He built a new database in his bedroom—now he powers Cursor, Notion and Anthropic. | Simon Eskildsen, Founder of turbopuffer

I was thinking about this from the second I woke up until the second I went to bed. During the evening when my wife and I would have dinner, I was running like different runs, right, to test different theories of how this would work and experiments and stuff like that. I think she might have a better representation of exactly what that summer was like from her point of view, but I remember it as being completely all-consuming. I need to go meet them. And so I flew to San Francisco. I showed up on a Monday afternoon and they were having problems with the Postgres database. And so I just tried to teach them everything that I knew about running Postgres at scale. That must have convinced them that maybe I knew a little bit about databases because they felt compelled enough that they started the migration. And of course, this company is Cursor. Notion had finally ingested their first workload and it was really slow. Justin found 300 milliseconds in the span of 3 hours that night. There was another time where Notion asked, can we do this, Cory? And Justin just told them in Slack, yes, we can. And then worked like I've never seen before for the next 24 hours to get it in their hands. When people talk about forcing something or willing something into existence, this is one of the first things that I think of. That's product market fit. Product market fit. Product market fit. I call that the product Product market fit question. Product market fit.

Product market fit.

Product market fit. Product market fit. I mean, the name of the show is Product Market Fit.

Do you think the Product Market Fit show has product market fit? Because if you do, then there's something you just have to do. You have to take out your phone. You have to leave the show 5 stars. It lets us reach more founders and it lets us get better guests. Thank you. Simon, welcome to the show, dude. Thank you for having me. There's very few people I interview in this show that happen to also be from Ottawa, but you're, you're close by, which is nice to see.

Yeah, I'm not from here. I moved here back in the day for Shopify. That's how I ended up in the new world, but it's actually an underrated place to build a company from. It has its advantages and disadvantages.

We'll get into all this. You're building Turbo Puffer. You've got some massive logos, right? Cursor, Notion, Linear, like massive companies using your product and you've done it all bootstrapped purposefully. So intentionally, so like you're keeping things pretty much under wraps. Let's start at the beginning. Like, give me just a little bit of your background. You mentioned where did you move from?

Yeah, I should specify that we're not bootstrapped. We have taken some funding back in early 2024. We raised from one fantastic partner, like sub-a-million or like around there, like kind of small pre-seed type amounts, a round that would look laughable in the face of what the rounds look like today.

Gotcha.

So I grew up in Denmark. I moved to Shopify to work there as a software engineer as what I thought was going to be a gap year back in 2013, moved over to work on the infrastructure team at Shopify. And Shopify was at the time only had an office here in Ottawa where we both are today. Of course, now today they have many offices around the world.

Yeah, 2013 was pretty small. Like how many people? They're pre-IPO, they're what, like Series C company or so?

That's right. They were around 150 people when I joined, maybe around 200 between when I interviewed and when I joined. Spent almost the 10 years there building infrastructure on the infrastructure team. It was a very small team at the time and grew quite a bit as Shopify grew. We were doing a few hundred requests per second when I joined and we went into, saw flash sales in the millions of requests per second by the time I left. The majority of what you spend your time on, if you're scaling infrastructure, are the databases and the database layer. And that's where I spent almost all of my time, caching databases and everything. The database that brought me the most trouble, both in terms of the product that we could ship and in terms of the operational experience, was the search engine at Shopify. And I never thought that I was ever going to work on that again, but after Shopify, I encountered the problem again of working for a friend's company, an actual bootstrap company. I worked there for a few months just helping them with infrastructure challenges on doing a short consulting stint. And they asked me to build a recommendation engine. And so I built a small recommendation engine and used some vectors because, I mean, vectors are great, right? They're a very good representation of what content is.

Yeah, maybe give like the non-technical kind of explanation of where this comes into play.

The way I think about a vector is that if you take an LLM like ChatGPT and others and you sort of cut off their head, then what you get is you get a bunch of numbers around around, you know, maybe 4,000 numbers. And what these numbers are really just a coordinate in a coordinate system. You can think about this in two dimensions where if you are Spotify, you plot all of the songs that you have into two dimensions, and the ones that are similar are gonna be adjacent. And this means that you end up with clusters, right? You're gonna have a rock cluster, you're gonna have a pop cluster, whatever. You zoom in, you're gonna get all those subgenres. So you can train a model that is very good at plotting unstructured data like text, songs, images, video, whatever. Some people call this multimodal. Um, and what really just is, is training a model to put content that is similar adjacent in a coordinate system. The example I often use is that in your Shopify, for example, when you search on a store, you search for a red dress and they only have a burgundy skirt. Well, that's actually a very difficult search problem. It's devilishly PhD-level difficult to do in a generic way with a classical model. ChatGPT can do this with ease, right? It can see that those things are very similar in the coordinate system. They'd be right next to each other. And so that is what a vector is. A vector is a coordinate in a coordinate system. It turns out that we can't just plot things in two-dimensional space. There's too much information to try to compress, which is what a vector is. And so we use these, these coordinates that are really in thousands of dimensions or at least hundreds of dimensions. And the problem with these things is that they are very large, but the benefit is that they're really good at doing recommendations and doing search. And it means that we can have just search that is really good as almost just like byproduct of training these LLMs. Not quite, but it's not a terrible way to think about it because there are key functionalities to—

I mean, in this case, in this use case, to just know what is alike, like what is this similar to in some of these dimensions.

Exactly. I mean, it's one of the things where ChatGPT and these LLMs are best at. You ask it for definitions, you ask it for more words that are similar to this word, or you ask it for what emojis are suitable for this particular thing. They're so good at finding information that is similar to other information. That is how they're trained. And so vectors are that, they're just a way to do that that is really applicable to search. And so what we do is that you take text, you take images, you take whatever, and you turn it into one of these vectors. And we don't do that. We don't take the data and create the coordinate. We just store the coordinate system. So if you were Spotify, Spotify is not a customer, but if you were Spotify, then you would take all those songs and put them into, into TurboPuffer, and then you would make searches like, okay, Pablo, we're going to make your Discover Weekly now, or some recommendation. Let's look at everything that Pablo has listened to. And then search for vectors or songs that are similar in space. That's what a vector is. And Turbo Puffer is the coordinate system. You can ask it questions. And the innovation of Turbo Puffer is that it is almost 100 times cheaper than the incumbents at the time to store that coordinate system.

