An Interview with Snowflake CEO Sridhar Ramaswamy About Knowledge and AI – Stratechery by Ben Thompson

Good morning,

This week’s Stratechery Interview is with Snowflake CEO Sridhar Ramaswamy. Ramaswamy began his profession at Bell Labs, earlier than spending 15 years at Google constructing out Google’s Search Advertisements Product. Ramaswamy then began Neeva, a subscription-based search engine, which was acquired by Snowflake in 2023; eight months later and Ramaswamy changed Snowflake Chairman Frank Slootman as CEO with a mandate to refocus Snowflake on merchandise, particularly AI.

We cowl all of those matters on this interview, together with Ramaswamy’s background and expertise at Google, and his present tackle the corporate and the challenges it faces in search. Then we dive into Snowflake and his sudden elevation to CEO, together with matters like enterprise fashions, go-to-market motions, and incentives. The remainder of the interview is about AI and Snowflake’s place available in the market: can Snowflake lengthen past its structured information warehouse roots earlier than opponents like Databricks leverage AI to wrangle unstructured information right into a extra compelling providing?

As an apart, I did beforehand interview Databricks Founder and CEO Ali Ghodsi about comparable matters, and whereas the 2 firms are bitter rivals, it’s notable how comparable their long-term view of their AI alternative is, notably relative to the hyperscalers on one aspect and SaaS firms on the opposite.

As a reminder, all Stratechery content material, together with interviews, is obtainable as a podcast; click on the hyperlink on the high of this e mail so as to add Stratechery to your podcast participant.

On to the Interview:

An Interview with Snowflake CEO Sridhar Ramaswamy About Knowledge and AI

This interview is evenly edited for readability.

Google and Neeva

It’s nice to get an opportunity to speak to you on the report. We chatted beforehand while you have been at Neeva, which I do wish to ask you about, however I at all times like to begin these interviews along with your background. The place did you develop up and the way did you get began in know-how?

SR: I grew up throughout South India in a state referred to as Tamil Nadu, after which I moved to Bangalore once I was 10 years previous. Lots of my formative reminiscences are in that metropolis and naturally, it’s now referred to as Bengaluru and it’s an entire lot totally different from the town that I grew up in.

Yeah, stuff modifications lots.

SR: It has modified lots. I acquired a bachelor’s from IIT Madras, now referred to as Chennai, after which I acquired a PhD in databases. In some ways, IIT Madras knowledgeable how I take into consideration computing. A tremendous set of colleagues that I studied with, professors that would have been wherever on the planet, however select to be in India to show college students like us. Brown College was additionally very useful in creating my important pondering, extra unstructured work. As you already know, getting a PhD is all about discovering attention-grabbing issues, which could be very totally different from taking programs. After which a number of years at Bell Labs. I’d say these three, most likely the most important early influences in computing, after which I moved out to the West Coast.

What did you do at Bell Labs?

SR: I used to be within the database analysis group. I joke to those that that is the constructing at Murray Hill the place the transistor was invented, I used to be a moderately pedestrian database researcher. I inform those that my declare to fame is I used to be subsequent door to one of many scientists that invented most theoretical facets of quantum computing, nevertheless it was an incredible place. And from there I got here to the Valley, joined Google early 2003 to a really totally different world of computing the place it’s important to relearn what you knew about software program and computing.

The database facets, you’ve come full circle in some regards being at Snowflake.

SR: 100%.

However Google, lots totally different. Inform me about Google, inform me about your time there. Like I stated, you have been there at a really attention-grabbing interval for the corporate there for a very long time, very distinguished profession. Stroll me by way of that.

SR: I joined in 2003, I truly ran a pretty big software program group for a small startup of over 100 individuals making $100 million income. However once I joined Google, I made a decision to return to being a person contributor as a result of I stated this can be a firm that’s rethinking, redoing how we take into consideration computing and distributed programming, and I needed to study from the bottom up. I used to be very fortunate to get by chance positioned, thoughts you, into the Search Advert serving staff. Little did any of us know that this is able to change into among the many greatest companies, if not the perfect enterprise mankind has ever invented.

Was there something related to doing conventional databases when now, I imply clearly on the finish of the day you could possibly argue all the things is a database, however at Google’s scale, it’s one thing else nearly solely.

SR: Nicely, sure and no. A variety of early Google’s method to computing was pushed by the truth that early Google largely handled read-only issues. For instance, internet search is usually read-only. There’s somewhat cache of your preferences and stuff like that, however the bulk of it’s read-only, and so there are approaches to distributing data, distributing load which might be very totally different from the way you assume say about database scale. We did have MySQL databases on the advertisements staff at Google, thoughts you, and I was part-time sysadmin for these databases for some time as effectively.

That feels like a really tense job.

SR: However it was superb to study computing. We additionally constructed a number of the most superb information processing techniques ever, log becoming a member of at planet scale. You could have logs coming in from each consumer on the planet that’s seeing advertisements. It’s essential be a part of them with clicks, I used to be speculated to do it on time.

I used to be going to ask that. You talked about search is usually going to be read-only, however with advertisements, there’s a write downside.

