Exa raises $85M Series B led by Peter Fenton to build search infrastructure for AI applications

Our next guest is in the reream waiting room already. We'll bring in Will from EXA with another massive fundraising round. Uh you can do the honors on this one, Jordy. Uh let's bring Will in from the reream waiting room into the TVP Ultradome. Welcome to the stream, Will. How you doing? How's your day? Hello.

How you doing? What you got? What are where where are you sitting right now? You're on a train. I'm in a phone booth within our Okay, cool. We need some of these. That looked like it was like some type of like tractor or something.

This is the uh this is the the physical instantiation of the lock the great lock in of of September to December. You get locked in booth. Fantastic. Uh you've clearly been locked in. What's the news? Yeah, big big day today. We just raised a series just announced our series B uh uh $85. Sorry, couldn't hear you.

Couldn't hear you. Oh yeah, Peter Fenton. That's exciting. What's it like working with him? Yeah, I mean, he's a legend. He uh he's super experienced. He's taken seven companies to IPO. He really he just honestly his vision for EXA was just like matched or exceeded ours.

And so that was really exciting to work with a VC who just like saw what we saw and saw how big this company could get. Yeah. Yeah. Uh so he's joined the board. Had you worked with him before or is this new? No. No, I met him like like a month ago, month and a half. Month ago. Nice. Yeah, fast friends and partners.

Uh, so yeah, take us through the the story of the business. I mean, only a couple years old, well on your way to an overnight success. I think if we check in with you any decade, it will be pretty clear that you're an overnight success.

Uh, but uh, take us through the history of the company, uh, where you are now, kind of what what what's been the key growth driver. Yeah, so we actually started in 2021, so summer 2021. This is way before CHBT.

And the idea was like, you know, at the time GB3 had recently come out and it was like this magical thing that could understand like a whole paragraph of text. And then at the same time there was Google which felt like it hadn't changed in a decade.

So the idea was what if we could build a search engine that was as smart as GB3, something like fully understand understood the web and fully understood your query and get you like way better results than Google. And that's how we started. Uh so not necessarily a search engine for AI.

It was more like a search engine for nerds because we were building it for ourselves. Uh and then when Chhat came out a couple years later, uh it turned out that like the search engine we were building for ourselves for nerds was actually perfect for AIS because AIS and nerds are very similar.

Uh but yeah, I mean it basically we had to go build a search engine from scratch for years. Uh which is really hard. So we did, you know, we bought a GPU clust a million dollar GPU cluster after YC.

We trained you know a ton of different models, tried a ton of different transform architectures until we had like a new type of search engine. So, so the uh so the GPUs are trained to actually So you're not doing you're not doing crawling on GPUs, right? That's like traditional CPU workloads, right? That's right.

So we have our CPU batch processing and then we train models on our GPUs. Okay. And uh and yeah, what there's been a ton of news today in the search engine world with the Google news like what what's been your interpretation of it? Oh yeah.

I mean there's so much like what's crazy like AI is very exciting but then search for AI is like now the most exciting uh space is really cool um I think everyone's realizing that uh like sure like we want smarter and smarter LLMs but really we want more knowledgeable LM like we want to connect our LM to like the best data in the world and search is a like basically the most important tool to connect your LLM to uh and so yeah I mean that's pretty exciting the there's a Google news like for example like uh yeah I mean uh they're sharing their data has one thing about the search space everything's very counterintuitive so for example like if everyone now has access to Google's data, people are less incentivized to build new search engines, which is interesting.

Um, so there's there's all sorts of counterintuitive like uh things in the search space. Yeah. So what so yeah, how how does that shape like your strategy? Because um it feels like right now Google's not only like the the they they have the search monopoly, they were kind of like, you know, found guilty of that.

Uh the remedies are not uh exactly putting them on the back foot. The stock's up 8% today. And it feels like they're partnering even they're going even deeper with Apple potentially.

