Snowflake CEO on Winning the AI Arms Race

Hey, Matt. Great to be here.

Thank you for having me. I think tumultuous is appropriate. I feel like we are in day two of the pandemic, where we all knew something was wildly different, but we didn't quite know what.

So it has that sort of something's big, but we don't quite know, kind of feeling to it.

I will take that all day long.

Yeah, Snowflake had a very different and cloud-centric view of analytics. That was the origin of the company now 12 years ago. And the core thesis was actually quite simple. It is that in a cloud computing environment, you could scale storage and compute, and implicitly memory, independently. Now you think, come on, that's not a big observation. But if you go back to how computing has always been purchased, it's always been purchased as boxes, which has meant fixed trade-offs between compute power, memory, and storage.

And if you didn't like any one aspect of these three dimensions, generally that's too bad. You would wait some more years to do your next upgrade. Snowflake basically created an analytic product that used the near-infinite scalability of cloud storage to create this amazing analytic platform where, if you wanted to do new analysis, deeper analysis on something on a whim, you could spin up a bunch of new machines, you could run this analysis, decide if you wanted to pursue it or not, shut down those machines, and go on your merry way.

That was the kind of stuff that it unlocked. And from very, very early on, it also used the fact that cloud storage is basically infinitely scalable, up to the level of many petabytes within a single deployment, which meant that what you thought of as different departmental databases suddenly had connectivity. You could take this data, let's say from Salesforce—I do this all the time, daily—and this other data from Workday, and yet other consumption data about customers from Snowflake, and join them, getting a single view that's as easy as just writing a single query against it.

So all of these were what contributed to the massive success that early Snowflake was. And the key thing that the company added was collaboration.

But it was again based off of an observation that most collaboration then, sadly even to this day, is people shipping files around either in email or by FTP drops. And the sophisticated people do SFTP drops. But Snowflake Collaboration was all about real-time data sharing, no data copying, data synchronization, cross-cloud. And what used to be an IT project, if you wanted to get data from a partner, turned into they can configure a screen in Snowflake. A very non-technical user, mind you, can say, I want to expose this data over to this customer, and I'm going to send them an invitation for a particular user.

As soon as that happened, data started flowing real-time cross-cloud. You didn't need to know where they were. It just showed up at your dataset. As I said, the magic again is it was continuous. That was the early inning of Snowflake, plus very clever and thoughtful go-to-market motions. We very deliberately went after asset managers as our early customers. Why? They're more progressive. They want to make more money. They're earlier adopters of technology and have less regulation to deal with, say, compared to banks, which just have—it's more complex.

Similarly, we formed key relationships with data providers, whether that is a State Street or a BlackRock or the New York Stock Exchange, because they were supplying data to everybody else. And so that's how we became a powerhouse that, to this day, powers much of financial collaboration. This is all almost pre-IPO Snowflake. And of course, it had an amazing IPO.

Absolutely.

But they're different slices. We also had a bunch of cloud-native companies, to this day, whether it is people like DoorDash or Coinbase or Instacart. They continue to be big customers of Snowflake. We had a big pop with everybody else during the pandemic, things like that. But as I said, this is very thoughtful go-to-market motions powered by some brilliant people that knew the finance industry, about how to drive progress there, how to create value, how to create more of an ecosystem of companies working with each other.

That's right. It is hard, don't get me wrong. It's still difficult. But the point is, there is still structure to these companies, and you can exploit that structure to structure your go-to-market. If, for example, you're a cloud provider, a cloud-native solution, and you're banging your head against a company that is still mostly a bank that is mostly on-prem, good luck to you. That is really hard. But there are people that are more progressive, and this is where ecosystem knowledge is very helpful.

Again, as I said, that is early Snowflake. And the next inning, which is still ongoing, was things around how do you create applications that run on top? As you know, there is a little bit of a hierarchy of needs when it comes to collaboration. You can share raw data; it creates some value. On the other hand, if you can share a predictive machine learning model that uses the data to predict something, you will likely get more value. If, on the other hand, you create an application that's on top of this data, then you create even more value.

That was a little bit of the third inning of Snowflake, which continues. These are what we call native applications. And what is magical about them is this ability to combine data from both provider and customer to drive new insights.

So Streamlit was more of a rapid prototyping platform. Yes, Streamlit was the Vercel of that time, but for developing front ends for applications, including native applications. But as I said, the magic with native applications is more that we have one from T-Mobile, as you know, biggest phone carrier in the country, 150 million customers. They created a data cleaning application that a different Snowflake customer can use to augment their data with the latest email address, with the latest address, but done in a way that that customer does not really get their hands on T-Mobile's database.

But T-Mobile doesn't see those customers as well, sort of do it in a privacy-safe way. And this actually has applications in the AI space that we will talk about. Now, when you're talking about agents and what actions can you take, this becomes an interesting primitive. But in all of these, Snowflake remained a database-centric company. Now, I worked at Google for 15 years, and what you learn over time is that there are very different kinds of software engineering that happens in what looks like a monolith.

The way in which you need to think about developing a core piece of infrastructure software, one that needs to run at incredibly high scale, especially if there are massive penalties to making mistakes. At Snowflake, in our storage layer, there are incredible penalties to making mistakes, which means you messed up customer data. It's an unforgivable sin. And so you have to be thoughtful. Similarly, when I ran the search ads team, getting those five nines of availability actually meant $100 million more because that thing was running at such scale.

Google.com was driven by entirely different elements of scale. There, it was all about how do you get developers to develop new advertiser experiences as quickly as you can and iterate rapidly on that. It was all about developer productivity. It was all about user workflows. The kind of people that loved that were completely different from the people that did infrastructure. You needed to figure out—we couldn't even get them promoted early on because all of the infrastructure people are like, wait, you run a five QPS service?