SR: Once more, sure and no within the sense of stats processing is delayed processing, you’re not attempting to try this in actual time. Completely, advertiser updates did go to the MySQL database, however that was a tiny a part of general advertisements information, issues like logs, logs processing, combination processing at scale. Among the many largest logistic regression techniques constructed early on the planet was a system referred to as Good Ass. There’s a variety of distributed techniques programming and naturally, extremely high-scale serving techniques as effectively, and so you actually acquired to study what distributed computing was about.

That was once I additionally rediscovered my love for truly main, and Google allow you to lead any which method you needed. I used to be a really hands-on technical chief for the longest time. I liked the combo of managing groups, serving to them ship nice merchandise, but in addition being tremendous technical, tremendous hands-on and it helped to work on an extremely technically demanding product. Over time that grew to become me working extra of the Search Advertisements groups after which extra of the Advertisements groups, which I did for 5, six years, Advertisements and Commerce. This included enjoyable issues like launching Google Pay to the world, which was an entire saga onto itself, but in addition the varied journeys of Google Commerce that you’ve got lined elsewhere. It was a captivating place, 15 superb and extremely rewarding years of my life and profession.

So then you definitely determine to depart, and never simply determine to depart however determine, “I wish to construct a product that has no advertisements”, referring to Neeva, a subscription search engine. We are able to get to Neeva in a second, however what was the motivation? Was it, “I’ve simply executed this for a very long time, I actually needed to construct one thing totally different, Google’s modified”, the place are these causes?

SR: It’s none of that. I had executed advertisements for 15-and-a-half one thing years and I used to be like, “That’s a very long time doing one sort of product”. Clearly serving advertisements for YouTube, which my co-founder at Neeva and I did could be very totally different from doing search advertisements, totally different issues, totally different advertisers, totally different objectives. However there was part of me that needed a reset button and I needed to begin recent, and I additionally knew that I had a restricted period of time earlier than I may go begin an organization. I used to be not younger even then, and largely I simply determined to reset and begin over.

I truly joined a enterprise agency, Greylock, proper after I left Google and the startup Neeva got here somewhat bit later. As I stated, working at Google was an unimaginable privilege, however there was additionally part of me that stated that I didn’t need that to be the interval of my skilled profession and so Neeva was a begin over and generally you simply have to begin new journeys they usually take you to attention-grabbing locations, and right here I’m.

What’s your notion of Google right now from afar? I believe it’s been sufficient years, you can provide some remark. Is Search tremendous? Nobody needs to be fearful about it? Or are they shifting too slowly? What’s your view?

SR: Look, that is the Valley, solely the paranoid survive. That’s as a lot true right now because it was 5 or 10 or many extra years in the past. This can be a time of extremely fast technological change, and we are able to go into element about how totally different firms are approaching it.

I personally really feel that there are phases of consolidation for firms when issues are going effectively and also you wish to be extra environment friendly at what you do. There are occasions when you really want radical change as a result of there may be radical change taking place exterior. Let’s face it, there are increasingly questions the place merchandise like ChatGPT utilizing internet search simply supplies a greater expertise than looking on Google. Completely I search on Google as effectively, it’s my default search engine on Safari, however I even have the ChatGPT app, I pay for it. I believe you’ll be able to say now that for a lot of, many frequent use instances of search, a extra conversational expertise is a greater expertise.

After which then again, for issues like trying by way of difficult areas the place you actually don’t need a conversational expertise, attempting to determine precisely which shoe to purchase in a conversational expertise, as you’ll be able to think about, that’s simply plain annoying.

However then again, I believe you start to have agentic techniques that may do far more difficult workflows than what you could possibly ever anticipate from a search engine. So I’d say Search is underneath siege from either side, from each the frequent issues that you really want easy solutions to, and the complicated issues the place you’ll be able to apply very totally different strategies from what you might have executed earlier than.

That’s a very attention-grabbing method to take a look at it. It’s like a pincer motion the place you might have probably the most difficult and the least difficult. Yeah, that’s a good way to border it.

SR: And so they all have very, very totally different options. And truthfully, in my thoughts, that’s a narrative of knowledge and Snowflake the place, sure, we began at one finish, however I believe the place we’re headed to is an growing blurring between product and repair, for instance. I believe this can be a quickly altering world and, sure, one at all times has to fret that Search just isn’t altering quick sufficient.

I used to be one of many key gamers within the cellular transition that Google made out of being a desktop firm to a cellular firm. We used to have this phrase referred to as RPM [revenue per thousand impressions] hole, which was principally the share of cellular RPM to desktop RPM. We began with that being 8% or 10% and spent a terrifying 5 years attempting to get that nearer to 100%.

Making an attempt to shut that hole, yeah.

SR: As I stated, it was really terrifying. I believe that is a type of moments, it is going to take 3, 4, 5 years to unfold, however it is rather tough to assume that how we devour data or merchandise 5 years from now’s going to be what we did three years in the past.

Most of this interview goes to be about Snowflake, however earlier than I get there, I did wish to ask about Neeva. I believe the high-level pitch is, “An ad-free search engine, subscription solely”. Is {that a} adequate abstract? What did you attempt to do and what classes did you study there?

