Turbopuffer CEO Simon Eskildsen: companies are building their own Google-scale search indexes on top of serverless vector DB

robotics, evals, and expert human data. Labelbox is the data factory behind the world's leading AI teams. And we have our next guest in the reream waiting.

Is it time to puff?

It's time to puff. We got Simon from Turop Puffer.

Waiting room. Let's bring him into the TVP Ultra.

Simon, how you doing? Good to see you.

Good. Good to see you guys. Hey, John. Hey, Jordy. It's time to puff.

It's time to puff. What's new in your world? Were you thinking about doing an over-the-top attack ad on a competitor or are you locked in focused for this Super Bowl? What's the strategy? How is the business growing?

I mean, you were so generous to give us a front and send logo uh to the Super Bowl.

Um as as a new entrant. Yeah. As a new entrant to the new world, I can't

I don't really understand American football.

Um

Okay. But it seems right up your alley because when I think of it from first principles, it seems like a sport that has been designed to show ads. Yes.

Which I think that is a great American invention. So I have to respect that.

It's beautiful.

It's very beautiful. Yeah.

Okay. Uh how has the uh the agentic coding boom been processed to you? What what's the effect? Are you seeing any slowdown? There's been hype waves and selloffs and all sorts of turmoil. Uh has the recent uh sort of reaceleration uh taken you to surprise? Did it hit you like a flashbang?

I whatever is going on right now is hitting us like something.

That's good. What is going on? Break it down for us.

You've been waiting to use this effect all day.

Yes, all day. And the chat demanded it. Anyway,

Bobby says billions billions must puff.

Yes,

we need a puff one. We need a puff one. Like a puffer fish that sort of just like

you know when you want me to get off camera, you get a puffer fish that counting down.

Well, we have a smoke grenade. We have a smoke grenade that kind of issues puffs of smoke. But maybe they need logos all over them. That would be good.

I mean, look, the world the world's puffing. Uh, I think

you asked about coding. I think coding was the first vertical that really started puffing really hard. And I mean in general the way that we look at the world is that we can certainly take all of the knowledge in the world and we can distill it into a few terabytes of weights and those terabytes of weights are extremely useful at reasoning over data.

But in some way they have to reason with the data and uh to reason with the data you need a search engine right that has all the canonical data and is that tool to search through the data.

Um and I mean that's what cursor and cognition and other coding tools are doing with turbopuffer. um is to search through the code. And so I think we're seeing that in other verticals now. We see it in legal, right? And we see

every vertical sort of starts by just pushing a bunch of things into context. But then really what they want is to allow the agent to search by itself. And that's where Turboer comes in as that as the search engine. It can index pabytes and pabytes of data to allow these agents to search.

Yeah, we were I mean we were talking about ads. uh is do you have a burgeoning ads business or not not ads in turbopuffer but powering ad tech systems because I can imagine a lot of new platforms they have a lot of content they have a lot of ma ads they need to match those and so uh your system should be valuable there is is that a stretch or is that a good application

I think most of the ad businesses already have extremely sophisticated systems for this I'm thinking new generally it's yeah

yeah no we we haven't seen that vertical really come alive for us yet?

Yeah, I imagine I imagine uh it becomes important uh uh and and uh what about other uh sort of uh applications or sectors that are growing in terms of uh this technology adopting this technology puffing broadly

something that we're seeing that's puffing pretty hard right now is that people want to basically build their own Google. They want to take the entire web and make it searchable and allow the agents to reason over either literally the entire public internet or data sizes that are that um internal data that's of that kind of volume, right? Hundreds of billions of documents. Um and recently we we we launched some product to support that where you can basically now build your own internet index on top of turbuffer for an extremely reasonable cost. So what we showed is that we can take 100 billion documents and make them available to you with a P50. Um so that's the immediate latency of less than 50.

So this is pretty remarkable for an offthe-shelf SAS product and something that we're seeing um that some of the most sophisticated customers in the world uh need and I don't think there's any other easy way to do it than to puff. Yeah, we talked a little bit last time on uh on just compute bottlenecks, AI bottlenecks, anything that you're that's keeping you up at night in terms of uh data center buildouts, semiconductor buildouts, even if it's not directly related to your business. Is there anything that you see that could potentially put a damper on the growth of the booming AI industry? I mean, I think everyone is everyone is is struggling to get compute once you when you get enough scale. Um, and I mean that that's it's it's good news in CPU land. Um, I'm I'm not contending for the GPUs yet. Um, but we are starting to see I mean you have to do that at any scale, right? You have to start requesting from the cloud providers, hey, I want hundreds and hundreds of this machine in this particular region for these customers. Um, so I think that's something that everyone's facing. Obviously, we're watching the DRM prices and things like that like everyone else. Um, but uh hopefully hopefully that gets that comes down.

