E53: Tomasz Tunguz on the Evolution of the Data Ecosystem

0:05Welcome back to Turpentine VC, a podcast where we discuss the art and science of building successful venture firms, VC to VC. In the episode ahead, Tom Tenguz, GP of Theory Ventures, returned to the podcast to talk about the past, present, and future of data businesses, specifically through the lens of Databricks. Tom offers a bull and bear case for Databricks, an analysis of the market and main competitors, and what's next for the data landscape in the AI era. Let's dive in. Tom, welcome back to the podcast. Thanks for joining. Pleasure to be here. Thanks for having me on, Eric. so we're going to cover databricks today maybe let's start with your journey with databricks and when did you think hey this is going to be a you know pretty spectacular company i remember meeting them at the a and the b i mean meeting ali and um uh i mean he's a force of nature uh and clearly like i mean he was coming out of the the berkeley labs and they were commercializing spark spark was a very difficult technology to run by herself and databricks was building software to manage it.

1:07Maybe it was the B and the C, not the A and the B. And, you know, many of the leading data scientists were using this technology as a way of manipulating enormous data sets to get them to, from whatever unstructured raw place they were into whatever place and packaged output they needed to be. And I think he's just done a phenomenal job building a really aggressive company. I think the story, one of our founders is building within the data ecosystem and he went to an event with Ali. And I think the story really sort of encapsulates his business aggression, which in my view is an extremely good thing.

1:46He said, he got on stage talking to all these different startups and he said, we would love to partner with you and we will be competing with you. So if you win, great, we'll work with you. And if we win, well, we win. I don't know if it's exactly what he said, but that was his sentiment, which is, this is an open, competitive playing field. We are playing to win. Let's be aggressive. Yeah. And talk about the company itself. What do you think was sort of the insight or what do you think about like what makes Databricks so special or what makes it work and situate it within a larger kind of context of the market?

2:23Yeah, great question. Okay. So let's say you're building within the data ecosystem. Then the classic data infrastructure has been, the output has been BI, right? Business intelligence. The CEO wants a dashboard of how many widgets am I selling per region per month? And the way that that's built is there are transactional systems. So let's say you have like a point of sale system that measures how many widgets you sell. That data is moved from the point of sale database, which is called the transactional database using ETL. It's basically just extraction, transformation and loading pipes like 5Tran into a cloud.

2:53data warehouse like a snowflake. It's then a snowflake gets manipulated, reconfigured, and then goes into a dashboard. That's what we would call sort of the modern data stack. And that works really well for relatively small data volumes. And it works really well for structured data numbers, things that you would find in an Excel spreadsheet. There's this entire other world of what's called unstructured data, which are things like raw text or video or images, and even really, really large files where that doesn't work because it's really expensive. And because they're so large, the latency doesn't matter so much.

3:27You may not need down to the second latency. And so what you want are bigger systems that are burlier, that can handle much larger data volumes and can parallelize them and manage them over large numbers of servers. And before the world of Databricks, there was a technology that was called MapReduce that Google had built that did this and which became commercialized in Hadoop in the form of Cloudera and Hortonworks and MapR. Those were the three companies that were sort of building this technology. And the way that it's been described to me as sort of the most tangible difference between these two technologies is Snowflake is really to deal with like data frames.

4:05And then what the Hadoop vendors, those were built by people who were file systems people. In other words, like how do I manage logs files or video files? and even when i was at google i was trying to play around with map reduce and i had candidly i had a really hard time it was very difficult for me to understand how do i set up a map and the reason it's called map reduces like a map this is how all these fields map together and then here all the here how all the files kind of compute and go out and get you to a reduction which is the that particular nomenclature for an output and so what spark said is okay file system people built this really sophisticated thing that can manage unstructured data and process it.

4:47But a lot of data people are having a hard time. So why don't we build this really sophisticated pipelining infrastructure at scale in a way that data people can understand? And so that's a very simple sort of reductionist view of why Databricks has had so much success is because data people now can take advantage of absolutely massive computing resources to build these data pipelines. So what are some of the applications of a big unstructured data pipeline. Let's say you want to train a machine learning algorithm to identify license plates on, I don't know, at the scale of New York City Police Department.