SR: The mental underpinning of Neeva was principally that an ad-supported search mannequin had reached its restrict when it comes to how a lot utility it may ship and that one wanted to begin over. The perfect that I may do in 2019 was to say, “An ad-free non-public search engine”, these have been the perfect ideas that I may discover to explain.

Do you’re feeling such as you have been in the suitable locations on the improper time?

SR: We completely have been taking up the suitable downside most likely two to a few years too early. By the point we launched the primary true internet index RAG[retrieval-augmented generation]-based search engine in early 2023, we had been round for four-ish years. My staff was drained, and we determined that we have been higher off making use of what we knew to a brand new house during which we could possibly be much more leveraged than proceed on. You know the way startups are, it’s important to be on the proper place on the proper time.

Yep.

SR: I believe the thought was tremendous, nevertheless it wanted one or two key items of know-how that didn’t exist after we began.

The brand new product paradigm, which was the chat interface. I simply talked to Sam Altman final week and he’s fairly anti-ad, he needs to attempt to make all of it work with subscriptions, which is consistent with what Neeva was doing, besides they’re ranging from being a chat interface and that’s most likely a greater place to begin from.

SR: That’s proper and 700 million customers, which can be useful.

I used to be going to ask that. Did you acquire extra appreciation for Google working at Neeva or perhaps extra disdain? It may go both method.

SR: Google is an incredible firm, I’ve stated this earlier than. A variety of Google’s early customers enterprise was executed each with an excellent product but in addition extremely shrewd enterprise offers with the likes of AOL and Yahoo and so forth. I’ve a variety of respect for what Google did.

And that rolled ahead to the offers with Apple and the Android and issues like that.

SR: 100%, the Apple deal, the Firefox deal, I believe the numerous PC producer offers again when Microsoft was asleep.

Yep.

SR: All of these have been what made Google the superb enterprise that it’s. However enterprise fashions even have shelf lives, and generally it’s very onerous to disrupt your self with a mannequin that may not make any cash while you’re making $200 billion on that mannequin. It’s simply not a factor that computes and that’s the place we’re.

Snowflake acquired Neeva in Might 2023, simply after you’d lastly discovered the AI RAG search concept, you have been CEO eight months later, was that at all times the plan?

SR: It was not the plan.

Snowflake purchased Neeva due to our experience in each search and in early AI. We used to tremendous tune, it feels like a joke now, seven nine-billion parameter fashions to do actually good summarization to have the ability to write the cited solutions that we now take with no consideration. We needed to do all of this on a shoestring as a result of we have been serving web-scale site visitors and definitely had web-scale indices six, eight billion pages giant.

That was the experience that Snowflake discovered enticing. We have been all very clear about Snowflake and Neeva that we might not proceed the buyer search engine. In reality, we shut it down even earlier than the acquisition closed, and we went to work on creating an excellent search product inside Snowflake. It’s now referred to as Cortex Search, after which pioneering issues like RAG-based search inside Snowflake, but in addition then engaged on strategies for structured information.

My unique take care of Frank Slootman, then CEO, was that I’d keep at Snowflake for six months, work on AI, assist set a roadmap for AI at Snowflake, after which work out what it’s that I needed to do. And in direction of the tip of the yr, I began having a number of conversations.

“Frank, I would like your job, that’s what I’d love to do.”

SR: (laughing) It didn’t fairly come to that, to be trustworthy with you. Come October, I used to be attempting to determine what to do and that is after we began having a number of conversations about whether or not there have been totally different futures. Clearly, this stuff are difficult and transitions are onerous, nevertheless it all got here collectively in February a number of months afterwards.

Nicely, when Frank Slootman retired, the inventory dropped 20% in a single day. It nonetheless hasn’t returned to the identical stage. Did you might have any concept what you have been entering into? Thrown proper into the fireplace doing earnings calls, persons are nervous?

SR: Yeah, even trying again, I believe it is very important additionally acknowledge that we guided the yr final yr to 22% when the consensus expectation was 30%.

Proper, there was a variety of stuff occurring.

SR: There was a variety of stuff occurring over there, and it didn’t assist that Frank, the legend, was retiring, and I used to be coming in. I believe it’s actually the double whammy of the change and the vastly lowered expectation, the steering that basically threw individuals for a tizzy.

Then again, sure, I’ve spent extra time with our traders, analysts which might be masking us, clearly the staff. Look, Frank has at all times been a straight shooter and he pushed for this transition as a result of he and the board strongly felt that it was actually essential for a product-oriented individual, anyone that breathed and lived product for a very long time to be the CEO at the moment of tumultuous change. I’ve to present them a variety of credit score, I’ve to present Frank a variety of credit score for not dragging this out to be a three-year transition, I’m positive you’ve seen lots of these in lots of different firms, he felt {that a} clear break was the suitable option to do it.

Sure, it was somewhat little bit of shock for the those that have been masking the corporate. It’s a exceptional firm, I spoke to a variety of Snowflake prospects earlier than I grew to become CEO and over simply the previous two, three quarters, you’ve seen how shortly the corporate has been resurgent, but in addition the pace at which we’ve been in a position to roll out new issues.

