic features and data that Sarah's just streamed in in the previous demo. With both of these steps, my table is now ready for training. With Snowpark Container Services, I can speed up the training of this XGBoost model with the integrated NVIDIA AI platform. That includes NVIDIA GPUs and secure end-to-end software as part of NVIDIA AI Enterprise. We can go ahead here and choose to select the data science libraries included with NVIDIA RAPIDS and NVIDIA AI Enterprise, and I'm going to do this training on GPUs. Now, for the sake of time, we did run this training a few moments ago, but you can see here we were able to train a 50 million record data set in only 17 minutes with the power of the integrated NVIDIA AI platform. Similar tests took 10 times longer without NVIDIA acceleration, and all of that speed boost actually results in a 2x cost savings. Minimal changes for me as a developer, and I get this massive boost in productivity. We've got our model, I want to take it to production, but I need to make sure it's secure, governed, discoverable, and observable. With the new Snowpark Model Registry in private preview, I can do all of this in a scalable way. Here's the Python code that we can run to now register this machine learning model into the Snowpark Model Registry. It has all of the necessary metadata, so now any team in my organization can go and find this model, pull it in, and do inferencing on our data to make sure we're predicting and preventing those machine failures. Let's recap some of the awesome tech we just saw. With Snowpark Container Services, I can easily deploy and run full applications like Hex entirely in my Snowflake account. I use Snowpark Python Data Frames to query and process the data. I was able to speed up model training 10x with the integrated NVIDIA AI platform. Finally, we deployed this model to a central registry for secure and scalable MLOps. If you want to learn more about some of these exciting announcements, be sure to check out our sessions, What's New: Snowpark Container Services, and What's New: Snowpark MLOps during the week. Thank you so much, Christian. Back to you. Pretty cool, right? Jeff is going to hang out because whatever I say, he makes it look awesome. This is how I want to make all of this come together from what you heard from Frank and Jensen yesterday, today, which is now hopefully with the different pieces you've heard about this morning, you understand how we're aiming to be the platform for AI applications, the platform for assistance, platform for copilots, and we have all the different elements. Most important element is data. You have your data in a secure platform, and we want to help you be able to do inference as well as fine-tuning of these large language models with safety, security, and privacy in mind. We will support some of our own models. You saw Document AI. We'll support a number of partner models, and we'll let you bring your own model, open source or otherwise, and you can create beautiful, amazing experiences with Streamlit. That's the broad vision, and hopefully, you will be excited about what's possible. Today, we're incredibly excited about the partnership with NVIDIA. Again, Frank and Jensen, that was a treat for all of us. It is not just what you saw from Jeff a minute ago, it's also some of the language models that NVIDIA has. Most important, we will use Snowpark Container Services to integrate the NeMo LLM GenAI framework that helps you do training and fine-tuning and detuning of models, again, with your data and the safety and privacy of the Snowflake context. We're also very excited about announcing partnerships with AI21l abs as well as Reka, industry leaders in having language models. That's what I said. If I say partnership, that doesn't sound that cool, but Jeff is going to make it look awesome. Jeff, take it away. Ready, Dash? Ready for one more? All right, you guys ready for a little bit more? Yes. All right, this is big. For the last piece of our Ski Gear Co that we want to do is around quality control. Now, this is actually a really difficult problem. We have a number of different product lines that we build. It can have a variety of different product issues. We actually want to build an experience that can be interactive with the quality supervisor, so we need a super powerful model to do this. We're going to use a large language model or LLM. Using the Snowflake Native App Framework and Snowpark Container Services, accelerated by the NVIDIA AI platform, third parties like Reka can now package their leading LLMs into a Snowflake Native App. This lets me install it in my Snowflake account that's secure for both of us. For Reka, their model, weights, and all of their proprietary IP is never exposed to me as the consumer. At the same time, for me, all of my Snowflake data that I use to interact with the LLM or even for fine-tuning, is never visible or exposed back to Reka. All of the data and processing stays entirely in my Snowflake account. It's a win-win for both of us. All I have to do is in the Snowflake Marketplace, go ahead and install this native app from Reka. You can see Dash is installing this now. It's spinning up that entire LLM app in my Snowflake account. This app comes bundled with a Streamlit interface that I can use to interact with the model. You can see here there's an image. Reka's model is multimodal, it works with both text and images. I've given this app secure access to my Snowflake stage that has images from our production line. Now, to help with this, we're going to interact with the model running now. We're going to do it without any fine-tuning. Let's just see how far we get without any fine-tuning. We'll start by just asking an easy question: What is this an image of? You can see it's your time back. Hey, these are ski goggles. Actually, a trick question. I noticed a few of you here thought those were a virtual reality headset. You're forgiven. You're forgiven. The LLM got it right, but let's try something a little bit more advanced. Can it give us any