No promotion for you. But these were people whose brains were wired differently. Similarly, the ads quality team, which was the one that decided which ads got shown by this high-throughput infrastructure, completely different animal. They were the iterative people. Five percent improvement in RPM? What kind of deep analysis do we need to do to figure out what are long-term, short-term trade-offs for revenue versus quality? And the reason I say that in the context of the Snowflake story is Snowflake mostly made all of its early mark as an infrastructure company.

And so there is a tendency, if you're a very successful infrastructure company, to think of everything as infrastructure. And so part of the difficulty that Snowflake had with machine learning and AI was they brought a similar mentality. We're going to write a design doc, it's going to take us 18 months, it's going to be great. Opposite of what you need to succeed in iterative environments like machine learning, where it's all about creating a platform for iteration. It's all about iterating things very quickly.

As you know, there's a new winner in AI every six months, and it's often a different one from the one that was there the previous six months. And so that's a lot of the sensibility that we have brought. My co-founder Vivek, who worked in ads quality for a long time, as did I, to Snowflake, we have a much more iterative approach. And you have to give a lot of credit to our founders, to the team, for embracing us and then saying, how do we become a more nimble platform?

Because this world of AI is moving so quickly. And a lot of what I'm trying to do at Snowflake is drive this sense of: you need to adapt to a changing world. You need to be stable where you need to be. At the end of the day, if we go down for 30 minutes, for example, I will get calls from a whole bunch of customers who basically tell me, listen, I get called by the regulator who questions my data strategy if I'm a stock exchange.

I don't get to give any excuses for why my availability is not very high. How do you simultaneously satisfy that while being a rapidly advancing company in the world of AI? That's the challenge that we have had to cross. But I feel very good about where we have come as a company in terms of what we have accomplished with AI, for example. And we will sort of get more into it. Our vision going forward very much is: we are an AI and data platform.

The idea is that being able to do meaningful things with data is a core skill that is distinct from, say, providing compute, from providing an amazing cloud storage system. It's a higher level of abstraction and stands somewhere between creating an application and creating a pure platform. And that's our claim to fame, and that's what we think we do exceptionally well. And we think this is a category that is also going to grow and expand over the next 10 years as cloud computing itself expands.

And that's the market that we are tackling, and how we have evolved to this. It's a combination of infrastructure strength, but also being able to iterate quickly where we need to.

Absolutely. I think, again, this is where I remember having conversations with Benoit, our founder, about this, where I told him, for any person to even understand what makes for an excellent machine learning engineer, machine learning scientist, or an AI scientist, that itself is a skill. Somebody that's very good at infrastructure cannot tell if somebody is amazing at AI. But it is this basic understanding that there are aspects to software that are just very, very different. And you will see this even with the founders that you invest in.

There are some founders that will show up that are very good at product iteration, that will rapidly know where to get to. While if you're building a core piece of infrastructure, which is all about the simplest interface to some hard, complicated problem, but solved exceptionally well, that's a different breed of people. But to be successful with any large platform or enterprise, you have to figure out how to recognize each of these skills and bring the best people in the world, make them feel valued, and go deliver on the solutions that you need to create this holistic product.

Yeah, this is a great question. And it also goes to your other point about how have we evolved as a company, especially over the past few years. Snowflake came of age as a proprietary data format company. What I mean by that is, we have a data format. It's called FDN. The name doesn't matter, but basically only Snowflake can read and write that format. And it's been very helpful for us. We drove an enormous number of innovations. There are features that to this day blow my mind and blow our customers' minds in terms of what is possible with core Snowflake.

You can write this very funny query where you can ask an analytic question: how many users did I have? That's easy enough. You can translate that into SQL, but you can ask that question with a time reference. You can say, how many users did I have as of February 1st, 2025? This is the magic of our custom format, where we not only remember what the latest version of the information is, but we keep a way to synthesize every possible time point or the previous N number of weeks or months.

And so we got a lot out of this format. But increasingly, what is happening is that the most progressive of the CIOs and the chief data officers, what they want is independence with respect to vendors, including Snowflake. So they want their storage, their data, to be stored in vendor-neutral formats, in open formats. They're very happy to work with Snowflake. On the other hand, let's face it, these are companies that have been around for 50, 60 years. They have seen Teradata come and go, they have seen Hadoop come and go, and there is a desire for, I don't want my data locked up in systems.

It is rational. It's also a movement, just like open-source software and open-source models is a little bit of a movement. And we have had to adapt to it. And as part of that adoption and adaptation, we started supporting Iceberg as an open format. Again, the name doesn't matter. The idea is that data is going to be written in a format that anyone can read. And Iceberg is an open format in the sense that no one company controls it. It's an Apache project with a set of contributors that come from many places, from Netflix, from Snowflake, from Databricks, from AWS.

But it is vendor-neutral because of that. And it provides a certain degree of protection for all of our customers: hey, I will be able to read this data five years from now, regardless of which compute engine we use. And Snowflake initially approached this as mostly a defensive move. We had a proprietary format, it was amazing. We had innovated on top of it. Why would you agree to an open format? But our big aha moment with this was that we realized that it unlocked many, many things that were impossible for us.

The way people typically use Snowflake is they would get data from source systems, whether it's a database or one of the apps you have on your phone that talks to a server that then writes a Kafka queue, a set of messages, and then it would get processed, it would get augmented, it would get de-spammed, then it becomes high-quality data that can be used for analytics. They would bring that data into Snowflake. The bulk of that business is data processing at what we call the gold layer.