So in that sense, I’m fairly pleased with the place we’re. Clearly we’re going by way of a bunch of macro modifications which might be miserable the entire inventory market however I be ok with the place Snowflake is each as a knowledge platform, however far more importantly as an AI information platform and what we are able to sit up for in our future. Was it a trial by hearth? Sure, however we are going to come out stronger.

Nicely, let’s spend the remainder of the time on Snowflake. Simply stepping again as somebody who was comparatively late to Snowflake, I believe you had in your Twitter profile or LinkedIn someplace, your job was studying Snowflake. What did you study? How do you describe what Snowflake is right now? What issues does it resolve? Why is it a greater answer? What’s your pitch to somebody who doesn’t know what Snowflake is?

SR: You already know after all that the adjective “studying” has a number of connotations. One is I’m actually studying Snowflake as in you might be utilizing it as a verb, however there’s a unique interpretation during which I’m a studying, a gift steady sense of simply studying on a regular basis, and it was meant to be tongue-in-cheek.

Snowflake got here of age as an extremely versatile analytic platform, a cloud information warehouse because it have been, and like different technical merchandise, it was born of the remark that everytime you and I needed computing — and it nonetheless applies after we purchase telephones — individuals at all times purchased bins they usually at all times got here with mounted parameters for issues like reminiscence and storage and compute, and when you needed to do extra, effectively, that was too dangerous, you wanted to attend for brand spanking new bins to indicate up. So we did this and our founders created Snowflake within the cloud to be this platform that would scale alongside a number of dimensions and it was an extremely environment friendly platform.

Proper. Separating the information and the compute.

SR: That’s proper, separating storage and compute. So when you needed to do tremendous fancy machine studying evaluation in your information set, you’d wrestle to try this in a legacy system. If it took an excessive amount of compute since you had manufacturing jobs working on it right here, that’s tremendous.

Or when you had two those that needed to research one thing on the identical time, powerful luck.

SR: Robust luck, take the ticket and look forward to it. Snowflake made all of that stuff simple.

Half two of Snowflake was then a collaboration platform, which we principally made the remark that companies work with one another, work together with one another, accomplice with one another. As you already know, lots of these interactions have been executed by issues like nightly switch of FTP recordsdata. To at the present time, I discover it dumbfounding that in 2025 ACH transactions take days to settle, that’s as a result of banks are sending recordsdata to one another.

I had the identical response once I was getting ready for this interview and understanding how these information flows labored. And I’m like, “I used to be a giant FTP consumer in 1998 when that’s the way you used to get MP3 recordsdata earlier than Napster got here alongside”, are you the Napster of knowledge sharing between firms? Is {that a} option to put it?

SR: No, it’s extra like we’re the circulatory system of the enterprise world. So firms like Constancy for instance, mandate that every one their companions give information to them by way of Snowflake sharing. We’re realtime, we’re cross-cloud and information simply exhibits up the place you need it to indicate up and there’s no programming concerned. What was once an IT challenge is now somebody configuring a display and alluring a consumer on the opposite aspect to get at it.

An unimaginable variety of information firms, firms that primarily promote information merchandise for a residing, whether or not it’s the New York Inventory Alternate or S&P International or State Avenue, all of them distribute their merchandise by way of Snowflake. So it’s been an extremely essential a part of Snowflake.

I believe it’s truly fairly attention-grabbing. One of many first large modifications that occurred underneath your watch was supporting Apache Iceberg, and letting your prospects make the most of storage exterior of Snowflake. So that you began out, we’re going to interrupt aside storage and compute which is a giant differentiator out of your opponents, nevertheless it was additionally, “Oh nice, we’ve two income streams”. Nicely, it seems you might have two separate income streams, you now have strain to really quit a type of income streams, which is sort of what occurred.

SR: You already know what I believe, I believe it was a mistake to think about storage as a income stream. I believe we should always at all times have bought that at price and inspired individuals to place as a lot as they might into Snowflake.

However impartial of that, although, I believe the open format requirements are right here to remain. It’s because the progressive prospects on the market, enterprises which might be on the market, at some stage don’t wish to undergo extra migrations once more, they wish to have their information in codecs that they will instantly run compute on. By the best way, that is additionally going to place a variety of strain on SaaS firms to present their information to their prospects. These individuals all legitimately say, “On the very least, I desire a copy of my information”, and Snowflake we embraced it.

I’ve made two large product modifications clearly with the staff over the previous yr. One is a wholesale embrace of open codecs. In fact, a technique to take a look at it’s storage income that we used to get, now we received’t get as a result of it’s sitting in open storage codecs on cloud storage. A unique method to take a look at it’s most giant enterprises have tons of, generally thousand occasions as a lot information sitting in cloud storage as they do inside Snowflake, and impulsively our superb compute engine can now be used for information engineering, can now be used for information ingestion.

So it considerably expands the worth that we are able to convey to our prospects and that’s been a change that we’ve embraced. Iceberg is the preferred format that’s on the market, however we’re pushing past Iceberg. We launched one thing referred to as Apache Polaris, which is an open catalog format as a result of we wish to make it simple for individuals to find information units as effectively. After which after all, half two of that’s AI which we are going to speak about quickly.