Are there any any kind of like nar you have an interesting view into the usage of pretty much every popular AI product? Obviously you can't talk about uh individual uh the the sort of usage of any individual product but is there any sort of narrative violations that you're seeing broadly every single day depending on new products getting announced people are saying you know it's over for this company it's over for that company or I just turned from this product but uh overall what are you seeing

I think I mean what we see is that the companies that are building the most exciting product are trying to operationalize very very large amounts of data and that's difficult difficult to copy and so I mean we have survivorship bias in the businesses that we see but they come to us because they want to connect more data to AI um than anyone else. I I would say yeah if you're a thin layer and you're not trying to vacuum up a lot of data then it's it starts to get really difficult. So I I mean I I do see as with how powerful the models are now, you you have to be investing in some pretty pretty deep tech um to to make it um Yeah.

Yeah. Are you doing anything on this on the on this uh on the front of like making Turbo Puffer intelligible to agents? So, uh, if someone goes to a vibe coding app or a CLI tool and they just say, "Solve this problem for me." The agent pulls Turbo Puffer off the shelf.

I mean, we we we try as much as we can to make the docs very easy to read. Um, I think it's not completely clear how the LLMs make the decision of what the best database is, but I think that we're trying to do what we can to make sure that the LLMs like like Puffin as well.

Yeah.

Um, I I we have made API decisions around what it would be easier for the for the LLMs to guess.

Yeah. Yeah. I heard an interesting anecdote about uh someone building a CLI tool. They wanted to make it easy for an L an LLM to grab or an agent to grab. And so basically they just uh rewrote the CLI to include every possible hallucination or mistake that could be made. And people often do this with with URLs where they will say, well, you know, some people might type the wrong URL sometimes, but LLM hallucinate in a slightly different way. they don't necessarily fat finger a single key, but they might, you know, drop the wrong token or use the wrong phrase or use the wrong term. And so they rewrote their CLI to include like every possible permutation that that that an LLM could could think about in terms of like update, edit, quit, exit, uh you know, all the different keywords. They just created hooks and functionality for all of them. So even if the LLM shows up and is like kind of clumsy, it can still use it and so it still likes it. But I don't know. I don't know where all that goes on the on the commercial side. This is more for like open source.

Yeah, I think I think maybe it's a little overblown how much you have to design for this. I think it's the same with humans. We fat finger stuff and the LLMs are a lot more resilient now than they were even just 6 months ago. So to me, it's this just always good API design. If you're designing a good API or a good CLI or good interface, you should sort of be able to predict like, oh, if I did this, it's probably going to work this way. And I think the LLM treats it exactly the same way. So I I I I would maybe take back that we're doing anything super intentional for the LLMs because I think the same design that's good for humans is good for LLMs.

Yeah. Yeah, that makes a lot of sense.

Uh what if anything has changed since the funding round?

Yeah,

not much has changed. Uh we got logo rights to ramp. So that's exciting.

Congratulations. Second ramp mention.

They were like finally you raised a bit more venture capital. We'll let you use our logo.

There you go.

I mean, you know, it's uh it's in and it's it's uh yeah, I think we're we're really really happy to work with that team and I think there's lots more logos uh waiting beneath the surface to be on Earth. Um so, the Murderers row will will continue to to evolve um as we earn the trust of more and more of the lovely companies that we that we work with. What was the what was early on when companies were discovering Turppuffer and realizing, hey, this is really good and this is going to be sort of key to enabling all these different features and functionality that we want. Was it actually a challenge getting logo rights because they didn't want their competitors to be super aware of it, right? You want like, hey, if I can have an even an extra month before

another uh company that I might have some overlap with, I'll take it.

Yeah. I mean, we have some verticals in um like there's a headfront that doesn't really want us to talk about uh um because they they see it as like a competitive advantage, right? Um and so I mean I we like that tension. Um, for sure. I think honestly in the beginning it was just um I'm sure that some of the companies that bet on us early were like, "Okay, this is like uh three or four people in Canada and we're betting our whole business and search on on them being able to make it." Um, and I've certainly heard from one of our early customers that that kept them awake at night.

Now the company is 10 times that size, right? and of a of a different caliber and we owe them everything to have bet on us that early on. Um, but it was certainly a challenge in the beginning, not just because they saw it as a secret sauce, but also because they were afraid of what their customers would think, right? Here's this like little like dinky company in the woods of Canada. Like what do they know about ranking search engines?