But open formats now have opened up a ton of new opportunities for us to participate in earlier stages of the data lifecycle. So we acquired a company called Datavolo. It has 120-plus connectors. And what that lets us do is bring both structured as well as now unstructured data into Snowflake. We will offer Google Drive, we will offer SharePoint connectors. It opens up what is possible with Snowflake. So tying all of these and AI together, what we mean by AI Data Cloud is we want to be that complete data lifecycle platform for you to ingest data, bring it either into cloud storage in open formats or into Snowflake, run data engineering on top of it, clean it up, make it high quality so that you can run high-quality analytics, high-quality machine learning on top of it, have it feed back into your system so it informs how your system should behave based on observed user behavior.

And then also use AI, for example, as a new consumption layer where you now can have chatbots. I have several of them from Snowflake on Snowflake data that I use on a daily basis for things like searching over all of our education, our enablement material, or how did Snowflake revenue do? That's what I mean when we call ourselves an AI Data Cloud. We help you with every aspect of the data lifecycle. That's the overall vision.

Obviously, this is a very big space. We work with a number of partners in each and every one of these. We have a ton of partners like Fivetran, dbt, Matillion, that specialize in data extraction, in data engineering. We work with the likes of Atlan, for example, for catalog support. Then we have a ton of partners that create native applications that run on top of Snowflake. Obviously, we run on top of the hyperscalers and we have product partnerships with the likes of Anthropic and OpenAI.

So it does take a village to create value for the customer, but we are thrilled to be that reliable platform on which all of these things rest.

We said we wanted AI to be a core part of Snowflake. Roughly, the way we did that in practice was, we said we will host a model garden inside every Snowflake deployment. Snowflake essentially runs in what we call deployments, which you can think of as a point of presence in every major data center that AWS and Azure and GCP have. These are instances of Snowflake that are running everywhere, but it's a single cloud. It's completely connected. Data can move seamlessly from one place to the other.

And we run a model garden inside each of these deployments. And we offer a set of products on top of it. That's the name Cortex. Cortex AI is an umbrella of products. In practice, what this means is anyone that can write SQL or Python in Snowflake can use language models just as part of their data processing. If you want to do sentiment detection on customer feedback you have in Snowflake, it's as easy as calling a single function. Similarly, if you want to create a chatbot on unstructured data, you create a Cortex Search index, and then you use Streamlit, which we talked about, to create a user interface and an application that you can deploy.

And Cortex Analyst is a structured data solution. It's the idea of being able to ask a question about a structured dataset that sits in Snowflake. And Snowflake Intelligence is the superstructure on top of all of this that helps with coordinating different elements. It's our agentic platform that stitches together these components into larger workflows that can happen. All of these fall under the general umbrella of Cortex for our AI products.

One is for unstructured data, and the analyst is for structured data.

I would say the aspects that have informed how I think about architecture and product, first of all, is with respect to elements of expertise that I talked about. How you think about delivering analyst-facing, end-user-facing functionality is a specialty unto itself. And so we are moving on to a modern micro-frontend environment powered by things like Node, with rapid iteration, half-hour deployments, so that we can develop UI faster. That's one aspect that is learning. But on the other side, Cortex Search is almost completely based on the Neeva search index infrastructure.

Asim, who's one of our most amazing engineers, called that system Koala. And Snowflake bought Neeva and so inherited all of that IP, and that codebase is still there. I can spend an hour just talking about the Cortex Search infrastructure. It is all SSD-based, and it offers the option, for example, to swap segments out so you can run massive experiments if you want, very, very quickly. What happens when you run search systems at the scale of Google is experimentation, quick loading of new data becomes a big issue.

And so there's a lot of cleverness that Asim put into how do we experiment, how do we put new bits of data into how search should work, how do you make it easy to try different aspects of search quality, how do you tune this? And so that has been very, very heavily influenced by systems like Mustang, which are the basis of how Google Search works. And so there is a direct line over there. Cortex Analyst is more interesting. It is a combination of our expertise as both AI people, but also with the intuition of an analyst.

Now, I've spent literally thousands of hours doing analysis. I've written so much SQL. And it's something that I really enjoy. I love playing with data. I love creating dashboards. And I knew that giving business users access to data without needing to create dashboards and without needing analysts to write SQL, it's been a dream. And Cortex Analyst was based off a simple observation that, in a business context, we are much better off with a very high-precision, modest-recall kind of product.

Most AI systems, including ChatGPT until very, very recently, for example, optimize for trying to answer all your questions. You ask any question, it's like ChatGPT will give you sometimes a very confident and completely incorrect answer. It is completely unacceptable when you're working in the world of structured data. If your CFO comes and says, hey, how much revenue did I make yesterday? That's not a matter of any subjective opinions. There is one factual answer. And so there is a lot of influence with how we did things before, but we also borrowed from the principles that drive Snowflake's success, which is precision really matters when it comes to structured data.

By the way, there is a lesson there for things like the deep research tools that even ChatGPT offers. There was this super cool interview between Ben Thompson and Benedict Evans about ChatGPT Deep Research, where he comments about how bad it is when you ask quantitative questions. Why? Because it consults the web, and the web often has really bad quantitative information. And that's where, again, that reinforced the need for high-precision access to structured data within Snowflake as a powerful tool for how do you drive the next generation of applications that you'll build with AI.

Early 2019.

That's correct.

Yes, two-ish years too early to market. If you were very—

That's fair. That is fair. We started along similar lines. The core thesis, again, was very simple, which was there must be a better way to do search. I'd worked on search and search-related products for 15 years, and I had come to the conclusion that if you start with things like, I am going to offer web links, I am going to show ads on top of those links, there was a limit to how much utility you could drive out of that product.