And by the way, you— why did you leave Shopify in the first place?

I left Shopify because I have been there a very long time working on infrastructure, and I wanted to see infrastructure in other places. I thought that it was time. You spend 10 years in one place and you want to see computers in a different place.

And you left for your friend's startup, or you left and then kind of that happened?

I left to get in the best shape of my life at first over that summer. And then I started working with my friends' companies. I called it angel engineering, where I went for a couple of months to my friends' companies, vested mainly equity, and helped them with some infrastructure challenge. I did some of that before I did this particular one, but this was the one that was part of the founding journey of TurboPuffer, where we mostly database scalability stuff.

Stuff.

And for this one, what I did was build this recommendation engine. This was a real bootstrap company that had spent $3,000 or so per month on their Postgres. And I ran the back-of-the-envelope math for this recommendation engine I built with the coordinates, right? You put these articles in a space and you recommend other articles.

What was actually— I mean, what was the use case? Like, what do they want recommendations for?

Yeah. Have you heard about this company called Readwise?

Readwise? Yeah, yeah, that sounds familiar.

So their first product was you take your Kindle highlights, you send an email every day taking some of your Kindle highlights and showing them back to you and helping you retain you're reading better, but now they have a reader product where you can save articles, you can save books, you can save PDFs, and then read them on your phone in a nice, like, standard web interface, right? And so you can imagine in a product like that, you might want recommendations. I think they still don't have recommendations, but this was the founding use case. So I built a very small recommendation engine and it worked okay for something very simple. It actually works okay. What I learned though was that I did the napkin math on putting this in production, right? They have hundreds of millions of articles. and it would have cost about $30,000 a month to put into production on one of the incumbents at the time. And that was just way too much, right? Like, they could not earn a return on that kind of cost. It just didn't make sense for the per-user cost based on what they were charging. So we tabled it. It was going to work, but we tabled it because the cost wasn't there. I just couldn't stop thinking about that. It's like, why? Like, the token costs will probably come down. They haven't really come down, but like, eventually, you know, they'll come down. Who was gonna do that for search? Because here was one use case in one company. There's gotta be thousands, if not tens of thousands of these.

Was there something that told you that it should be cheaper, like much cheaper than what it was?

The fact that they didn't ship a feature because the infrastructure just fundamentally wasn't at a price point where they could earn a return on what the infrastructure provider was charging them, not because the infrastructure provider was bad, just because that's how they implemented it. Just made me, you know, I, I have this GitHub repo called Napkin Math, and in Napkin Math, I have a list of numbers of what a computer can do. How many gigabytes per second of memory can you read? How many gigabytes per second can you read from a disk? How much can you read over a network? What's the latency from Sydney to New York? All these kinds of numbers. So the way that I approached the problem was, okay, well, is there a fundamental compromise that you could make here where you give something up and you get the cost in return? It turned out to me as I stared enough at the numbers and, and drew this enough time that there might be a way now that you could build a database that you wouldn't have been able to build 10 years ago that would make this really cost efficient. There is this, um, have you heard of S3? Do you know what it is?

Like the S3 bucket, like with AWS stuff?

Yeah.

Yeah.

So for the readers who don't know what it is, S3 is this incredibly cheap way to store data. If you put data in memory, it costs maybe $2 to $5 per gigabyte of data you put in memory, right? And if you put it in an S3 bucket, the data costs about 2 cents per gigabyte. Wow.

Okay.

Like a true 2 order of magnitude cheaper, but no one had really built a database that would be fast enough for recommendations and for these types of— for storing the coordinate system on top of S3. The fundamental thing that you give up is to write into S3 takes hundreds of milliseconds. That's not acceptable if you're building a checkout system at Shopify. Or you are browsing the Netflix catalog or some other consumer use case, it is not an acceptable latency to write new data.

Because that comes up when, like in that use case, like you're checking out, now you have to wait 200 milliseconds for that checkout to go through or whatever?

When you do a checkout, you're doing much more than one write to the database. You're potentially doing hundreds, right? So you multiply 200 milliseconds by 100 and you end up in tens of seconds. A normal database can do a write in hundreds of microseconds, right?

Gotcha.

3 orders of magnitude faster than you can do it to S3. So, okay, you might give up 3 orders of magnitude in write latency, but actually you don't have to give anything else up. And then you have this true 2 orders of magnitude decrease in the base cost of the system. That's a great compromise if you're building search, because when you're building search, you are okay with, you know, if you add a new title to the Netflix catalog or you're adding a new song to Spotify, or even if you're adding a new product to Shopify, It's okay that it takes 200 milliseconds to make it into the search engine. It's an acceptable latency.

It doesn't really matter, cuz the, the thing about search is when you search, you want it to be fast, but when you add something to that database search, no big deal.

Exactly. So we could build that database, S3 and disks and a bunch of other things that are kind of boring and nerdy, sort of meant that we couldn't have built this database 10 years ago, but now we can build it with this particular architecture. And then if you query it a lot, right, we just put it into, into memory, right? If the data is active. And pay a little bit more for it. That's the fundamental architecture that I came up with in 2023 and then sort of locked myself in and came out with the first version of the database over the summer of 2023.

How quick was this, by the way? Like from the time where you hit this, okay, this is not going to be affordable for Readwise until you have the epiphany, like just walk me through that timeline, the epiphany, and then you lock yourself up.

Yeah, so I work with Readwise on this in around September, October of 2022. Right? So this is around when ChatGPT was coming out. The problem bothered me for the next 6 months, and I was working with some other companies during the time. And then finally in May of 2023, a consulting agreement fell through. It's just like it was not going to work out. I was talking to a friend and I was explaining to him, I was like, I think you could build a search engine with this, with these compromises. And God bless this guy, he's one of my best friends. He actually helped me design the website. Well, he designed the website. I mostly just provided really annoying feedback on the website and I think he would agree that he did not understand what I was going on about. Hopefully I'm able to articulate it better now than I was then. And he just said, you should just do it.