5:21Well, you use Spark. You'll use Databricks as core open source technology, and you'll piece together a constellation of different open source projects that take those video files, put them into a folder, effectively process them with code, and then spit out an output. And a lot of the times that output is just either refined files, but now you're seeing Databricks actually starting to compete with Snowflake. You can look at the data warehousing revenue has 4xed in the last year. It's gone from about 100 million to 400 million in a single year. And two years ago, we would ask enterprise buyers, are you using Databricks for BI?

5:57And the answer broadly was no. It was very difficult to find anybody. And now we're hearing about them more and more. So these two worlds of the structured data and unstructured data are starting to converge. And that's why there's increasing competition between Databricks and Snowflake. Why didn't the solution exist before Databricks? Like what took so long? Was it sort of this had to be the right time for the company or was it just too hard to do? Or talk about the time, the why now? Well, most companies were not dealing in huge volumes of data, right? Google was the first one. And, you know, the apocryphal story is that Google, when they started, they were going to use very fancy Oracle databases, but because they were crawling the web, it was just too expensive.

6:34So they needed to build a new technology to build MapReduce. And that's what started. But Google, there were only a handful of companies that were really operating at web scale. Yahoo also was operating at web scale. Yahoo created Hadoop. And so Hadoop and MapReduce kind of came to be. Then over time, everybody, once we saw the business model of Google, everybody in the world said, okay, now this is the era of data. We need a lot of data in order to compete. And so data volume started to increase and the market size, the addressable market of people needing massive data infrastructure solutions started to grow.

7:07And as that started to grow, the way to capture the market was to improve the user experience so that the data people could actually use these technologies in a much more sophisticated way than they could with the MapReduce Sadoop world, which was targeted to basically a different user. Did Databricks have, what were the major competitors along the way, or is it really just such a project that they didn't have any there it's hard to point to a competitor um it really is i i don't know i think at this point you have snowflakes uh snow park which spark spark is databricks's technology snowflake came along a little bit later and created a technology called snow park very deliberately which is um snowflakes equivalent there's been some adoption but it hasn't sort of derailed the locomotive that is Databricks, that's probably the most viable competitor of any sort of scale.

8:05Yeah. And was Databricks obviously investable? At what point did it become obvious that it was going to be a pretty big company? At the beginning, I would say the venture ecosystem as a whole believed more in structured data than they believed in unstructured data. And that's because, remember, we were coming off this wave of Cloudera and Hortonworks and MapR. And there had been massive commoditization in that market. When Cloudera started, list price for a node, in other words, running Cloudera on a single server was about$4 ,000. And then at the end of like five or six years, it was about$1 ,000 a node.

8:39And there weren't that many players, but they're all trying to win relative share. And so they just kept cutting price and price and price. So there was a fair amount of skepticism, I think, in the venture community overall that in this part of the ecosystem, somebody could charge a premium. There was also, I think, a fair amount of skepticism that was a lot of value within that unstructured data. So you think about, okay, what is unstructured data? Well, maybe it's logs, right? The value per gigabyte of a log relative to the value of a gigabyte of a point of sale system, massively different. Log is not that valuable.

9:09Very few people actually look at it. The value of one frame in a video is not that valuable. And so you're basically pressing these giant lemons, not getting any juice in unstructured data, or at least that was the perception. And whereas the converse was true within structured data. And now with this AI wave, the reality is the transformer model, well, that's totally enabled by extremely large compute and extremely large data sets. So you could argue Spark and that technology is, and the related technologies are basically a necessary or requisite to getting to where we have been in AI. AI is just massive tailwinds for Databricks, basically.

9:48Like as the market has matured or developed, it's only just better for Databricks. Yeah, that's right. I mean, I think the reason that there's so much interest in the company in the late stage markets and the reason why I think when they go public, you'll see a significant amount of demand is it's thought of as the AI company, the AI infrastructure company, not in the way that OpenAI is thought of as the model provider. But if this is like the data movement company, if you're moving data, if you're manipulating data, if you want to set up pipelines to train machine learning systems or modify them, you will be using Databricks predominantly.