I'm proud, especially looking back at how well we ran Google Ads, especially search ads with that focus on quality. If you compare more recent platforms for search platforms, for example, App Store searches, they're truly terrible. Even when you ask highly precise questions, they will show you completely irrelevant ads. We were proud of the work that we did in search ads, but I also knew that the model had its limits. There's only so much you can push something, and the pressure to make money is always there.

You're always tempted to take that extra line of pixel because it was going to make you so much more money. In many ways, Neeva was a pure intellectual exercise of search is a really important function and we should be able to rethink it, but we didn't have the tools. I think we were two years too early in terms of creating, for example, a truly conversational experience. As soon as GPT-3 came out in 2022 and we began to understand—remember, in much of early—actually, it was '21 December.

In much of early 2022, you had to go to their silly playground and enter commands for GPT-3 to get answers. But we would start doing things like, hey, can you write what's called an abstractive summary of a page full of information? One of the hardest problems that there is. There is a technique called an extractive summary, which is you have a piece of text, you ask a question, but the question's answer is there in the text. The model just has to fetch that answer out.

Much easier problem to solve. If, on the other hand, you had like a 1,500-word blog and you told the model, in three sentences, in your own words, tell me what this blog is about, three, four years ago, this was an impossible problem. You just could not get sufficient quality. And GPT-3 was showing signs of actually solving problems like this, which is one—that was our light bulb moment for, hey, now we can actually create a search product that is all about summarizing a great answer from the best pages on the internet in order to answer your question.

We had working prototypes by mid-2022. And the launch of ChatGPT in December 2022 basically spurred nonstop work. I almost stopped sleeping, and we were cranking on what does the solution look like? And we launched sort of the true AI-powered search engine on January 6th, 2023, roughly four years. It was an amazing product. There are things that we did for disambiguation, for example. People's names are often very confusing if you use web search because there are many people with similar names.

And we figured out how to disambiguate different kinds of names. Lots of cool little techniques that made that product better. But I think you know this, Matt, companies have shelf lives. Once you've spent a certain number of years going at something, if you haven't achieved enough success, you also tend to accumulate a lot of cruft that becomes very hard to discard. And we decided that it was going to be very hard to sustain ourselves as an independent company and, much more importantly, to catch up to our valuation.

We had raised a lot of money. And in retrospect, it's again another one of these. I think there is virtue in taking, for example, just a few million dollars and saying the first version of a company needs to be built with five, six people in a relatively short period of time, and there needs to be some magic. And if there is not that magic, there is a structural problem. There is a lack of some kind of insight. I think the fact that we were in a zero interest rate environment blinded us to some of these things.

We raised too much money. Our valuation was $300 million for a company barely making a million. That just was not okay. And I had serious doubts, not about raising the next round. Unfortunately, we were an AI company. We could have raised another round.

But I had serious doubts about being able to catch up to our valuation. If you're saying, if I need to be a responsible company in 2024, and I should be valued at no more than, take your pick, 20 to 30 times your revenue—you're not even talking profit—that just looked harder and harder to achieve. And that was the main reason why I said we were throwing kind of good time after bad and good money after bad. And we basically sought a place where the expertise—and we had a lot of it, amazing people, amazing expertise in search, in AI—and we looked for a place where we could deliver a lot of value.

And Snowflake became that place. I was an afterthought for the acquisition. My original agreement with Frank, the Snowflake CEO, was that I would stay with Snowflake for six months. I said, I will help you get the team integrated. I will help you set up a strategy for AI, and I will leave end of '23 and figure out what I want to do. And that was very much my plan. This thing of becoming Snowflake CEO was very much an accident and a little bit of a happenstance that came together.

That's correct. Epiphany was an analytics company turned into a customer relationship management company. We were all about big data and analytics back then, built on top of systems like Oracle and DB2. And absolutely, we worked on machine learning algorithms there. And so this is—

Yeah. Snowflake is something that brought a lot of my interests together in life into one place. I have a lot of expertise, a lot of experience with big data going back to my PhD days, going back to the work in query processing at Bell Labs, also at Epiphany. Much of the early work that I did at Google was again on data-centric systems, on serving systems that were high performance, but also data processing systems that were incredibly high performance. I had the privilege to work with people like Shiv Venkataraman, who is now at Google, on building some of the largest log joining systems in the world.

If you got logs of every impression that we serve, just imagine how much that was. And we had to join them with clicks. This was back in 2006 and 2007. All of that experience then combined with machine learning and AI, being at Snowflake felt very much like coming home. I was like, this is a place I love. And I met a lot of customers. They love Snowflake. That's part of what made me decide that I'd be happy spending five or ten years at this place.

He did Marketo. That's right. That's right. They had amazing people. Craig Weissman, who was my first manager at Epiphany, was the first person that told me that I could become a manager. That was the farthest thought in my head, but he encouraged me to think about it. He was Salesforce's first chief architect.

And then their CTO as well. Marc and he had gone to school together at Harvard. And so a lot of pretty amazing people came out of Epiphany. Mehran Sahami, who did machine learning at Google and now works for Stanford, also came from the Epiphany mafia. A lot of really good people.

John was, I think, our head product manager. He was a VP of product.

So, small world.

First thing is, the power of defaults still matter. I'm obviously a sophisticated user of products. I know which ones I should use for what. I've become a big fan of ChatGPT with search for any kind of quick reference. I still use Google via Safari quite a lot for things I'm not quite sure that ChatGPT is going to answer, or out of sheer force of habit. The fact of the matter is there are billions of people that have this habit. They're going to keep using Google.

I don't think it's an immediate thing. But on the other hand, ChatGPT has numbers that begin to look like that of the big tech companies. My take is that any product that has a billion users is world-changing.

They're beginning to approach that. The last numbers that I have read are 750 million. And that's a lot. That's a lot of users, and their ability to monetize that, especially the high end with paid products. I'm sometimes ashamed to admit that I pay for their $200 product.