What strikes me, although is you introduced up the community results that Snowflake has, and certainly one of my senses is firms get in bother after they maintain on to too many factors of integration, and so that you had, “We’re going to promote you storage and compute”, and yay, two income streams. However to your level, that really restricted what was extra of a lock-in. Lock-in is a flowery option to say an attractiveness of your platform, which was this information sharing layer, and it’s like you must be one or the opposite. In the event you’re going to be a community, then all the things else round it must be free.

SR: I fully agree with that. I believe Snowflake’s core worth is that of an incredible information platform for large-scale information computation. We have now efficiently introduced parts of machine studying AI into it, so it’s somewhat little bit of one-stop store for a lot of totally different sorts of computation that individuals wish to run on that, and the community of shoppers that we’ve and the way they work with one another, that’s the enduring worth of Snowflake. Sure, we began as a proprietary format firm, and it’ll take a while for it to play out, however I believe there’s an enormous alternative within the open information house as effectively.

Nicely, it’s additionally a little bit of a threat, as a result of when you can simply get compute, in case you have your information impartial, you could possibly very simply go and use a competitor as a result of the information is simply sitting within the center and anybody can lock into it.

SR: There’s much more to compute than working a SQL question. It will get into all the things like, “What’s the governance help that you simply present? What’s the superb collaboration services that you simply present?”. We offer catastrophe restoration for our greatest prospects. There are prospects which might be required by regulators to run a major in a single cloud supplier and a backup, which by the best way prices solely 10, 15% of the unique one, in a unique cloud supplier, and all of this comes built-in in a single tight, easy to make use of product. So there’s much more to Snowflake than, “Right here’s a SQL question that you would be able to run”. Does this imply that we’ve to maintain competing and maintain getting higher at what we do? Nicely, welcome to competitors.

Google’s billing was at all times 30 days arrears sort of billing, it was at all times invoiced, that’s not fairly the way it works at Snowflake, however lots of the strategies that I discovered each optimizing Google Search but in addition working with the gross sales groups are strategies that I’ve adopted right here. I’m a giant fan of issues like Boolean metrics for measuring effectivity the place as a substitute of measuring a mean, in case you have a staff of 300, you wish to know what fraction of your staff is past a baseline of excellence that you simply set. So there are some strategies that switch over and different new issues that I’ve needed to study, however that’s life and that’s enjoyable.

SR: I believe that is precisely the sort of factor that we’d like. We have now printed metrics on Cortex Analyst, which is our product for unstructured information. And actually, in very a lot the design side of Cortex Analyst that I used to be very concerned in, a part of what I advised the staff that we would have liked to consider was the precision-to-recall trade-off. As you already know, search operates on this mode of pretending that it has infinite recall, it doesn’t matter what query you ask, Google is like, “Yep, I acquired a solution for that”, and it’ll by no means inform you how good the reply is and whether or not it’s best to consider it or not.

That was fantastic thing about the ten blue hyperlinks that on the finish of the day, you the human needed to make the ultimate resolution about what was the reply.

SR: You’re the human, you might be one deciding that. Although I do discover that there are extra instances the place Google will inform you there will not be nice outcomes for this question when you sort in like dumb stuff, which is somewhat little bit of a change over the previous few years, I truly assume that that’s change. However sure, early fashions of ChatGPT for positive would simply reply any query that you simply requested even when they didn’t actually have any data or any confidence about it.

So we very a lot construct in, we name it a verified question repository of constructive and adverse examples for Cortex Analyst, as a way to sample match each towards constructive examples that an analyst has vetted that this product ought to really feel assured answering, but in addition a set of questions that we shouldn’t be answering.

As a result of roughly talking, when you put a chat bot in entrance of an individual, one of many first questions they’re going to ask is, “What is going to my income be 15 years from now?”, which no mannequin has any enterprise answering. So I believe having metrics like that and making specific trade-offs for when are you able to convey worth is a very essential side of any AI-driven product that merely doesn’t get sufficient consideration.

What’s extra priceless do you assume? Perhaps I’m most likely main the horse to water right here, however to have say a GPT-3 stage mannequin with 99% reliability or a GPT-4.5 mannequin with 95% reliability?

SR: I imply, I’ll provide the identical reply. When individuals give me selections about pace or high quality, my reply is, “Sure”.

(laughing) Nicely no, it’s important to select one.

I imply you’re not constructing or publishing huge foundational fashions, and also you truly have been fretting in some public feedback a yr in the past that, “Boy if it takes billions of {dollars} to make a mannequin and we’re decreased to solely a few mannequin suppliers, that wouldn’t be nice”. Has this yr been a really thrilling yr in your perspective with these open supply fashions which might be so able to popping out and are these fashions tremendous important to your corporation going ahead?

SR: We work lots with open supply mannequin suppliers. We have now a superb partnership with Meta and the Llama staff. I believe the quantity of progress being made on the planet of fashions, I believe it’s been fairly exceptional, I believe it additionally primarily prevents fashions from turning into an oligopoly. I believe having a world during which there are just one or two large mannequin suppliers or three is simply not that nice an consequence. I believe final yr has been fairly good from that perspective.