Uh, and I think they were concerned what their customers thought. And we've just tried to do everything that we can by keeping keeping our uptime as good as possible and walking everything carefully and building good software to earn that trust. But I think every startup that gets logos, I think the founders have to really fight for it. Um, and they have to build the relationships to make sure that people trust they're going to use the logo in good hands and they're not going to plaster it all over contexts um that they wouldn't feel okay with.

How are you guys approaching vendor selection today? when are you running the calculus of should we build this ourselves or pull something off the rack?

We were we were just discussing this recently where we've started our our data stack now is just using like cloud code or cursor agent in a repo and just giving as much information as possible what's going on. we've ditched we're we're starting to more and more ditch just the tools because it's so easy to compose the charts and things like that with agents. So the the shallow SAS software I think it's it's it's dire there because it's just so easy to replace it now. Um but I think there will also be a bit of a hangover from that right there is real value here in like permissions and enterprise like all these enterprise features. Um, so it will probably go very much in the in the direction of everyone's going to build their own little shallow thing and then some of them are going to realize that this is a this is a bit of a nightmare to maintain and then pay for it again. Um, could be dire for a little bit, but I think that that I don't think it's going to swing. I I I I still think there's a lot of value in that.

Yeah. What's the uh what's the worst like logo crime that you've ever seen someone do? Is that just uh go in the customer database, search has any employee from a hyperscaler uh signed up for a trial account and then slap the logo on the landing page. Uh what what would you tell a young founder not to do in terms of throwing a logo on a landing page?

I think we we like I think logo crimes is actually great diction. I think we see this a lot where you see like just you go on a page and it's like Microsoft, right? And it's all these it's like okay like

we were talking about this internally because so for example um there's a there's one small team inside of Verscell that uses Turbo Puffer

and I think a lot of companies might then just put the Versell logo on the website and say Versel's puffing.

Yeah.

But I don't think that's fair right. It's an accilliary use case. It's one person who's using it for go to market and doing a phenomenal job over there. And so what we're trying to do is like okay that's that that logo might not make it to the front page because that would not be an honest association. And then on the use use case page we're going to be having a tag that says core and accelerary. And then under accelerary right it will say okay it's this particular team at this enterprise that's using it. Um just to be completely honest about it. And I think that's and then I think just sending a note to the teams that you work with of like hey is this a cool uh place to use your logo. You have to be very respectful about it.

Yeah. Ask for you.

Yeah, you can't speedrun it. It takes time.

The chat the chat's saying Verscel's puffing.

Breaking news. Put up a TBPN trading card. We got the scoop here.

A case study.

Yeah,

case study will come out. But as I said, it's an accillerary use case and and we love the trust, but it's not like V0 is puffing or something like that, right?

That makes a ton of sense. Well, thank you so much for hopping by. We'll talk to you soon and have a great day.

While we were live, Google reported earnings. Revenue grew 18%, search 17%, Gemini has 750 monthly users, and the stock is absolutely nuking. So,

say just don't don't look at your portfolio. Stay in Slack.

It's just barely a$4 trillion company, though. That's so rough. When are you coming when are you coming to California next? After

hours or Southern California? I'm going to California next week.

All right.

But when are you when are you coming up to Canada? I keep like every time we talk about this and you guys are like, "Oh, it's so cold. It's the whole country."

I just feel like the whole country is unAmerican from my perspective. It's just it's just

Have you met American Simon? I'm right here.

Okay.

I'm ready for you.

Okay. You got me. You got me. Well,

we would love We would love to go fishing.

Uh

we need to do a weekend weekend trip. It's too hard to bring the show.

Honestly, yeah. Family trip.

Yeah, that'd be great. Let's do it.

Realistically, we bring great bring in everybody.

Awesome.

You You dumped them in the lake and you have a great time.

Sounds like risky with some one-year-olds, but we'll give it a try.

Anyway, thank you so much for stopping by. See you, Simon. We'll talk to you soon.

Great update.

Goodbye.

We talk about Phantom Cash. Fund your wallet without exchanges or middlemen and spend with a Phantom card. Uh chat says we need red suits.

Red suits. Yeah, we were talking about that today because we've been wearing the white suits a lot and now uh it's time for the red suits. We're going to have to call the tailor.

Capital says Alphabet sees 2026 capex 175 to 185 billion versus 115 billion expected.

We go there. Uh, for the past decade, Mag 7 were free cash flow monsters who