Actually, same here, because I don't want Snowflake to pay for it. It's too much guilt. I use their deep research a lot. I'm careful about which question I ask, but I consider it such a privilege that I can think about a new topic, or a new topic comes up in a conversation like this, and just walking back or driving back, I go, I really should learn a little bit more about this. And then you get this three-page report that combines the best of what's there on the internet.

It is wild.

But to me, that combination of incredible tail combined with the ability to monetize at the head, I think this is very powerful.

And so, given that they have a free product that can begin to address many casual use cases and an effective way to upsell people into the paid product—and remember, unlike Google, starting from zero is a magical thing. That was a lot of motivation for why I started Neeva. It's not only can you rethink the product, but starting from zero is a gift because every dollar you make after zero is money you didn't have before. And ChatGPT can afford to take that kind of an attitude.

I think the other thing that would be worrisome from a Google perspective is Google makes most of its money on highly commercial but largely undifferentiated search queries. What I mean by that is the queries with the most RPM, revenue per 1,000 queries, are queries like auto insurance that don't have an obvious answer. That's where people duke it out. You'll be shocked to know that there are queries that have $1,500 to $2,000 RPMs, which means Google makes $1 or $2 every time you or I, like a user, types the word auto insurance into search.

We are not talking clicks. We are talking about every time somebody puts that in, on average, Google makes that kind of money. But here's the thing: in five years, in three years, if you and I want auto insurance, we are going to go to some conversational interface. We are going to say, hey, I live in California, I have three cars, I live in Cupertino, and I drive about 5,000, 6,000 miles a year. My previous insurance agent was State Farm. What are my options?

And it's going to go to different sites. It's going to enter information on your behalf. Even if it doesn't go to different sites, it likely will go to agents that are provided by each of these providers. It'll supply information to them, take that information back, summarize it, come back to you with, okay, here are your top four options.

This is not a hard prediction. Technology is inexorably headed to this kind of an outcome. So to me, the conundrum of Google is: how do you fight off a world where the easy increasingly has just straightforward conversational answers and not necessarily the 10 blue links, and where the difficult becomes much more of a research task that people should go spend time on? And this applies even to products. You know this. I find the act of buying anything next to impossible.

Because if you like a certain kind of a shirt or a certain kind of pant, you have to figure out all the right sizes, do this, do that.

But if you're willing to trust some agent with this information and it's actually good at telling you, here are the options that you should be considering, that is going to be a preferable experience. I think that's the conundrum of where search is headed, where for all of the easy queries, you just want a simple answer. For all of the hard queries, you want a new set of techniques that basically don't really conform to the 10 blue links and click on this model. Google will survive, don't get me wrong.

But I think search needs to be very, very different from what it is today. And the question becomes, if you are Google, what are products you can create that strike the right balance between reaching for the future but not cannibalizing huge amounts of revenue? Mostly they've tried to address this with standalone things like Gemini. But part of the problem is, in a world in which Gemini is not 10 times better than ChatGPT, there's a first-mover advantage.

The rules that killed Neeva, which is, at the end of the day, until AI search came along, you simply didn't have a product that was 10x better than Google Search. That's just the reality. That's the reason we couldn't compete. But Google is bound by the same rules. They can't make Gemini into a consumer hit unless it's 10 times better than ChatGPT. And that's a really tall order.

If you look in terms of what AI can enable today and what we think AI can enable in the near future, and broadly break that down into categories of structured and unstructured information, Snowflake, as you pointed out earlier, made its name as the best structured data store for analytical purposes that there is. And our use cases run broad and deep, for everything from customers using us to close their books at the end of every month, to get their financials in order, to running anti-money laundering systems, to things like next-best prediction, which customers like Disney do.

On top of Snowflake data, that list goes on and on. But back to my point about what are things you can do with AI: business users inevitably want faster access to business data that there is. There's always been pain associated with getting data out to business users, especially new data, different kinds of insights, very quickly, because they live in tables and people have to write this language called SQL or bits of Python to get at this data. And so you have analysts, basically, that understand both the business context of the data and the SQL language to then write queries to populate dashboards.

And so there is a desire on the part of lots and lots of customers to short-circuit this process. How do I get data to customers faster? And so this is where products like Cortex Analyst come into play, because they make this act of getting data to end business users just a whole lot faster. But here is the rub, and here is where the entirety of the data strategy comes in, and where things like Iceberg still have work to do. It is one thing to say, I can write SQL reliably to answer a business question, but you quickly go from there into, well, who can see it?

What's the governance around it? Not everybody can see everything in a business context. A multinational bank, for example, has pretty tight rules around who can see what customer data, even if they've done the work to put it into a single table. And all of those governance rules apply out of the box if you use Snowflake. And so if you have an AI strategy of, I want to get data to my business users faster, it's a good strategy unless you have governance rules, unless you have role-level access control already defined.

This doesn't work. And a lot of our customers in the financial sector, for example, also use things like replication and disaster recovery so that if there is a problem, they can switch over from a main Snowflake instance to a disaster recovery instance within their SLA. Quite often it's things like 30 minutes for the most critical applications, without any loss of data, mind you, in between. This is actually switching stuff over. That's a little bit of what we mean by: you need a data strategy.

And so products like Pinecone—it's a great company—came along. They started providing vector indices. That's useful, but you run into the same issue. If you're indexing documents, you have permissions on documents that you have sitting in Google Drive. It's not the case that every document is visible to everyone. How can you have an AI system that respects that out of the box so that, again, those kinds of governance rules are applied? That's a little bit of what we mean by: you need a data strategy.