10:20uh and so that's why there's there's just so much demand for it it's interesting so let's go to your your post a bit so you were starting to explain basically how databricks and snowflakes are intersecting but let's compare and contrast the businesses a a bit more well so i was at both the databricks summit and the snowflake summit of databricks event which was two two two and three weeks ago respect well snowflake summit first and the databricks summit san francisco and you can get a sense of the difference in the buyer universe snowflake summit probably director level and above a lot of the topics covered are about transformation of a company from wherever they were to the future of data data bricks is attended by engineers a completely different prototype completely different phenotype of individual coming to those two events and five years ago they didn't really overlap why was that well we talked about the modern data stack and how most of data movement, particularly for structured data, is the creation of reports.

11:18Well, it turns out that there are three categories there. There's business intelligence, Omni, and Looker. And then there's data exploration where I'd put Hex. And then the last is machine learning systems. Historically, machine learning systems have all been post hoc offline. In other words, CEO asks, can you cluster the customers? Who are the customers most likely to churn? What is the revenue prediction for next month? Those are not part of a product that is delivered to customers. that's something that's presented to a ceo as a report and then when we had the launch of the transformer everything changed and now all everybody wants ai and in the old world the people building the reports for the ceo were not part of the engineering organization you had a vp of eng and then you had a vp of data and those are two parallel structures where they might use the engineering teams were producing some of the data and the data teams were consuming it but That was it.

12:05That was the end of the data pipeline or the data supply chain. And now what's happening is the engineers are saying, actually, you know that data that we're producing around points, whatever, people buying widgets? We want you to send it to the data team. We want you to make a machine learning model out of it or an AI system. And then we want to put it into the path of production. We want to make a recommendation system or a chat bot for customer support. And so instead of the data dead ending within the data team, it's now basically a circle. The two teams are working together and the two teams are converging.

12:37And so the Snowflake people are now starting to talk to the engineers. And the engineers are saying, well, we need these production grade systems. And so these two teams are coming together. And as a result, Databricks and Snowflake users are starting to see each other more and more and more. The big question strategically is if you're a large company, do you standardize on one? Do you standardize on Databricks? Do you standardize on Snowflake? Do you have some teams running on Databricks, some teams running on Snowflake? And the rationalization of that infrastructure, I think will be sort of the defining, the battle of the giants here over the next five years.

13:10Hey, we'll continue our interview in a moment after a word from our sponsors. How deep do you go to seek out an answer to a question? Maybe you've spent hours clicking the source links on an obscure Wikipedia page, or maybe you're even the type of person who checked out the entire shelf on the topic at your library. If you're nodding along, then check out GiveWell. an organization that researches questions about global health and philanthropy, even if a satisfying answer might require years of reviewing studies, talking to experts, and over 300 footnotes. GiveWell has now spent over 17 years researching charitable organizations and only directs funding to a few of the highest impact opportunities they've found.

13:45Over 125 ,000 donors have used GiveWell to donate more than$2 billion. Rigorous evidence suggests that these donations will save over 200 ,000 lives and improve the lives of millions more. GiveWell wants as many donors as possible to make informed decisions about high-impact giving. You can find all of their research and recommendations on their site for free. You can make tax-deductible donations to their recommended funds or charities. And GiveWell doesn't take a cut. If you've never used GiveWell to donate, you can have your donation matched up to$100 before the end of the year, or as long as matching funds last.

14:17To claim your match, go to givewell.org and pick podcast and enter econ102 with Noah Smith and Eric Torrenberg at checkout. Make sure they know that you heard about GiveWell from econ102 with Noah Smith and Eric Torrenberg to get your donation matched. Again, that's givewell.org to donate or find out more. Present kind of the bull case and the bear case for Databricks. Like what's the, if everything goes well, what does that look like? And if things don't go according to expectations, what does that look like? Well, the bull case, so the company grew 50 % last year and then is on a 60 % growth rate at scale.

14:57So there's a meaningful inflection in growth. Very rare to see a company actually accelerate, primarily driven by AI. The bull case is that Databricks becomes synonymous with AI infrastructure. Anybody building AI systems will be building on Databricks. combine that with starting to win some significant share away from existing Snowflake structured data customers, and it becomes a dominant platform. I think the counterpoint to that would be Sridhar now as a CEO of Snowflake is able to make pretty meaningful inroads and retain and continue to grow the existing Snowflake customers by presenting an extremely compelling AI story.