The the Bloomberg reporting that just came out uh is that there are rumors that uh that Apple will be partnering with Gemini to test out if Siri can be powered by Gemini which is obviously built on top of uh the Google crawler and Google's search index.

And so is the future for you look like a really long tale of you know tons of businesses that need to use some s some form of like best-in-class search and they can't go to Google for uh for a particular reason at like an API level because Google doesn't want to sell that as a product and you're there to sell that.

Yeah, that's exactly right. Like a big difference between us and Google is that Google is not trying to be search infrastructure. They're trying to be a platform like they want people to go to Google. uh and uh but we are uniquely trying to be search infrastructure.

So we want a long tale of companies like every company to be able to have the highest quality search inside their applications and so that makes us very unique like when we said we were a search engine for AI like a couple years ago uh no one really got what that means but I think that world has really played out where now we have you know thousands of companies using Exa uh powering all sorts of applications uh whether internal or external uh with high quality search and and like you said like Google doesn't like is not doing that like they're not going to have an API because well there's a lot of searches that are valuable in the context of specific applications that are not even that valuable to Google, right?

Like yes, like like when you're trying to retrieve information that you can then take action on, that's not necessarily something that Google is like act, you know, able to run a bunch of great ads against. Is that is that right? Yeah, that's right.

And the business model determines basis model determines uh like how the search feels or how the algorithm works. And so for example, like Google's really bad at recruiting.

Like you can't use Google to find like give me all the engineers in San Francisco who have like a PhD in machine learning and get a list of those things. Like why why isn't Google good at that? Well, it it doesn't make them more ad revenue.

Whereas we're very good at that because uh all sorts of customers AI applications want that high quality knowledge. And so like you're the the end user you're selling to like really determines the search algorith which also makes us very unique. Yeah. What is the value of actually training a model?

I feel like there's enough of a business. I don't know if now's the best time for this for that particular business, but but like just being the best web scraper index. You have all the every single web page perfectly indexed in a database that people can query. That feels like a valuable product in and of itself.

Um, and then someone else brings their foundation model to bear on top of your search index. And you're a partner to the big labs that train the big models. Why are you training your own model? Yeah, I mean well the easy part is actually gathering all the data.

I mean it's hard that you need like crazy scale infrastructure but it's it's an engineering problem once you gather all the you know trillions of pages. The hard part is how do you filter those trillion pages to the right 10 or 100 pages in real time and that's where all the secret sauce is.

And so like we train our own we train embedding models because we see that as the like the the way to get the highest quality search. Like how do you get like perfect search over the world's information?

you need to train all these like crazy neural uh models which is in contrast to the old world which is like Google for example mostly uses like keyword based method obviously they use a mix of things but like we we're very bitter lesson at exa and so we believe if you have the right uh you know training data set and like uh feedback signal you can get an extremely good search engine for the thing you're training for.

Yeah, I mean when you say you're bitter lessons pilled, I feel like that's rough because like who's more bitter lesson than Google? Like they have the most data, the most compute, deep mind, all this other stuff.

Like like what is the can walk me through like the counterpositioning against against like Google's like business here? I I the whole like Google will do it is like a complete trite over I don't believe in it.

I I'm bullish on you, you know, but like there is something weird about like what you just said in that I would expect Google to be able to move off of keyword pretty quickly. Uh at least from a technology perspective. Yeah. Well, there are two things.

So, one, it actually is kind of hard for uh for a giant shift to switch to a different algorithm. It's like have that works really well. It's it's very reliable and there are a lot of benefits to a keyword search algorithm. Uh and so it's it's it's hard.

You know, you have a lot of people at the organization who are like been there for decades who like have a certain thing. uh you have all sorts of like uh advertisements that are connected to the keywords actually. But and I think the more important thing is that it comes back to the the way the revenue model.

So they are serving humans with ads and so to ser like and by the way Google is fantastic at that like Google has built amazing consumer search engine for humans to make money from ads and uh and optimize for what humans click on but that's not what AI applications want.