You need to figure out what it takes to create a layer of data that is analytics-ready. And just having raw logs, for example, is not helpful because, first of all, raw logs will not have the business identifiers that are useful for you to look at. Is this data actually, like, how do I look at it in terms of terms that matter to me as a business user? Having the IP address of a user logged is not that helpful. What you really want to know is which country they're from so that you can look at country breakdowns.

And similarly, pretty much every log source, every raw data source, has an element of spammy data in it. There is some bot that decides to send a pile of queries to whatever interface that there is, but you have to figure out how to get that data out, how to dedupe that data. I had entire teams devoted—it's called an ad spam team. When I left, I think they had more than 150 people. Their job was to get out bad impression logs, bad queries, bad click logs, bad conversion logs, because we needed to be careful about what we charged advertisers, what we counted as a conversion.

But every company has some version of, I need to clean this data out. That's part of what we mean when we say there needs to be a deliberate data strategy before you can then say, I'm going to use AI. And even when you can use AI, you still need business governance rules to apply on top of this data. And we have any number of customers, whether it's the likes of Morgan Stanley or JPMC that are big customers of ours. Siemens, funnily enough, one of the first AI products that they created for internal use was a chatbot that indexed all the PDFs of the manuals for all the devices that they make so that their customer service people could get their hands on relevant fragments very quickly if they are out in the field.

We have other customers like Bayer that are early adopters of Cortex Analyst for getting at structured data. There are an enormous number of companies that have large parts of their data estate in pretty good shape for them to be able to do this. By the way, this is also where Iceberg, or just a data lake strategy or a lakehouse strategy, is not an immediate answer because formats like Iceberg do not understand governance rules, do not understand row-level access control. They don't understand users.

They don't understand what it means to replicate data. And so this is where data platforms like Snowflake come into place to put that additional business sense on top of the data. But you also asked about what do people do if they don't have that?

My take very much is this is a journey, and you need to deliver value one step at a time. At Snowflake itself, for example, I am very flat with my team that I do not have any tolerance for 18-month projects. If you tell me that you're going to start on a rewrite and it's going to take you 18 months before it shows any value, they go, "No, I can't actually deal with that." And so they need to start on small projects.

Here is where we are hopeful. We have a team that is working on what we call AI migrations. We have written tools to make it easier to migrate data from a source system, whether it's SQL Server or Oracle. We are using AI to make things like unit testing of these conversions faster. We are actually hoping to apply some of our technology to the tools that we provide our customers so that some of these migrations are faster. My advice to all customers about AI, about data, is one project at a time and one dataset at a time, always showing value, always moving the business forward.

But I have met customers. I met this—it's a bank, mind you, in Charlotte—that said, "In six months, we have gone from having no data on Snowflake to being able to launch 24 different datasets on Snowflake in production, in use by customers." And mind you, this is a bank that has regulators asking them all kinds of questions about every aspect of how they design their systems. But this is also a little bit of, like, an amazing dude. He's the one that told me, "Yes, I work from 7:00 to 11:00 because my family life is from 7:00 AM to 11:00 PM."

Because my family lives in a different city, I just work Monday to Thursday, but I prefer to work seven to eleven every day. He's a little bit of an intense person in terms of how he gets work done, but rapid progress is possible. These tools are getting better.

It's easy to pass judgment in retrospect. I'm the first person to tell you that. But I like looking back and seeing whether we should have done things differently. I find that to be an intellectually rewarding exercise. I'm often critical of decisions that I make. You learn from them, not to place blame, but to just get to the bottom, perhaps, of why things went wrong. At Neva, we talked about why I should have called it quits earlier.

Or what were the structural flaws in Neva? The thesis that there needed to be a better search engine was a fine one. But at the end of the day, if you were to ask me even back in 2019, "Sridhar, can you create a better search engine, or at least show me you can create a better search engine for some slice of search with six people and six months?" My answer would've been, "That sounds hard. Probably not," but it illustrates the structural trap that I fell into, where at one level it was a good idea, it's a big market, but on the other hand, it's not like I had the insight to create that 10x product, knowing that getting people to change behavior is always going to be hard.

So, as I said, I apply that kind of critical thinking to decisions that I have made in the hope of learning. My take is that us deciding to monetize storage was a long-term strategic mistake. I think we simply should have said we will pass through storage costs, whatever rate we are able to get from cloud partners. We should have gone and told our customers, like, you pay that much for storage. It's not free, but it's very inexpensive. And the consequence of not doing that was people said, "Oh, I will move only the gold-layer data into Snowflake."

We lost out on earlier ingestion of data into Snowflake because we made storage a product. We monetized storage. But remember, storage is an undifferentiated product. It eventually ended up creating the pressure, in my opinion, for things like Iceberg, even though you can argue about whether Iceberg would have come even if we had made storage dirt cheap, because the movement to be independent of vendors is one that has legs, and it has sound structural reasons on the part of all customers who don't want to be doing migrations every five years.

Well, that's a little bit of the history of how we are here. But the flip side of betting on Iceberg, similar to what could have happened if we had sold storage at cost, is that the aperture of what we can now sell to our customers is a lot larger. We can go aggressively invest in earlier parts of the data lifecycle, like ingestion, because we feel more confident about our ability to do processing on it. And my take to my team on this topic is, we live in a competitive environment.

I expect us, me and the team, to drive way more revenue in new things we can do with Iceberg than we are going to lose potentially by some of our storage revenue going out to cloud storage. And at this point, we have pretty much lowered the cost of storage in Snowflake to match that of what the hyperscalers provide. And it's pretty inexpensive. There is a portion of that that we do make as revenue, but we don't particularly think of that as a growth driver and things of that nature.

But we do have to win more business with Iceberg than potentially what we are going to lose from storage going out. And we have to create a world-class product that people are delighted to use so that we continue to make money. But that's a little bit of welcome to life. You have to compete.