15:33And you've seen that, right? At the announcement when Slootman appointed SweetR and announced it, there's been significant recruitment of people with material AI backgrounds. You have the NVIDIA partnership. You look at what they've launched with Cortex, which is their AI product, and the announcements at the Snowflake Summit. In addition, just the cadence of PR has been completely different, much more aggressive, trying to position themselves within the world of AI, developing their own model that competes with Databricks' model and others. So I would have to say the pace of innovation from Snowflake is pretty significant.

16:10The question is, will customers move? And the answer so far, at least according to our diligence, is that it's very hard for a company to move from Snowflake to a Databricks. You're talking about pretty significant change management. And so at least today, we haven't seen much of that. And then, if I could just go on for a little bit longer, one of the accelerants of that transition and one of the justifications for the tabular acquisition, which was Databricks buying a company that commercialized a technology called Iceberg, is the ability to move data and particularly compute. So, okay, what do I mean by that?

16:49Well, a database is, there's what's called a query engine. So you ask it a question, it does a bunch of math, and it comes back to you with an answer. But when it asks those questions, it reads a bunch of files. And a database is basically the combination of those two things with a bunch of other stuff, security caching. But it's a query engine and a bunch of files. And Snowflake is a database that's optimized for a particular thing, structured to data analysis. And Spark's Databricks, or Databricks' Spark, is the same thing. Well, what's happened is we're starting to see the unbundling of the database.

17:21Well, now there are many different kinds of compute engines. There's Snowflake, there's Spark, there's MotherDuckDuckDB, there's PrestoTrino, Dremio, and graph databases. And what customers are asking for is, I want to pull my data out of the database. And why do big customers care about that? Well, it's expensive. It's expensive to store data two or three times in two or three different database centers. It's expensive to be compliant with international law if your data is in a whole bunch of different places. And now with these AI workloads, these companies actually want all their data in a simple place because if they decide to build some cool AI feature, it's just much easier to have it in one repo, one repository.

17:58And so Databricks bought a company called Tabular, which commercializes this data format called Iceberg, which allows you to basically move data across different workloads. So if you decide, I need an dashboard for the CEO and it has to have 60-second latency probably going on Snowflake. But if there's something that's a little bit slower, a little bit much larger, you might move it on to Databricks Spark. And the more this trend happens where the customers are in control of their data, the more compute workloads will start to work around, the greater the competition at the query engine layer.

18:29But you're much more bullish, right? Like of the two cases that you outlined, you think it's much more likely the bull case or say more about that? Oh, I think it's, yeah, yeah, for sure. I mean, look, to see a public scale company growing at accelerating their growth rate, I think they'll do just fine. If you, we ran a linear regression based on five key variables, which predicts about half, you know, has about prediction actually about 50%. That model basically predicts that Databricks would trade at about a 14x forward multiple, snowflakes roughly around 12. but I think if it were to trade today, if it were to list, it would probably trade significantly higher than that because of the public market appetite for AI stocks, which doesn't exist.

19:13I mean, the only way you can get AI exposure today is Microsoft, which is a massive conglomerate or NVIDIA, which is a fantastic chip company, but chip company nonetheless. And so there's a void there. And so I think when it lists, it will trade incredibly well. Yeah, it's really interesting. If you had to put like a basket of companies that if, you know, you just that similar to Databricks, you're making a similar kind of bet. What other companies would you put there? Well, public is pretty hard. I mean, I think, okay, so let's say you had a dollar to spend,$100 ,000 to buy any stock and you wanted to have maximum exposure to AI.

19:50Databricks would be in there. Snowflake would be in there. OpenAI would be in there. Microsoft would definitely have to be in there. They have a business that went from zero to 5 billion run rate in about 18 months. And that's just, I think, the beginning of their ability to cross-sell. I think deeper than that, it's hard. It's hard because you have commoditization at the model layer. And we can sort of talk about that. But if you look at like the large language models, they range in sizes from about 2 billion parameters, let's say, to about 175 billion parameters. And there's a benchmark that's called the MMLU, which is high school equivalency.