So a critical part point of Exa is that like we are optimizing for a different thing than Google. Google is optimizing for humans and clicks and ads. we are optimizing for AIs and AI applications and like all these complex searches that this new AI world makes. Yeah.

So that that explains like even if we both have a ton of compute uh we'll build very different search algorithms. Yeah, it does seem like with the uh what what is it AI search overviews like those have at least anecdotally been hallucinating like crazy.

So um they definitely have some gaps to bridge between like some phenomenal stuff going on in Gemini and some frontier level uh like reasoning models and then some very reliable Google search results and then the the AI search overviews still they haven't fully solved the hallucination problem and some grounding in truth there obviously a problem they're working you guys compete with with you.

com we just had on Richard on is is that is that or is it is it it's not a lot not a lot of if I see him it's on site Well, it was funny if you both announcing fundraisers today. I don't know if that would saw that. Yeah. Yeah. I I mean they they mentioned that they do search and they also do AI agents.

Uh so uh like we are definitely like in a similar space. I would say like the the space is very hot like the market everyone's realizing the market is massive. Sure. Uh and so like you know I welcome other players there. Uh I think you know we've been doing this for many years.

You welcome them but you encourage them to take the full Memorial or Labor Day weekend off. Yeah, you guys should It's a hot space. It's going to be a long road. Just take the full 3-day vacation. We grind. Yeah. Uh, okay. Sorry. Last question for me. Um, you bought a GPU cluster.

Uh, that can mean a million different things. Uh, what does that mean? Do you own the land? Do you own the data center? Do you own the cards? Do you own a space within Azure? Like there's so many different are you partnered with the Neocloud?

What can you tell me about what it means to actually own or manage a GPU cluster? What are the best practices like what actually works? You're down dealing with depreciation. There's so much going on. Love to just learn more about like owning and managing a GPU cluster as a startup, not as some hyperscaler. Yeah, sure.

So, in terms of like getting GPUs, like there are a couple options like you could spin up on demand clusters. Yep. Uh which is more expensive. Uh but if if your workloads are like very like spiky, then that makes sense. But for us, we're doing constant research.

So we want something that we're like constantly using get to like you know 80% utilization.

Uh when you start to have that kind of utilization then you want to get like either reserved or your own cluster and you know do the math and like uh you know it often makes sense to buy your own cluster if you're crazy enough to like set it up if you have the you know if you have the expertise to set it up and and uh you could customize in all all different ways.

Um and so yeah we kind of went that route. Also it's just freaking badass and like we know we're going to get bigger and we want to have experience like building clusters. Um, I would say like we're not. So, you have servers like in your office like the cluster exists physically. Okay.

It's like George Hod literally has like server racks in his comma AI uh spot in in San Diego and it's like the craziest thing you've ever seen.

Then Nat Freiedman when he was doing NFTG they he spun up the Andromeda cluster and that was more like in partnership with someone else which made a lot more sense because like why do you have to learn how power routes necessarily like that's not necessarily the differentiator. Yeah. Yeah.

So we don't we're not big enough yet or our previous one wasn't big enough where we had to buy the land or buy the data center. We have a part data center but you know as we expand at some point we'll have our own dude. I see land and on your horizon. Well I think you got to vertically integrate.

You got to buy oil and gas tract. You you got to buy the the natural gas resources in the ground. Buy the sand that you can turn into silicon. Fab your own chips. Design your own chips. You got to be vertically integrated. It's the only way to win. From sand to search. Exactly. from Sam to search.

That's a great That's a great line. Iron t-shirt. Give him I' I've seen enough. Give him 10 billion. Give him 10 billion. Thank you so much for hopping on. Jordy, you got anything else? No, that's great. Awesome. Awesome progress. Congratulations for uh breaking it down for us. We will talk to you soon.

Congrats to the whole team. Have a great one. Yeah. Thanks, guys. Bye. Cheers.