BI continues to be an important category simply because most business users on the planet interact with a BI tool. And there are more modern ones like Sigma. We are an investor in Sigma. We partner with Sigma. We share a heritage in that it's also a Sutter Hill company.

Mike Speiser special. And they're making remarkable progress. We are happy for where that is going. However, as you can imagine, business users being able to directly get at data always has powerful allure. And Snowflake having a direct relationship with users within an enterprise that consume Snowflake data is a clear strategic priority for me as the CEO of Snowflake. By the way, as a person that worked on consumer products for 15-plus years, it was really strange to come into Snowflake and realize that they didn't own the bottom layers of the modern data stack that you talked about.

They went, yeah, yeah, partners do extraction and transform. They didn't own the consumption layer of the data. It just blew my mind. Wait, wait, wait, wait, you don't own the front door? You don't own the back door? What are you people doing? It was just the weirdest of experiences because the first rule of consumer software is, you own the front door and you prevent anyone else from taking that front door. But on the other hand, I also felt like simply taking on BI as a category was a fool's errand.

Without that 10x insight, you're not going to—big companies don't get into new categories with products that they have and somehow win. It's not a thing. There are any number of examples of companies that have tried it and gotten nowhere. And remember, not only are you competing with an entrenched category, you're competing with an entrenched category that also has a free product.

Power BI is quite good.

And it's essentially part of the Office offering. And so I've been skeptical of saying that we can just compete in BI with a new paid product at scale when the bar is high quality, no price. It's a little bit like competing with Google. On the other hand, to my point earlier about AI, how we consume information absolutely is going to change. And things like Cortex Analyst—we don't pretend that it's a BI product. My aspiration for Cortex Analyst is, if 95% of the simple questions that any business user wants to ask is satisfied by Analyst, I see that as a significant step forward.

And this is the reason why we obsess about high precision. They need to believe in the product.

And for recall, I'm perfectly happy for all of the complicated questions to be, for this complicated question, you need to email this analyst or click on this button to send this to an analyst who will give you insights on which dashboard you should consult. I think that's an interesting step forward. And from that perspective, I also think that the value of the semantic layer, the metadata that describes tables, that describes the business metrics that are going to be used within that business, how they come together, I think that'll become more and more valuable.

So we actually see the catalog as a really important layer that we need to get better at in a hurry because that becomes the layer that enables AI at a higher place. The modern data stack is there, but I roughly say BI has to be significantly different tomorrow from where it is today. And back to the point of zero being a good place, the fact that we make $0 out of BI—

—is a good place. We can focus on what is innovation with AI that can drive broad use and sort of build up from there without getting into competing in—I don't know if you've used BI tools, but have you? Yes. Yeah, it's a dizzying amount of detail to get right before you can claim to be a BI tool. It's not for the faint of heart.

I have met customers that employ 3,000 analysts, and there is incredible pressure on them to hire more analysts. And tools like Cortex Analyst that let them do higher-level work of annotating the schema, creating the semantic layers, setting the grounding for how these tools should be used in bulk is much more fulfilling than writing endless variations of SQL queries for every little task that somebody wants to do.

We actually work with Redpanda. It's a partner.

That's a lot of different things. But I do think that things like streaming ingest, which we are absolutely planning to support—we support something called Snowpipe Streaming, which is low single-digit-second latency—I think these are things that we will continue to invest in. There is one aspect of real-time, real-time, which is, call it sub-50-millisecond real-time for messages and things like a streaming solution as a backbone for agentic applications. To increase the reliability, generally we're partnering in this area to provide this.

As I said, we have many customers that use Redpanda for the ingestion, and then it deposits data into Snowflake. I agree that it is an important area, and streaming as an AI backbone, I go back and forth. As you can imagine, if you have agents that take five seconds to come back, and that's not an unreasonable thing if you're going to use an AI model, consult some data, and stuff like that, reliability with RPCs is a little bit of an issue if you're operating at very high scale and you want to make sure that failure rates are low, because the streaming solutions provide protection against random dropped RPCs because of overload and stuff like that.

So I can imagine those kinds of use cases. So my rough take is it is an interesting area. The practical uses of true sub-50-millisecond streaming for analytics are quite limited and somewhat expensive. The AI backbone, we are continuing to explore the area, and we are in a pretty good place with respect to what we support with respect to real-time and ingestion and scale and stuff like that. But true streaming, Kafka-style, is an interesting problem that we are looking at.

And what has happened over the past few years, and Confluent demonstrated this with WarpStream, is that there are second-generation solutions that rely on things like cloud storage as their backing store that are considerably simpler than the pure Kafka isolated set of processes talking to each other. Or Redpanda is a similar kind of architecture. So there's a bunch of architecture choices as well when it comes to how do you do streaming.

I think it's useful to step back a little bit and talk about the priorities that we as a product team had for AI last year. Remember, we were playing catch-up with AI, and we wanted to bring products to market, but we needed to be deliberate. We could not be everywhere at the same time. And so the priorities that we set out for the team, first and foremost, were: build amazing products that stand on their own, that we could blog about, that we could publish benchmarks about, and say these are world-class products.

And we have done that for everything from what we call AI SQL, which is the ability to seamlessly blend in both LLM functions, language functions, but soon multimodal models, into how you write SQL, and be able to run processing with it. Batch benchmarks on what can you do with SQL. We publish benchmarks on Cortex Search, which is our unstructured data solution. World-class. As I said, origins with things like Mustang at Google in terms of how we think about search.