He just saw your passion. I mean, he knew you well enough to be like, dude.

Yeah, I was like, should I? Like, this guy's kind of being annoying with this consulting agreement and I don't know if I want to do it, but like, I have this other idea and like, it's percolating in my head. He's like, you should do it. And I said, I think, I think we could do this cool thing with a website where it feels ASCII and sort of retro. And I think there's something there, but not take it too far because then it's going to feel in the uncanny valley. And we're sort of riffing on that. And he's like, you need a name. And I was like, well, how are we going to come up with a name? And he said, you should choose a name that has an emoji that doesn't have any other connotations. So we're just like, you know, looking at our phones on a couch in the Airbnb in Copenhagen, and we find the pufferfish emoji. And we both agree that, okay, well, no one knows what this emoji means, or it doesn't have any other connotation. And then we're sort of throwing it around. I think our wives wake up around this time and his wife says Turbo Puffer, and she said it in an Australian accent, which is, by the way, that and the French accent are the superior accents to say the name in. I'm not going to try to—

You're not going to try, Kwame?

Not going to try. Not going to try.

Turbo Puffer, something like that.

Something like that. And it just made us happy. That's really what it is. Today, the way that we explain it, right, is that you put the data on S3. and it's slow and cheap. And then when you query the data, it gets faster and faster as we sort of puff it up into RAM, right? And disk is an intermediate layer, but that is the origin story. So, you know, September of 2022, October, I'm working on this recommender system at Readwise and various things to integrate the first GPT models that were really good into that experience. Then I did another consulting agreement with another company called Replicate, and then I went and started building this in May of 2023 is when the great lock-in. Happen.

Did you have a sense of just how big the market was, or was this just like, this needs to exist, I'm going to build it, we'll see kind of what comes of it? You know what I mean? Like, how much did you size things out in terms of the opportunity in front of you?

I didn't do any of that. I can articulate why now I did it, but at the time it was— your personal neural nets gradients are trained on your experience. And what I just saw was that there are these vectors, they are enormous, right? So what we didn't talk about with these coordinates is that because they're in thousands of dimensions, the coordinate in the coordinate system is very counterintuitive, but the coordinate in the coordinate system is larger than the original data. If you take a paragraph of text, the coordinate that represents it is actually larger than the paragraph of text. And so it is just so much data. It's so large that the costs are so great that it's very difficult for most companies to earn a return on. And that was the fundamental first insight. So these are, this is really big and it's just too expensive. And the second thing was that S3 buckets and disks and a bunch of other things, there's some technological advancements that meant that we could build a database in a way we couldn't build before. The way that I articulate this now is a little bit more structured. If you want to build a generational database company, you need 2.5 things. The first thing that you need is that you need a new workload, because if you want to make it as a really large billion-dollar-plus database company, you need a reason for more or less every company in the world to need a new database, or at least use new products that use your database. That new workload is that we have AI that wants to be connected to enormous amounts of data. We are the latter part of that, right? We are the coordinate system, right, that you— that AI searches in. When you train an AI model, you're trying to basically compress all of human knowledge into a few terabytes. That's impossible. You can't do that, at least not by any computer science or any science really known. It does not make sense that you can compress all of human knowledge into a few terabytes. What you can compress into a few terabytes is a way to reason about the world and a mental model of the world. But in order for you to use that, you have to have access to tools. and the tool that we are is search over all of the unstructured data potentially in the world, inside of every company in the world. And that's the new workload, connecting data to AI. That's the first thing you need to build a very large database company.

How much was this front and center at that time that this would be the new workload for you?

The new workload was completely overfit on Readwise, right? It was like, oh, Readwise has this new workload, maybe other companies do, right? It was completely overfit on that one customer. So I assumed that that, you know, latent demand is when there is demand in the market that is pent up and you can't access it because pricing or availability, or there's some force at play that means that you can't get what you want. And what Reezwise wanted was a search engine that was about 1 to 2 orders of magnitude cheaper because then it would have been a no-brainer to search it. So I felt, well, there's gotta be 100 other companies. This feels worth at least writing an article about, right? It's the output of the summer and I was in a good position where I was looking for a project and Project Felman. So that was a new workload. Second thing you need to be do to build a really large database company is that you need a fundamentally new way to store the data. That's what we just talked about. Put it on S3, 2 orders of magnitude cheaper than RAM, and then the stuff that you actually need to put in RAM, you put into RAM when needed, but you only puff it up when you need to actually search the data. That's the new storage architecture. None of the existing databases have that storage architecture because otherwise there's no reason why the other big database companies like Snowflake and Databricks and MongoDB and Oracle would not just get good at it unless it fundamentally challenges assumptions that they built on top of and relied on for decades. The third thing, or the two and a half thing that you need to build a large database company, is a way to suck up a lot of data really quickly. You don't strictly need this, but if you're impatient, then you want this, cuz otherwise you need to wait a long time for the compounding function to work on a smaller base. That's what you need. And I could not have articulated at the time, but it felt like that was in the air, right? The way that we think about innovation is that, I mean, it's more of a discovery than it is about doing anything genius. And the discovery here was that there's a set of compromises we can make, and it's uniquely suitable for this new workload, slower writes, and no other real compromise.

The other framework for that, the innovation piece, the Mike Maples thing, right? About like kind of living in the future. And so this idea that you're kind of out there doing things like you're not actively trying to start a company or trying to find a problem, you just doing things as you were doing things, and then you came across this search situation and you just went out and solved it. And then AI happened to be this wave that has made the opportunity so much bigger. It's, it's just crazy how things work out.

That's right. So that summer, it was never a business. It was a project. And it was a project where I thought that I was maybe publish a really interesting article. And I actually did. And I wrote an article about the experiments that I'd done at the sort of like a few, a month or two into it. And I sent it to a friend. And he said, you should turn this into a real service. And that was enough encouragement for me because he was one of the industry leaders in this space and practitioners. And so I just kept working at it and there were some real problems to solve, surely, but it was not something that I felt like I know a lot about databases and so on, but it was nothing that was like uniquely really complicated, right?

And it was just you.