You might also like

Who Owns America | Enterprise Technique Hub

20 April 2025

Enterprise Technique with Cameron Bishop

19 April 2025

So the place do you see your utilization of this taking place? Do you see prospects in the long term simply switching these out and in willy-nilly? Do you must do fairly intensive post-training on a per particular person buyer foundation? You could have a competitor that has their very own mannequin and their promise is to coach a mannequin for his or her prospects, or is that this simply going to be fancy RAG with a verification system round it? I imply, the place do you see leveraging your place however remaining aggressive in the long term?

SR: I have a look at AI as a pure addition to what we do as a knowledge firm, and so for this reason we construct merchandise that complement what we do. The search product for instance, for unstructured information, together with embracing issues like open codecs. We’ve invested in a giant method in connectors, so extra information from extra functions might be introduced both into cloud storage or into Snowflake in order that they’re available for question. Equally, on the subject of structured information, we wish to make it tremendous simple for individuals to get on the structured information by asking pure language questions, but in addition having the ability to create issues like dashboards on high of the information.

The place we’re headed is now in how are you going to compose these totally different information parts, these agentic operators as they’re, to construct extra complicated workflows, and this was a degree that I made earlier about how merchandise acquired set into firms. As you already know, there’s a really giant ecosystem of system integrators that take merchandise after which create actual utility coping with the entire element and messiness that exists in each firm. I see a world during which increasingly of actual use instances that you simply and I do is instantly solvable on a platform, that’s the angle that we very a lot see Snowflake pushing in, it’s driving workflows on high of knowledge. I believe we’re redefining the road between product and companies due to AI and because of agentic capabilities.

To reply your query, by the best way, with respect to which mannequin would I choose, the best way I attempt to not have that be an either-or is for instances the place you really need exact quantitative solutions, you desire a system that’s tuned for max precision. If then again, you’re coping with somewhat bit extra of an analytical activity that doesn’t have a set reply, you desire a extra highly effective pondering mannequin which can make extra errors however can even give you higher plans. It’s a query of getting the issues which might be in your basis be dependable. So the place I see Snowflake is constructing on high of the information platform, of the information estates that Snowflake has and utilizing that to ship quicker worth for our prospects.

You type of have been driving at what to me was certainly one of my large takeaways serious about Snowflake, and Frank Slootman is a legend, it’s onerous to take over for him. He had sure axioms, certainly one of them was, “Slender your focus”, I believe is one thing he’s well-known for. One factor that happens to me nonetheless, is that if the general Snowflake proposition is we make your information accessible and simpler to work with, you will get far more utility out of it. There’s a consumer expertise part to that, and infrequently a option to ship a greater consumer expertise is to be extra built-in, to do extra merchandise. Generally when you have been utilizing Snowflake, you’d hit a wall, “Sorry, you must go get another person to fill this product want”, and it’s like, “I like Snowflake, can’t you simply do it for me?”.

You’re referring to that it feels like on this regard, the place there’s extra stuff that the product ought to simply be capable to do as a substitute of going exterior. Have you ever felt like one factor you must do is definitely Snowflake must increase its focus somewhat bit right here? It must be broader and must have — you must not hit these partitions as a lot as you probably did beforehand?

SR: For positive broader on the base. Snowflake restricted itself to primarily what’s referred to as the gold analytic layer that information that has been cleaned, that has been processed that you would be able to run analytics and machine studying on. I believe what we’ve executed completely with our embrace of open codecs, with our embrace of connectors, is that this potential to behave on a lot bigger quantities of knowledge. We’ve made the underside an entire lot beefier, which once more performs to our core energy.

Does it although? I imply you might have other people that started off — you had the Warehouse versus Lakehouse, all the varied — and also you talked about initially about totally different opponents coming in at totally different angles. So you might have a competitor that begins out very broad and that seems to be nice for AI as a result of it offers effectively with unstructured information, Snowflake’s extra slender, however simpler to make use of. Do you actually really feel your benefit in that case goes broader, or are you combating on unfamiliar terrain and it’s important to get your information out of Snowflake truly to make it extra accessible and that’s been a battle you’ve been having to struggle?

SR: It’s not an both/or. I believe there may be, to start with, it’s essential to grasp that we’re within the midst of a secular transition over from generations of legacy techniques over to the cloud.

So your competitor is definitely the information nonetheless in information facilities on premise.

SR: There’s a enormous quantity of that, this can be a quickly increasing enterprise and so there may be lots extra. And in areas like migrations, for instance, you’ll be horrified at how tough and the way lengthy a few of these migrations take. A part of what we’re very enthusiastic about is, “How do you apply comparable strategies?”, primarily the AI-assisted human loops for making issues like migrations go quicker.

However again to your fundamental level, although. I do assume that the world of AI, the world of interactive consumption of data does create attention-grabbing new alternatives for merchandise that aren’t competing with the Tableaus of the world, as a result of that’s a wholly new class.