A team that has built search multiple times and really knows what it's doing. This is the third or fourth-generation system that Asim has worked on with respect to search. And so, a world-class product there. Then Cortex Analyst, as I said, a unique take with an early focus on maximum precision. How do you put in place a software system that can get better over time? And our second priority was getting marquee customers to adopt AI. And so this is where getting customers, whether it is a Disney or a Siemens or Elevance Health that we work with, or Bayer, getting a lot of these brand names to adopt AI was important for us to establish credibility.

And then the third directive that we agreed on was, we said we want very broad adoption, not revenue, very broad adoption. Make the product so easy that tons and tons of customers can create and deploy POCs, and them spending lots of money is not necessarily the first priority. To me, this part is very important because I wanted to demystify AI for our customer base. I flatly tell people, if you have data in Snowflake, creating a chatbot, an unstructured data chatbot, is literally five minutes of work.

It's that easy. Cortex Analyst takes a little bit more effort because you need to have semantic information about what do the tables mean, what does a measure mean, business metric mean, things like that. But again, in a few hours you can get a perfectly passable solution that you can put in front of a business user and ask them to try it out. We have work to do in terms of getting it to very high precision. The net of this is that we have thousands of deployed production use cases on our AI products.

And as I said, the benefit of this is there is broad, much broader awareness. We had to be very deliberate with our sales teams about how we educated them, who are the AI experts, how did we establish regional presence? Remember, I need to make sure that there is somebody sitting in Japan locally that knows about Snowflake AI products that can talk to local customers about it.

I also got grief about whether we have GPU capacity in Australia. We have this wonderful woman who is our solution engineering VP, and she always gives us feedback about what we need to do better. But we drove that broad adoption. And so we have a number of use cases that have been deployed to production that run the gamut from just write SQL to create different kinds of chatbots. And we are now beginning to do things like look at scaled adoption.

What does that mean?

You have customers now asking questions about, wait, if I release Cortex Analyst to 1,000 business users, talk to me about how much money this can be. What's the upper ceiling? These are good conversations to have because they want to make sure that these products are going to scale over time. There's also this issue of us scaling our capacity to be able to do this. Again, having models like Anthropic and OpenAI as the control models helps us a lot because they're just better—they do better planning, they reach better decisions in terms of where it's going.

That's sort of roughly where we are. My take is that revenue will come. And margin will also come. I was very clear with my team to not focus on margin early. Back to my point about storage and margin, I just think that the world of inference, not foundation models, the world of inference is going to go through so much change this year, both in terms of GPU availability, which is easing up quite a bit, but also in terms of people like Groq.

I think the one with the K or the Q, Jonathan's company.

With fast inference, much cheaper inference. I think there are a number of options like that. So I like our strategy for how we are taking AI to market. We are now beginning to have conversations with our customers about, can you create native applications that can join provider data with customer data and then be able to do things with AI that combine this data? And so we are very much at this phase of demonstrate the art of the possible. Get them out to production, get them actually used by people that are getting value from it, and then revenue will follow.

And so this has been very beneficial because I can go to our customers and say, listen, I'm not asking you for a $5 million commitment on AI. You can create a chatbot with $10 or $20. Try it out. If you see value, we will figure out how to scale it. I think this is a much more sustainable approach to create a very broad business that, over time, will put roots, will drive revenue for us.

We got out of the foundation model business early last year. It just looked like an impossible challenge. At the end of the day, we are a smallish public company compared to the likes of Google and AWS and Microsoft, or even OpenAI, in terms of how much money we are able to put for things like model training. We shifted our folks to focus more on things like post-training, where we felt we still had leverage. We also have a very good inference research team.

That specializes in making inference super cheap and super fast for our needs. I think that's the right place to be. That also drove our partnerships. By the way, these are deep, meaningful partnerships, meaning the Anthropic models run within our deployment. So I can with confidence look at our customer and tell them that their data is not leaving the Snowflake deployment. And it's similar with Microsoft and Azure and OpenAI. And so these are meaningful integrations that we have done working with these companies and the cloud providers.

I feel like that's a much better use of our resources, which we have to husband a little bit, than continuing to try to invest in foundation models. We got priced out. It's fine.

That's right.

We love open source. Why? It drives competition. It's that simple. And we launched DeepSeek quickly. It's more a demonstration of, can you run that 11-second sprint? This model's there. It appears to be good. If the worry is where it's hosted, we can host it. We actually hosted the full version of DeepSeek, not their small model. These are the ones that are not going to fit within an H100. And we were proud to do it. Is it a company with staying power?

I'm not yet sure in terms of their ability to innovate. If you were to compare them to xAI, I would say they are definitely one or two steps behind in terms of their ability to come from nothing and train a world-class foundation model. There are lessons for us in terms of how nimble they've been and how scarcity actually leads to innovation. Back to my point earlier about there is such a thing as having too much money, even as a startup. I think there are lessons that we all should take away from it.

And open models are great. I'm super excited for Llama 4, which I think is coming out imminently.

So we partner with Meta as well. So I think the more models there are, the more options that you and I are going to have as consumers.

This is just about us giving back, us helping educate the next generation of people on data and AI. A part of it is also about educating people on Snowflake, but the goal is very simple. There are a new set of tools that are going to be coming, and the more people that are educated in it, the better it'll be for everybody, including Snowflake. And that's the reason why we started that program.

That's right. We have a beautiful new campus in Menlo Park that's opening next week, and we reserved space in it, one floor of it, for startups that we collaborate with. We invest in startups out of our balance sheet. We work with some VC partners, but also directly with startups, generally in areas where there is alignment that we can be working together. And the AI Hub is just a continuation of that, where, in addition to things like funding, we can also offer office space, the ability to work closely with Snowflake engineers that are in the office, and a great new campus that had excess capacity.

Thank you, Matt. This was a broad and deep conversation, and I really appreciated chatting with you. Thank you. Thank you.