It's just me. And inside of the company, you know, we make a lot of fun of what we call V1. There's almost none of V1 left now other than some pieces, and we all make fun of it. And my joke is that they make fun of it, but it's the reason they have a job. Yeah. Now this doesn't really resonate because these people are so good that it would not be an issue for them to get another job. I was obsessed that summer. It's all I did, and we would have guests and I couldn't stop thinking about it. I was just in a notebook. My wife was very patient, but also a little bit annoyed, rightly so.

How all in? Like, what are we talking about? What would a normal week like be, or day like be for you in those 2 or 3 months where you're building?

I was thinking about this from the second I woke up until the second I went to bed, right? And then if during the evening when my wife and I would have dinner, I was running like different runs, right, to test different theories of how this would work and experiments and stuff like that. I think she might have a better representation of exactly what that summer was like from her point of view. But I remember it as being completely all-consuming. And I remember not being as present as I wanted to be when we had guests and we were hanging out with people because I couldn't stop thinking about it. I was truly encapsulated and I felt that if I didn't do it, someone else was going to do it. And every day the opportunity felt larger and larger. I did 3 rewrites during that time. I tested every single way that I can think of to do it to make sure that what I did was the best. I changed platforms. I tried so many different things. And finally, it was the 4th of October, 2 years ago and change from when we're recording. I finally released it on X.

Did you have a following? Let's start there. Like, what's the context of you on X? Like, are you big? You're posting a lot.

You've got people that follow you, technical people, I just have an egg avatar and 200 followers. No, there you go. Um, no, because of my articles on napkin math and when I was at Shopify, I did a lot of talks at conferences about scaling databases and things like that. I had accumulated a small following on X and so it got really good traction on X. People were encouraging and it was enough for me to feel that maybe I have something here.

Do you remember what, like, what are we talking like 1,000, 10,000 views, 100,000 views? Like what do we, what kind of order, order of magnitude of? Attention did you get?

I can pull up the tweet right here because I was just reposting it the other day. Uh, post engagement, almost 400,000 views.

Okay.

Yeah.

So it was pretty big.

The website was really simple, great encouragement, and I hadn't shared with anyone that I was working on this because the problem is when you talk about what you're doing, talking about it feels almost as good as actually doing the thing. So you've got to just do the thing. And I launched it and I got a a really fun email. I got a fun email from a company in San Francisco, small company in San Francisco. They were just 6 people and they more or less said that, hey, we're working with this peer of ours and the trade-offs that they're making and the economics just don't make sense, right? They were spending 6 figures on this and the fundamental problem they had was that the per user cost was, let's say, a dollar per user. I don't know what it actually was, right? But like, something that was untenable, and maybe they were charging them $20, maybe it was $5 per user. It was growing at a trend where the per-dollar cost was not something that they could earn a return on.

Very similar to your Readwise original kind of issue.

Exactly. And Readwise wasn't even the first customer there. They only became a customer actually a year later, um, which is very funny. But, and I think that if I hadn't gotten so busy, I would've just gone inside of Readwise and then shipped the thing right on TurboPuffer. But we got really busy really quickly. So I got this email, hey, we're spending six figures, the economics don't work for us. We're very interested in what we're building. And knowing that team now, I know that they would have sat around the dining table and said, why hasn't anyone just put all of this on S3? And when you query it, you put it into RAM, and it's as fast as anyone else. And you just paid a little penalty of at the very beginning, when you start interacting, it's a little slow, and then it gets fast, and then high the write latency, exactly the compromises we made. And for whatever reason, I had the conviction at the time to do two things. The first thing I did was that I texted who I thought was the best engineer that I worked with during my time at Shopify, Justin. And I asked Justin, Justin, do you want to come on? I can't pay you anything, but maybe we can work something out. This is what I'm working on. And then I sent in a bullet point list of the things that they would work on. And Justin said, let's take a walk. And we took a walk and Justin started working on it in the beginning, a little part-time, and 3 or 4 days in, they said, I want to be all in. And then Justin came on as co-founder. The other thing that happened around that time is that for whatever reason, I had the conviction with this small company in San Francisco that I'd never heard of. I'll reveal the name in a moment, that I should fly to San Francisco and meet these guys. And so I did. I flew to San Francisco.

Why was that?

They were impossible to track down. They were so busy that the only call that I could get with them was that at one point, one of the co-founders emailed me at 5:20 AM on a Thursday and said, hey, can you meet now? I'm free. After missing two meetings because of outages.

5:20 Ottawa time, by the way. Which time?

Eastern. Yeah. So for him, it would have been 2:00, 2:20 AM. Right, right. And I happen to be up. And so I was like, yeah, we can jump on a call. And we hopped on a call and explained all of it. And after that call, I felt okay, like we have this email exchange, like I need to go meet them because it's too hard to get them on a call. I just need to go see them in person. And so I flew to San Francisco and I didn't tell them that I was in town for them. I said, I'm in San Francisco on Monday. Let me drop by the No problem. I showed up on a, on a Monday afternoon, like basically the day that I got there. And when I showed up at the office, and this is how I remember the story, I need to actually confirm this with the team, how they remember, but this is how I remember it. But you can't trust your memories on these things. And the way I remember it is that I showed up and they were having problems with the Postgres database. Maybe it was down, there was some issue with it and they got it back up. And then I, doing my consulting, I'd spend a lot of time with Postgres. And so I just tried to teach them everything that I knew about running Postgres at scale, installing these different things, just helping them as much as I could. That I must have convinced them that maybe I knew a little bit about databases, because then when I told them what TurboPuffer, where it was going to take it, and all of that, they felt compelled enough that they started the migration.

Which is a massive deal, to be clear, like just in terms of trust, like it's not like trying out a normal product.

And of course, this company is Cursor, right?

We have tens of thousands of people who have followed the show. Are you one of those people? You want to be part of the group, you want to be a part of those tens of thousands of followers. Hit the follow button.