quality issues with these goggles? You'll see it's identified there's a few scratches on the lens. Now let's ask something very specific. Let's ask us to give it the model ID or number for this product. You can see it responds back, it doesn't have that information. It makes sense. That's our own internal nomenclature that we use for our products and our product lines. With the power of Snowflake, I can fine-tune this model securely with my data and make it even smarter. Dash, let's switch to the fine-tuned version of this model. We'll go ahead and ask one of those questions again. Let's ask if it finds any quality issues with the image. Here it's even a little bit more specific, a few scratches in multiple locations. Now let's ask that more specific information. Can it give us more details about this specific goggle? This is amazing right here. That is our model ID. That is from my Snowflake data that I was able to integrate with this LLM from the Snowflake Marketplace entirely in my Snowflake account. This info is just what I need as a quality supervisor. I can now follow up on this item and take any necessary steps to make sure our ski goggles stay great. To recap, I was able to install an entire LLM-powered application directly from the Snowflake Marketplace and run it in my Snowflake account next to my data, without compromising any security or governance of that data. The application was built entirely with Snowflake. Reka's third-party LLM was hosted using the NVIDIA-accelerated Snowpark Container Services. We have a Streamlit interface that we can use to interact with the model, and the entire thing is packaged as a Snowflake Native App. If you want to learn more about using LLMs in your organization, be sure to check out the session, "Unleashing the Power of Large Language Models with Snowflake." Thank you, everyone. Back to you, Christian. It's quite exciting, right? Yeah. Yeah! What's possible is now orders of magnitude bigger than what used to be possible before. You saw a lot of AI, ML. We can extend database capabilities, we can build applications, host applications, languages, GPUs, orchestration. This came together, the private preview, a few weeks ago, maybe 5, 6 weeks ago, not long. We said, "You know what? If any of our partners or customers is able to take the preview and build a real use case, not a demo, but like a real thing that we can go make available to our customers, we wanted to show it to you in the keynote." We thought it was a very small amount of time, so it was unclear if it was going to happen, and we ended up in a little bit of a problem. More than one partner showed up with an amazing solution. We were chatting internally, and we said, "Maybe we're going to show one, maybe we show two." We said, "Maybe if we show two demos, we can show three." You know how that line of thinking goes. At some point, the production company was like, "When I called someone earlier, like, crazy for talking about 10 demos," like, yeah. We had a similar reaction internally, externally, on trying to show too many demos to you. We are so excited about what is possible with Snowpark Container Services, that we have a few more demos for you. In particular, we have 10 more demos for you. Live, there's no recording, so we need the demo gods to be really kind to us. Far, they behaved. Are you excited and ready to see 10 applications of Snowpark Container Services? Let's do it. Demo number one. All of you familiar with Airflow? We have Astronomer, the company behind the Managed Airflow service, and you're seeing Airflow orchestration happening in Snowflake. Alteryx, company well known for analytics automation platform, and what you're seeing here is advanced analytics, workflow, all pushed inside of Snowflake. SAS, company that needs no introduction, enterprise analytics for a long time, and what you are seeing here is the ability to publish and deploy scoring of models again, inside Snowflake. Dataiku, many of you are customers. Company known as the everyday AI company, and here you see deployment of models to be able to do inference at large scale. Hex, amazing notebook. You saw it earlier in the demo. The entire UI, the whole product, is running inside Snowflake securely. Your data doesn't come out. Support for SQL, for support for Python in Snowflake. NVIDIA. You know now that we're great partners with NVIDIA, here you see some of the NeMo LLM models and framework running inside of Snowflake again, with a Streamlit UI. Did any of you order a vector database? We have some Pinecone inside of Snowflake for you. How about some spatial analytics? Big part of the modern data platform, CARTO, the entire UI, in this case, fleet optimization inside Snowflake. Weights & Biases, many of you know them, company to do MLOps, very, very important to be able to trace and evaluate LLM models, again, in Snowflake Container Services. Last but not least, if you want a graph database inside of Snowflake, RelationalAI running inside with fraud detection. You want to see them all at the same time? 10 demos running concurrently. Give it up. By the way, we have many more amazing partners and customers do the integration with Snowpark Container Services. Thank you to all of you. Like, wholeheartedly, I would have loved to have, like, all 30-something demos. It was just too crazy, but you are in the pavilion, you're in the booth. Go visit them. We have amazing capabilities running inside of Snowflake. Cool? This brings us to the end of chapter three, but this is the beginning. Not only the beginning of Snowflake Summit, but the beginning of a new era of new types of business logic, use cases, applications, services, running inside of Snowflake. Hopefully, all of you are a little bit in the, "My head is spinning on what is possible now. How do I simplify my environment? How do I consolidate my environment? How do I get more value out of my data? How do we collectively change what is possible with data?" We're extremely excited. Thank you for joining us. Have an awesome Snowflake Summit. Bye!