However again to your factors about innovation that you’ve got made in different podcasts, I believe AI-driven consumption goes to be very broad and considerably shallow to start with. It’s by no means going to have the sophistication of a dashboard that somebody spent six months creating. Then again, many, many extra customers can get information method quicker than needing to attend for an analyst.

Over time we’re completely going to have a chat to my pivot chart widgets the place you’ll be able to manipulate a pivot desk with out actually needing to know programming by merely saying, “Hey, drop this attribute because the X attribute”, or, “Compute a measure like this”, the place you might be extra assistively creating issues like this. These are the alternatives that I believe Snowflake is creating. And once more, tremendous aligned and complementary to our strengths as a knowledge platform, we’re not attempting to reinvent ourselves to be another new firm.

You talked about a pair occasions the mix of services, and it’s humorous you talked about system migrations. Are we in some respects going again to the previous days the place truly what must occur is individuals will undergo this honeymoon interval of, “Yeah, simply throw an AI mannequin at it”, however truly what is critical is 6 month, 12 month, 18 month migrations, refashioning your information, after which the AI is then helpful? It’s truly downstream of doing a variety of work up entrance that everybody thought we have been executed with. However now it’s again to the ’70s, again to the ’80s, because it have been.

SR: No enterprise individual goes to inform you that they’re executed with their enterprise migration from on-prem techniques. I problem you, you’ll be able to go speak to any giant financial institution, they are going to all inform you that there’s a ton of that work. What’s totally different about right now is I believe we’ve the potential to dramatically lower the timeline of those sorts of migrations. We’re doing one migration for a really giant financial institution that, it’s an 18-month migration. It’s simply crucial information set that they’ve they usually’re terrified, and so are we, about making any mistake with that migration.

Isn’t there a bit, although, the place if we find yourself in a world the place migrations take a very very long time and that’s downstream from, “We simply realized information must be actually well-structured”, and that’s a giant a part of the method. Isn’t {that a} world that’s good for Snowflake since you’re on the extra well-structured information aspect of issues already?

SR: That’s appropriate. That’s proper.

So migrations for the win.

SR: Migrations 100% for the win. Migrations quicker for the win. I’ve even met those that have advised me, “I’ll have you ever begin on one occasion of this legacy database that we’ve, do job fast and you’ll have 50 of those”, and generally that finally ends up taking three months, which doesn’t sound like a very long time, however 50 occasions three is a really very long time.

Do you want a fair bigger service group, then, going ahead?

SR: I don’t assume it’s a matter of individuals, I believe it’s a matter of a lot better know-how, and I essentially consider within the human-assisted AI loop the place AI can resolve courses of issues which might be simple to unravel and the tough ones might be solved in a extra assistive trend. I believe that may drive a variety of progress in these gnarly areas which do have a variety of nook instances.

How do you see this taking part in out in the long term? Clearly you’re going to be optimistic about Snowflake, however how do you overcome this? Your promoting level so far has been comparatively easy-to-use however completely predictable and having that be a bonus versus the beautiful unstructured, however truly that’s factor as a result of, it’s like there’s the previous saying there again when individuals have been much less bullish about Netflix, “Can Netflix grow to be HBO earlier than HBO turns into Netflix?”, each ended up being pretty improper in sure regards, however there’s a bit the place, can Snowflake grow to be unstructured and broad-based earlier than unstructured and broad-based options grow to be precise and helpful and predictable for making your stunning dashboard? Is {that a} honest method to consider the competitors?

SR: Nicely, we’re energetic gamers on this, we get to jot down the historical past as a lot as anybody else. We’re not recipients of this historical past and I’ll argue that the traits that make Snowflake nice, which is a simple-to-use, tightly-integrated product, just isn’t one thing that you would be able to post-fact bolt onto issues.

I like the hyperscalers that we work with however I inform those that they’re actually 300 GMs which might be all competing to have their product line succeed. No person is considering, “What does it take to create one unified, tight information platform?”, that’s been our obsession from day one, and that’s the place we predict we’re set as much as win. Clearly there are various different facets of computing that the hyperscalers are going to be nice at however, much like how there are two or three nice mannequin makers on the planet, despite the fact that the hyperscalers, once more, clearly have extra money than God, I believe creating a good unified information platform is tough.

And I’m biased, however I’d wager my cash any day on Snowflake turning into a lot better with unstructured information with a good product than a sprawling product constructed on unstructured information turning into amazingly good and tight and simple to make use of on all information.

So in case you have the hyperscalers, they’re higher at infrastructure, they’ve benefits there, they’re going to finish up sticking there, it’s truly tougher to go up into the platform layer and also you’re on the platform layer — that leaves the SaaS layer. And also you talked about earlier than, individuals wish to get their information out of their SaaS apps. On the identical time, when you’re a SaaS app, that’s fairly terrifying for comparable causes that it’s considerably scary so that you can lose management of the information that you’ve got.

Do you see the market pressures on SaaS firms resulting in this, the place they are going to construct integrations with Snowflake, extra simply get their information into there? Or are you going to should extra vigorously apply that market strain and having your personal functions, your personal vertical approaches that claims, “Look, simply change over as a result of they’re going to be sluggish, they’re hoarding the information. If you would like it built-in, we’ve a greater answer and it’s higher as a result of it’s tied to all the things else”.