Little company, 4 founders. They were 6 people then, and I don't know how much revenue they were at, right? This is in, um, what is this? This is in October or November of 2023, and they needed 2 features, and Justin and I worked very hard to get them those 2 features. They started onboarding. They had a few billion vectors, which at the time was one of the largest workloads in the world, and they just did it over the course of a week. They migrated everything that they had, and their first bill was 95% smaller than their 6-figure bill that they had before, right? So they went from a 6-figure bill to a 4-figure bill. And of course, Cursor has grown a lot since, but it started a very special relationship. And in the beginning, I was so honored that they called me once sometimes and asked me to help them with infrastructure candidates, uh, to explain the opportunity at Cursor. I was probably one of the few people looking at our dashboards that could see their growth. and I sold them as hard as I could to the few candidates that came through.

I'm sure they're very, they're very happy now.

I don't know why they needed me for this, frankly, because they are so good at recruitment that I always felt honored when they asked and happy to do it. And they even asked me for infrastructure advice and how to lead teams and things like that. And I think they're probably much better at me than that at this point, but they were humble and they asked for advice where they could, and it forged a very special relationship at the time. One that I think continues to this day. We were in Slack and these are not their words, but we truly felt like we were part of the cursor team and we had our little function and we just tried to make sure, because my goal was like, one of the things that almost brought me to tears at some point was just like Swaleh, one of the co-founders, just saying sort of, you know, this is one of the only things that we just don't worry about. We just don't, we have not had to worry about it as we scaled up. And as someone who has been on call for a pager at Shopify that lost millions a minute if something goes down. And it's an absolute honor and privilege when one of the fastest growing businesses of all time by revenue tells you that you are one of the pieces of infrastructure that they've had to worry about the least. So we leaned in heavily, our only customer to begin with, and we built everything to make sure that it worked for them and they used every single line of code in the product. And we just worked very hard, Justin and I, to make sure that this was gonna work.

Just the two of you kind of at that point. And how, how quickly did that cursor take you to a million ARR?

Um, not just Cursor alone, right? We reached a million in ARR about a year after we launched, but we had more customers then than just, than just Cursor. We also had Notion and a bunch of other customers then.

So we'll get into how you landed those others, but just on Cursor, I'm curious going back to that Readwise thing, because with, with Readwise and these other examples, it was at the point where without TurboPuffer-like thing, they just couldn't launch the feature right now. Like you made it way cheaper for them. If you hadn't come along, could they even have scaled to what they've scaled to without either TurboPuffer or something like TurboPuffer to make the numbers work?

I think that if Turbopumper hadn't come along, they would have built it themselves in-house.

But they would have needed to have that sort of decrease in cost to make everything else make sense.

They would have built Turbopumper themselves, and I think they would have done a great job, but it would have meant that they would have had engineers that were solely dedicated on that, um, really good engineers who could be working on making Cursor better. It was a great partnership, right? And that I think was very mutually beneficial. But now they're many companies that have launched Turbo Puffer, right? Again, it was a discovery. It was not that it's like fundamentally the hardest thing to do, but of course you need a lot of good engineering for a long period of time to make sure the product is really good. But Cursor could have built it for sure. They have one of the best engineering teams in the world.

I mean, to be fair, that goes for many products. Like many products could in theory be built, but most people want to focus on their own product and especially the infrastructure, you know, kind of use others. So you and Justin, you're working on Cursor, just Cursor for how long?

You could sign up in December. So we, we had a bunch of other smaller companies, believe it or not. But the second larger deal in the thousands of dollars per month was TELUS. So for those of you who don't know, Pablo knows, but TELUS is one of the largest, if not the largest telco in Canada, and they are very much what you would consider an enterprise company. But someone inside of TELUS, this guy Justin, just saw TurboPuffer and similar, he'd had all these problems with vector search and search in general with these other providers. And he saw something in us. To this day, I give Justin so much credit because it's very rare that an enterprise gets that amount of leeway to bet on a startup.

At the database layer too, yeah, no less.

Yeah, we spun up a Canadian region for them within a few hours after them asking. And we're just like, look, we're going to work very hard for you. We know that you're putting yourself a little bit on the line here. Tell us what you're paying, and I think we can do a lot better than that. We just had a frank conversation with him, and they're very happy customers still.

What do they power, by the way? What's the use case for them?

They have a bunch of internal systems, but they have this platform that they work with their customers on integrating called Fuel IX, where they use TurboBuffer to search, right? And then they have a pluggable model layer, and then they sell that to their customers. And I think they use it for a bunch of things internally. They call them like internal copilots, and that they use for, for a bunch of different things.

But so it's AI use cases, but a lot of them internal. Okay.

Yeah. Inside of TELUS, they were our second larger customer.

In December, you said, like 2 months after?

No, this is in maybe February or March.

Okay. So I mean, 4 or 5 months.

Yeah. And then there was a bunch of random small startups. There's this company called, um, they, this Chrome extension called Merlin. There was a company called, uh, Fixie, and Dave wrote an article. He was like an ex-Shopify founder. Now they're doing voice agents. They built this, like, it was like, you know, when you watch like a sports roster where they trial all these databases against each other and that ended with TurboPuffer winning. And that article was really generous for us. It was a bunch of startups like that, like a dozen or so back then. I mean, Cursr was not that big. I didn't even know them back then, right in the beginning of 2024. They just launched in May of 2023, I believe it was. So they were not that big in October 2023.

How did TELUS, how did these other startups find you? Is it through X? Is it just like word of mouth? Or are you doing anything actively to get customers?

It was just X. That's just how they found it. We didn't have any other channel.

Did you keep posting though on X? Like you were doing that pretty consistently?

For sure. Yeah. I was like, ah, just improved performance by 20%, you know, things like that. Yeah, for sure.

We were—

I was very active on X. Justin was active on X, and the Turbo Puffer account was active, right? I think one of the things early on that also caught people's attention is that the website looks very similar now to what it looked like back then, but it had a very particular brand and flair. And me and Joachim, right, who is the designer that helped me come up with the name and the initial website, we spent so many hours on that, right? Joachim was basically the first outside person other than Justin and I on Turbo Puffer, just working on the design. And that design got a lot of attention. Back then. And now I think it's, I don't really know, like we didn't have any inspiration that was like, that was super close to it, but we just spent an enormous amount of time on it. And I think now I would probably look at it and be like, how do, why do we spend hundreds of hours on this?