20:21So how good is this particular LLM relative to a high school grad? and they're all basically converging to somewhere between 70 to 80 percent of the performance even it's for the very very big ones so i think it's it's hard to kind of invest at the model layer particularly for and the capital intensity so then you probably have to go application by application okay let's take a look at code automation let's take a look at legal automation let's look at accounting automation let's look at security automation there the companies are much younger so i'll probably stick with my my mega cap basket for now that makes sense what do you think people most misunderstand or underappreciate about the company?

20:57What do they not fully understand or what do they get wrong? I don't think the technology and what it does is broadly understood. You know, a lot of people think about databases and an easy mental construct for it is just a really big Excel spreadsheet. And what Databricks does is imagine you had like a million Excel spreadsheets and they all had images in them. And what you wanted to do was, and all of those images had license plates on them. And let's say you wanted to count the numbers of A's and B's and C's and 1's and 2's and 3's, you would use Databricks to do that. And that's not a very sort of like grounded example, but just to give people a sense of scale about what the ultimate technology does, that's a really hard technical problem.

21:42And it's a hard technical problem not because the optical character recognition, like looking at the 1's and 0's and A's and B's and C's are hard. It's a hard technical problem because you're talking about thousands of computers all working together at the same time to achieve a single result. And so it's like I was a rower in college. And so it's like a rowing boat. But instead of eight people in it, you have a thousand people. How do you coordinate that? Yeah, that's really interesting. What else about the company do you feel like we haven't gotten into that is maybe interesting for people who are either students of the company or students of business more generally?

22:17I think a big push is the data catalog. And the data catalog was a strategic announcement for both Snowflake and Databricks at their most recent summit. Data catalog is the address book of the data ecosystem. So let's say, Eric, you create a table about number of podcast listens by episode by date. You might own that table, right? And I would create one about venture capital investments. I would own that table all within a company. Each of these data fields are related in some way. They all start from the same source tables and they all kind of evolve in a big company to the ultimate product that they'll be.

22:59And there's this strategic push to control that data catalog. Okay, that data catalog can be used for lots of different reasons. One is I need to look up who is the subject matter expert in a 10 ,000 person company who owns the podcasts table. Okay, it's Eric. There's another reason to understand the data catalog, which is lineage. How do all these fields actually come together? Lineage is important for compliance. So if there's an external data set that touches social security numbers, I really need to know that to make sure it's super secure. there's data quality which is is the data that's being fed through all these pipelines actually accurate are the distributions remaining the same are the data volumes constant so that the people who are just using deciding using data can be confident in it there's um bi anyway so that if you can think about like the data catalog it's basically the nexus it's an understanding or it's a roadmap of how all the data within a company fits and if you can control that and understand that, you can build a whole bunch of products around it.

24:02The challenge with the data catalog historically is nobody wants to update. It's like a wiki. Somebody creates a page once and then 10 years later, somebody comes back and visits it, but it is super valuable to a company. It's like an internet. And so now there's a strategic push to control that data catalog for both Snowflake and Databricks. There's an open versus closed source dynamic here, which also we haven't talked about, but maybe I'll just pause there on the data catalog. Yeah. Let's talk about the open-close dynamic. Sure. So Databricks is an open source project. I mean, Spark is an open source project, came from a lab, very difficult to implement.

24:37The commercialization motion for Databricks is the infrastructure to be able to manage tens or hundreds of thousands of machines operating this particular open source project, a little bit akin to Kubernetes, which was a technology that came out of Google. The company has done, Databricks has done a really nice job commercializing it. When they started to move into structured data workloads, they created a technology called Delta Lake, which is effectively a cloud data warehouse. And that was closed. That was closed source, which was a bit of a departure. And I think, and I don't know this for, you know, this is just my conjecture.

25:13I think that was basically a way of ensuring that other players within the ecosystem couldn't leverage the existing technology that Databricks might have built. And it was also a way of keeping people, their existing customers within a closed universe. What's interesting is Snowflake started closed source and is now pushing more into open source. And so they're basically like complementing each other. One was closed source moving into open Snowflake. Databricks was open. It's now moving more into closed. We'll see what they do with the iceberg acquisition. And so there's this like, it's like two sumo fighters trying to figure out, okay, where am I weak?