SR: Initially, we consider in partnerships. We work with quite a lot of SaaS suppliers, we frequently have bidirectional information integration with them the place our prospects can convey information from the SaaS platform into Snowflake, but when they wish to take a knowledge set and put it into that SaaS platform for analytic functions, we completely facilitate it.

I’d truly say that on the subject of the SaaS gamers, whether or not it’s ServiceNow or Salesforce or SAP, motion has shifted over to, “What do agentic workflows imply for these merchandise?” — there’s a motive why they’re all stressing that half as a result of I believe agentic AI, agentic workflows provide an alternate to the individuals which might be doing work on these platforms. That is a part of the rationale why they’re leaning much more into these sorts of options. We work actively with them, it’s a big market. We have now some prospects that transfer information from Snowflake to these platforms, we’ve many others that convey their information from these platforms to Snowflake for central evaluation and followup actions like we do.

However there may be additionally one other layer of motion and interplay that is happening. As you already know, the world of agentic AI could be very complicated, it’s onerous to seek out a precise definition for what an agent is or what a part of an agent is, however there are efforts at making this stuff interoperable, so I do assume we want to a world during which, for instance, an organization makes an agent that makes it simple to question one or take motion on one information set, and that being interoperable with one other agent that’s sitting on high of knowledge that’s in Snowflake.

In reality, we wish to assist individuals create these sorts of composite workflows. One in every of Snowflake’s superb choices is that this factor referred to as a Native Software, the place a supplier, say S&P International, cannot solely ship information but in addition ship code as this native utility right into a buyer deployment. Now this code can not have a look at something within the buyer deployment aside from what the client says this app ought to look, however the buyer can not see any of S&P’s information, and now you might have an utility.

To do this, does that buyer, it might have a look at information they’ve — that must be on Snowflake?

SR: That must be on Snowflake. However the buyer has to grant permission to this app after they set up it.

However the stunning factor a few mannequin like that is that this app can now take motion and provides insights based mostly on each S&P’s personal proprietary information and the information from the client. So assume S&P International has nice details about all shares and let’s say your portfolio is sitting with some asset supervisor. Now this app can provide a view of, “What are essential metrics to your portfolio?”, that’s now a mixture of information that solely the asset supervisor has and what S&P has.

Now take into consideration how this mannequin works in an agentic setup. Now you can have brokers, each from the client aspect in addition to the applying supplier aspect, that may come collectively to do significant orchestrations so I do assume that there’s going to be integration at that stage. In reality, we’re working with Microsoft to have Snowflake Intelligence parts be out there as a part of their Copilot and Energy BI and issues like that, so this can be a quickly altering world. And sure, the SaaS gamers should lean into it, as a result of I believe agentic AI gives choices for the mix of individuals and merchandise which might be their mainstay.

So what does success seem like? If we’re speaking once more in 5 years and persons are like, “Look, salute Frank, he took the corporate to IPO, however wow Sridhar actually remodeled what Snowflake is to be greater and higher and higher”, what does that Snowflake seem like?

SR: As I stated, I believe we’re in an period of unprecedented change and alternative. We made $3.5 billion final yr, rising at 30% and if we fulfill our mission to actually assist enterprises mobilize information, we needs to be rising at quicker than that clip for the subsequent 10-ish years. I used to be a part of Google Search rising from $1.5 to shut to $100 billion the yr I left. In the event you’re asking me about an audacious ten-year objective, that may be an audacious ten-year objective.

Headlines: Snowflake To Be The Subsequent Google. That feels like an excellent, nice plan.

SR: (laughing) Clearly aspirational, nevertheless it’s additionally the journey alongside the best way, Ben. I believe making a staff that prizes excellence, that’s actually good at what it does, that may take care of change itself and assist prospects navigate change at the moment of each alternative and hype can be a fairly highly effective mission and I’m reminded of that each time I speak to our prospects, our sponsors inside our prospects, they put a variety of belief in us and to be that accomplice that’s the information arm of what they do, I believe that’s the actual privilege.

What I inform individuals is that probably the most celebrated firms of the twenty first century, like Google, like Meta took information nearly as critically as they took their fundamental product. In reality, it’s these suggestions loops which have created greatness and I inform our prospects that it’s our imaginative and prescient and my dream that Snowflake is that information accomplice for them to grow to be that environment friendly and that insightful with their information as these nice firms have been. To me, fulfilling that mission with increasingly prospects is, I believe, reward. However sure, monetarily or when it comes to simply development, aspiring to issues like mid-30s development for a decade, that compounds.

Sridhar Ramaswamy, nice to speak with you and I sit up for speaking to you once more quickly.

SR: Thanks, Ben. Take care.

This Day by day Replace Interview can be out there as a podcast. To obtain it in your podcast participant, go to Stratechery.

The Day by day Replace is meant for a single recipient, however occasional forwarding is completely tremendous! If you need to order a number of subscriptions to your staff with a bunch low cost (minimal 5), please contact me instantly.

Thanks for being a supporter, and have an excellent day!