That was going to be my question. Why did you? You can go both ways on websites. Like there's people who think it's huge, you know, matters so much. Other people could say easily, like, who cares? Like the name and the whatever. Like if your product's great, your product's great, you know, like you could, you take both sides of that debate and I'm sure find examples for both.

During my time at Shopify, I have looked at, if not hundreds of database websites, right? And again, what you care about with a database is just what does it do? What are the trade-offs? What are the guarantees? What are the limits? Who's using it? What does it cost? That's really all you care about. And I just wanted all those questions answered on the homepage. And I wanted an aesthetic that said, we care about design and UX, and you're going to have a great time. And we've spent some time to make sure that everything is organized in a way that makes sense, but not so much that you're not sure if you're buying a database or a sneaker. So that's really what I wanted with the website. And then we have this particular monospace type of aesthetic that we've iterated on in many, many iterations since then. But it just felt to me that it was so important to position as we care about your experience, we care about DevEx, we are hardcore database engineers, and we want to make it very clear for you to understand whether this is the right database for you. And if not, that's cool.

I mean, you literally had— I don't know if this is how it started, but you have like a— you can toggle around depending on your workload, see exactly how much it's going to cost you.

That's right. Yeah.

So Eli, tell us through X now. Now we're in 2024, right? This is— yeah, this is like February. When do you realize that Cursor is like, like, holy shit, you know, like these guys are growing at an insane, insane rate and they're going to be a, you know, huge customer for you guys.

I think we were just always wondering, like, when are they going to stop growing? Compounding is a very weird thing where it's very slow until it's incredibly fast. It was not clear to us, like, we were so focused on Notion because at the time, right, they sort of, they started POCing with us in the spring of 2024, and we were so focused on them.

They also came inbound, like, same thing through X?

No, outbound.

Okay.

Yeah.

Walk me through that. What were you doing on the outbound side?

So I mean, maybe we can— the other things that were happening around then at the beginning was that we were— a bunch of VCs got interested. Generally these were the VCs who were very technical, who also saw, but maybe also couldn't quite articulate that there's something here. And we ended up choosing this investor called Lockie. And the reason that I loved Lockie was for 3 reasons. The first reason was that I met him on the first trip that I went to San Francisco. I was actually introduced by the Readwise co-founder. He said, just go talk to this guy. Guy, and I went on a walk with him and he told me on the walk, Simon, if you're serious about this company, you need to get this lawyer because he's the guy that always leaves me basically crawling out of the room. And it's just like, it's such an odd recommendation for a VC, right? Now I'm like, that's incredible because yeah, if you are serious about your company, you need one of the best lawyers in the world. It's completely underrated. And this lawyer team that we work with at Wilson Sonsini, just, they take a small bit of equity, so they're completely aligned with you, common stock, and they just see more transactions. Actions than anyone else. And people underestimate how much this matters. But then I talked to one of the funds, and funds like, do you have a lawyer, right? Because a good sign whether a startup is serious. And I said, oh yeah, this guy, Rob Broderick and Damian Wise at Booz Allen City. And he said, oh no, no, no, no, no, don't talk to those guys. Here's my guy. That's how I knew that this is a good lawyer. That's how you know.

And they've been helpful. I'm curious on this, where have they had the most value for you?

I think the concept of having lawyers for different things before being a founder was a very nebulous concept to me. But a corporate lawyer is essentially the person who cares about stock option grants, like your cap table, anything that revolves basically exchanging equity for cash, right? Or for more equity or whatever, right? So these people become very important on partnerships. They become important for optimizing equity for employees. They are important for optimizing taxation for employees. They're very important for if you want to sell your company, then you want people who are your coach in your corner. And if you are raising money, then and you of course need someone who sees a lot of that. If you are thinking about your corporate structure and like, what are you— should you have in the US versus Canada, you need really good corporate lawyers. We have corporate lawyers in Canada and in the US, but you want a corporate lawyer who sees a lot of activity and ideally a corporate lawyer who never works with VCs, but only ever works with founders. And ideally, your corporate lawyer is available on WhatsApp within 5 minutes at all times for any questions that you might have. And that's what these lawyers are, and that they're incentive aligned with you on anything that might happen, whether you're IPOing, whether you're selling your company, or whether you're doing investment because they have the same stock that's diluted exactly the same as you and all of the employees. I talked to other lawyers at the beginning, and when they told me they're hourly, you can just see their daddy issues come out through their eyes. That is not what you want.

So this VC makes that recommendation. When do you go back to him and decide, do you want to raise?

Yeah, so the other thing he did was that he was just helpful immediately, right? I got to taste before I bought, and he was just incredibly well and responsive. And then there's another thing, which was that he didn't know anything about databases and he was honest on it and he just said, yeah, I said, have you ever invested in a database company? No, I've never invested in a database company. And he just said it with such conviction that I was like, oh, maybe that doesn't matter. Maybe it's actually a good thing because Justin and I should know about databases. The people that we hire should know about databases. It doesn't really matter if the investor knows about databases other than they might find us earlier. And he was just incredibly well connected. So yeah, those were the reasons that went into choosing him.

When you went and you decided to raise some small pre-seed that you won't say exactly how much, like, why did you need that money for? I assume you want to hire, like, How many people do you want to hire with that?

Yeah, so in the beginning, we were like, everything just ran on Justin and I's credit cards. So we were trying as hard as we could to not lose money. I mean, we didn't know how to raise money. Like, we didn't know how all this stuff worked. So we're just trying to not lose money. So we were optimizing like the cursor workload like crazy to just not lose money. And then when I was in high school, I went to this competition called IOI. Have you ever heard of this?

No, I haven't.

So IOI is an international competition where like 90 to 100 companies send their 4 students. It's basically like Computer Science Olympics in high school. Yeah, the problem is that no one's heard about it, but it contributes more to GDP than the real Olympics. And let me give you an example. The founder of Cognition is a gold medalist. The founder of this company called Modal Labs is a gold medalist. A lot of Silicon Valley and the founders of this generation are people who did really well in these competitions in high school. I did not win a medal, but Bojan did. And Bojan and I met in 2012 in Italy where we both went to the competition. He was on the team. The North Macedonian team, I was on the Danish team, and we got along. The Macedonian team called him God, because he was very, he's very good, very good at the competition. And he was a silver medalist. And we got reintroduced by a friend. And he was the first employee at Turbo Puffer. But we didn't have the money to pay him. And we'd already paid enough money out of our own credit cards, Justin and I, we need to have someone on board. And that was the reason why we raised.

So the other thing I want to talk about the outbound piece. So you said you actually, you got Notion through outbound, what outbound motion did you start running?

We just knew that Notion was such a good fit for our architecture. And we got introduced, I forget who exactly introduced us the first time, but we got reintroduced many times. Like most people just ignored us to begin with. And then eventually someone took a call and then nothing happened for like 5 months. So it's hard to attribute because Lockie is also an investor in Notion. So, you know, I don't know what exactly he worked, but at some point they just reached out and said, hey, geoeconomics are holding us back a little bit on product. And so we want to try on some other suggestions. Later I learned I was just at dinner last week with one of the people at Notion that worked with TurboPuffer, and apparently they said that him and a guy on his team that we now work closely with had sketched out what they thought that the right architecture for Notion was for search. And they had basically come up with TurboPuffer and then sort of matched that. And so I don't know exactly how we ended up on this call, but we had a call and there's, you know, 6 people in the room and it was just me on the call on our side. And they just berated us with questions about can you do this and can you do that and so on. Now I understand that they had already had the product vision that they're executing on today. Notion AI, you know, it's like the cursor for pros, and they needed the right search engine to power all of that with the right economics that they could earn a return on. And so they asked for a region in Oregon, and we spun up a region in Oregon within a day, and they just started testing. And Justin just pulled tricks doing this POC that— I mean, Justin and I worked together for almost 8 years at Shopify. And Justin was always one of the best engineers I'd ever worked with. What I saw Justin pull to make sure that Notion was gonna work was beyond anything I'd ever seen before. I remember distinctly, um, Morgan, uh, our second employee had joined at the time and we were doing a little team offsite in Quebec, Canada. And we came back from dinner and Notion had finally ingested their first workload and it was really slow. It was like 500 milliseconds for a query and we wanted less than 100. Justin found 300 milliseconds or something like that in the span of 3 hours that night. Day, just kept banging out PRs from looking at the workload. And there was another time where Notion asked, can we do this, Cory? And Justin just told them in Slack, yes, we can. We just need to expose it in the API and then worked like I've never seen before for the next 24 hours to get it in their hands. This is how Justin worked on this deal. When people talk about forcing something or willing something into existence, this is one of the first things that I think stuff. And it was weeks and weeks of that, POCing different angles, right? And then I was working on the pricing because we didn't even know how to charge for this at the time. Bojan was working on a complete rewrite of TurboBuffer on the side, and Morgan was working on the full-text search. We were just, we were so focused. And finally, towards the summer, we got the feeling that maybe they were going to commit. We'd felt a little like it was slow, and they had some commits with the other vendor that they were riding out. And on the 25th of July, it was getting really close. My wife was 7 or 8 days overdue with our first child. I was working so deeply on the billing system because otherwise we were gonna lose money on Notion because we needed to put this new billing in place. Finally, they signed on the 25th of July in 2024. I invited Justin and our virtual CFO, Mike, who's now on full-time, who were the only ones in Ottawa. And then we had Morgan and Bojan in on a video call and we're all just like, you know, awkwardly on like a call. My wife was there as well. Very pregnant. And Justin says to my wife, "Wouldn't it be funny if your daughter was born today?" And this is the biggest jinx of all time because this was, you know, late afternoon. They all leave. Justin and I hang out for an hour. Like, I cried a little bit that day. It was just like we'd worked so hard on this, and Justin in particular. And Justin left, and then my wife came down and said, "I think something is happening." No way. And 3 or 4 hours later, our daughter was born.

That's amazing.

First daughter? First daughter. That was a wild, wild day.

It would be hard to craft a better story than that. And did you know at the outset, like when you're sizing these deals, like Notion, obviously it's gonna be a massive customer, but do you know how much revenue that's going to generate exactly? Or is it an estimation? 'Cause you're not really selling like pre-seed or whatever, obviously. Like, so how much do you know about how big it'll actually be?

Yeah, I mean, at the time I was spending probably a third of my time doing engineering. And a third of my time on pricing, you know, an updated pricing that was going to work for them and making sure that we could stay afloat. And then the other third of the time selling. What we did, right, is you sent them a form saying this is like the approximate size of your workload, and then we generate a quote for them. Notion came back and said, are you going to lose money on this? Now I know that Notion was legitimately concerned that we were going to go out of business because it was so much more cost effective than what they were doing. So yeah, we knew exactly what we were going to earn on it, right? And we knew like the rough margin profile. You never really know until the rubber hits the road. Yeah, we knew exactly. And at this time, like that workload was like, was bigger than Cursor.

This is where I'm curious to dive into this because at this point, like, so you've got TELUS, you've got Cursor, which you don't know how big it is, but it is, you know, it's a growing customer. You got a bunch of little customers. Now you sign Notion. I would argue 99% of founders in that situation, especially these days, they're going around, they're raising like a $30 million Series A. I mean, you've got all the, like everything you need to have to go out and do that. You, you don't do that and you haven't done that. Why not?

There's 6 reasons to fundraise. Reason number 1 was the reason that we fundraised in January of 2024, exactly how much money that we needed to prove to ourselves that we could find product market fit. If Justin and I could not find product market fit by the end of 2024, we were just gonna close up shop. Justin and I were very clear about this. We even told the investors who were invited for that round in 2023, that we were gonna close up shop by the end of 2024 if we'd not found PMF. That was terrifying to everyone but Locky. We raised enough money that we knew that we could fund R&D for 2024. The second reason to raise is to fund growth. It's to fund marketing and other ways that you think that you could grow the business once you have a proven way to turn dollars into more dollars and mindshare. The third reason to raise, and this is probably the most popular reason, and it's kind of what you're getting at here. It's for ego. It's also known as raising because you can, or because there is momentum, or because you got preempted. These are not first principle reasons to raise capital, which has real downsides, right? It dilutes employees, it ups the strike price. It has all kinds of ramifications that you now have to live with that might put your business at risk. The fourth reason to raise is to fund liquidity for the employees, right, that have believed and that have been here for a long time. The fifth reason to raise is for trust and/or publicity. So you raise basically to build credibility for your business. That would not have been an illegitimate reason for a database company to raise. And the sixth reason to raise is for strategic partnerships. This can either be, you know, a VC that you really think is going to make a dramatic difference for you, either because they've proven it or they have connections or whatever that you need. But it could also be institutional investors, right? Where you, you know that if you get this customer, they get on your cap table, it's incredibly important for your business, right? Like if you're building software for movie studios, you might do that for Disney, right? Or something like that. Those are the 6 reasons to raise. And if you're raising for any other reason than those 6 reasons, you are raising at reason number 3, which is to fund ego.

There's a lot of reasons there that I could argue, yeah, like why not? You know, you get more credibility, you'll get more go-to-market, you know, whatever. Like all those reasons could in theory apply to you.

I mean, I could go through every single one of them and we can see why they didn't apply at the time.

Or even today. I'm curious on maybe not every single one, but the ones that would be most obvious out of what you said would probably credibility, which would in theory help you, you know, hire and things like that even faster than you already are. And, you know, go-to-market. I mean, those would be the two that jump out at me as like, yeah, you could raise for that at, you know, small dilution. Why not?

On the growth piece, we don't have a growth motion where we can put dollars in and get more dollars and mindshare out the other way. That may be true over time, right? That's the team, that's the team I'm working on building right now. And then we might raise for that reason alone.

Because it's mainly inbound still that your growth is coming from?

Yeah, mainly inbound is the majority of our growth. And then on the, on the trust piece, we have found that that has not been a limiting factor. Now we got very lucky, right? We worked so closely with Cursor and they grew from being maybe not a particularly known logo to an amazing brand to be associated with.

Huge.

In a very short order. So we got incredibly lucky. But there's lots of downsides to raise. I don't have a neat list for you. When we hit one of those triggers, that's when we would consider raising. And we would look at the people that we've worked with, that we've liked working with, when the time comes.

How many people are you in the team?

The team now is, um, like as of October 9th, we are 17 people.

17 people. And so I assume obviously you're profitable.

If you want to build a database company, it is a longstanding endeavor. And so we are profitable for that reason because we want to build an enduring, large independent business.

Maybe my last question on fundraising, like, you don't need 20 or 30 more people on the product and engineering side and then be unprofitable but not have to worry about longevity, for example. That's just not a need for you.

If you find me 30, like, database engineers of the caliber of the ones that I already have, then yeah, like, then we need to raise. I'm constrained in finding those people and getting them to join Turbopuffer more so than I am on capital. And by the way, these engineers want equity. That's what they're most interested in.

You purposely, I mean, you started off the call, this I think was before we started recording, but you're like, you know, Otto was a great place to hide, right? You're kind of purposefully somewhat stealth, like you're buying into the stealth piece. And even on revenue sharing, you don't want to share revenue. Like, what's the strategy behind that? What's the thinking behind that?

We want to build the best database. How does disclosing various things, how does, like, how do these things help me be a big database company? If something is difficult to price and it has externalities that are difficult for me to price, I just don't do it, right? What is very clear is building a world-class database team and working on the database and connecting our databases and many companies as, as possible that need it. That's very clear ROI, and that's why I spent my time doing that. Of course, we need everyone to know what TurboPuffer is, and that time is coming.

Got you.

Cool.

Well, listen, let me stop it there. Let me ask the 3 questions that we always end on. The first one is, when did you personally feel like you'd found true product-market fit?

It was that day that Notion signed because Cursor was still not that big. They were an incredible customer and we loved working with them. But when Notion signed, that's when I realized that this could be a really big business.

Second question is, was there a point— I mean, things have gone really fast, so I feel like the answer is no, but was there ever a point where you thought things just wouldn't actually work out?

About 3 months after we had gone to market, our biggest peer basically launched Turbo Puffer. And that's when Justin and I had to look each other in the eye in January of 2024 and decide whether we were going to go up against this or if we're just going to go and do something else, right? And we spent a lot of time debating that and ultimately decided that we're going to do our darn hardest here because like we're excited. And we have a great set of customers, but if we don't find it by the end of the year, we'll close up. We want to build a big database business, and if we can't do that, then there's no point.

By the way, how do you go up against that? Is that a product thing? Like, you're just— your database is just better, like, or that's what you're trying to do, just be better than them, be cheaper than them? Like, is there big ways to differentiate?

We thought that it was going to be incredibly difficult, right? When an engineer sees something, I think a good engineer will always assume that it's the best version of what it is. And we knew what we were building and we knew how to get to the best version of it. And we assumed that that's also what they had. But for some reason, customers kept coming inbound, right? Even though we were a small company. So like why that happened, like, I don't know, right? Maybe it wasn't as suited for some of the workloads that they had optimized it for in R&D. Maybe they rushed it. I don't know. Right. But I know that customers were testing us against it and we were performing better. and that's why they chose us.

And then last question, what would be like your, your number one piece of advice for other early-stage founders?

You are going to encounter a lot of very articulate advice from VCs and other advisors, but if you truly have the founder, like, company fit or whatever, founder-market fit, you will really have to trust your intuition. That is a lesson that I've, like, had to retell myself many, many times because I'm— it's very easy to be persuaded by someone who's very articulate and spent a lot of time in the market, you got to just lean into the things that make like your business and what you think is right weird, right? And that's what makes your business unique. Otherwise, it becomes a cookie-cutter thing, comes out of Silicon Valley exactly as the VCs think that the playbook is. But every playbook that's worked has always been a little weird.

Simon, thanks so much for jumping on the show, dude. It's been awesome.

Thank you, Pablo.

Wow, what an episode. You're probably in awe. You're in absolute shock. You're like, like, that helped me so much. So guess what? Now it's your turn to help someone else. Share the episode in the WhatsApp group you have with founders, share it on that Slack channel, send it to your founder friends, and help them out. Trust me, they will love you for it.