He got rejected by 40 VCs & had 6 months of runway—2 years later, he raised $100M from a16z. | Edo Liberty, Founder of Pinecone

Edo Liberty left a high-paying job at AWS—where he was building AI at the highest level—to start Pinecone, a company no one understood. He pitched 40+ VCs, got rejected by every single one, and nearly ran out of money. Then, he flipped the pitch, raised $10M, and built one of the most important infrastructure companies in AI.

Then ChatGPT dropped.

Suddenly, Pinecone was the must-have database for AI apps, with thousands of developers signing up daily. The company exploded, leading to a $100M round led by Andreessen Horowitz and a 10x revenue surge.

If you’re an early-stage founder, this episode is a must-listen.

Why you should listen:

•How he went from from 40 VC Rejections to a $10M Seed Round

• Why he quit a High-Paying Job at AWS to start a Startup

• The game-changing shift that made VCs finally “get it”

•What really happened inside Pinecone when AI took off

•Why most founders misunderstand market timing and what to do about it

Keywords
AI, Machine Learning, Startups, Entrepreneurship, Vector Databases, Fundraising, SageMaker, AWS, Technology, Innovation, Pinecone, vector database, seed funding, ChatGPT, startup growth, business model, AI, infrastructure, early stage founders

Timestamps
(00:00:00) Intro
(00:07:50) Edo's Story
(00:12:27) The Early Days of Machine Learning
(00:32:23) Seed Funding
(00:42:09) Unsustainable Scaling
(00:53:41) Told You So
(00:59:24) A Piece of Advice

Chapters

• 7:50 Edo's Story
• 12:27 The Early Days of Machine Learning
• 32:23 Seed Funding
• 42:09 Unsustainable Scaling
• 53:41 Told You So
• 59:24 A Piece of Advice

Transcript

Edo Liberty (00:00:00):

I think the one thing that founders tend to misjudge is how much control they have over anything. in both directions. It's your company. You can just decide how things work and tell people what it is and just if they don't like it, they can go to help. You don't already experience that. You're not our customer. It doesn't matter. It just doesn't matter. It's not what you're going to build. It's not why it's important for your business. It's not how it's going to propel your application forward. It's not about the time you'll save. It's not about your developers, it's nothing. It's an API. It's a vector database. Here's the API, here's how much it's going to cost you. Here's the performance. If you like it, great. If you don't like it, go to hell. All the metrics, all the adoption, everything just completely, it was like a discontinuity.

It went from essentially nothing to something that looks like adoption. I remember towards the end of the journey I was, I was giving up, we had six more months of runway and effectively failed to raise the seed round. We're like, okay, now what do you do? Do you just stop and try a Hail Mary again in four months when hopefully something changes? There was one more meeting, screw this deck that we worked on that we tried and failed 40 times. I'm going to just rewrite the whole thing and suddenly this clicked.

Previous Guests (1:20)

That's product market fit, product market fit, product market fit. I call that the 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.

Pablo Srugo (00:01:32):

Do you think The Product Market Fit show has product market fit? 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 five stars. It lets us reach more founders and it lets us get better guests. Thank you. 

Edo, welcome to the show, man.

G(1:47)

Thank you man.

P(1:48)

 Dude, so I was just looking through and we were chatting the other day, but I'm looking through your profile here, you were at Yahoo and then you were at AWS and you're building SageMaker. It's like 2018, 2019. You're the peak of AI in AWS and you leave to start a startup. Why would you do that? Dude, what happened?

Edo Liberty (00:02:10):

That's a great question. I think it comes down to the vision of the market and where things are going, and I'll start by saying that except for that, everything else was exactly the wrong time to start a company. I had a 4-year-old and 1-year-old twins. My wife was working full-time at a hospital. I had a pretty fun and exciting job at AWS. We were building a lot of amazing technologies. I had like a hundred and some engineers and scientists working under-

Pablo Srugo (00:02:44):

I'm sure you were being paid quite well as well.

Edo Liberty (00:02:47):

Yeah, I don't even want to go into that. Obviously there's something really nice about a nice lump of cash landing in your bank account every month.

Pablo Srugo (00:02:59):

I say that by the way, all joking aside, I just wanna comment. It is a real thing in the sense that if you do the math, when you're thinking of starting a startup and you start from a position like you were at or at Google or Meta or one of these kind of top tech companies, it's actually quite hard to make the leap if you focus a lot on it just because in order- and again, not getting into what you were specifically making, but just as a general point in order for your startup for you to make as much from your startup, given the 10 years it normally takes, given the success that they normally have, et cetera, it's got to be a very big outcome. In other words, the math is almost not really on your side, not maybe the number one reason to start a startup.

Edo Liberty (00:03:43):

No, a hundred percent. to add to that, the rationale for starting a company being bad, I was 40, I had already had a pretty successful-, I was an adjunct professor and a director and this and that. I was already in a place in life where going back to walking in the desert and working my ass off, not seeing my kids and making less than an intern at Facebook is not necessarily the wisest decision and I'll give my wife all the credit for this.  I remember the kitchen table discussion where I said, I really do want go do this. She said, well, what's the worst that could happen? And I said, well, it's the worst that can happen is what happens with a lot of startup companies, which is I'm going to work my ass off for several years, make no money and it will end with heartbreak and then I'll have to go and find another job and we will have consumed some big fraction of our savings. And she was very cavalier about it. She said, sure, our kids go to school, they have clothes, we have a car. Just promise me we will be able to keep paying the mortgage. And so we took out the Excel spreadsheet, we looked at it, it's like, yep, yeah, we can do that. she said, go do it. That's your calling. That's what you want to do, you believe in building your own company and doing something huge and let's do it

Pablo Srugo (00:05:12):

Cool story. But it is actually on a serious note, very important to be aligned. A lot of founders, I mean- some founders are in their twenties and it's just like boom, go for it, why not? But then when you're a bit older and you have a family and you have a wife or a husband, and if you're not aligned, having that kind of short personal runway always eating at you is a recipe for disaster. It always takes so much longer than you think to kind of get off the ground and the last thing you want is to have to kind of forcibly fail because you just ran out of personal time or just got misaligned as a family and all of a sudden it's like, okay, I've got to go back to work. I thought in six months we would do it, but you needed two years or so.

Edo Liberty (00:05:54):

I agree, and also I think it has to do also with a personal lifestyle choice and what's important to you in life. To be honest, until I was 30, I was a scientist, I was a PhD and then a postdoc and then I started my first company and until I was- all my friends around me and my age group were already making big salaries and sort of living a very different life than I was living and I had absolutely no problem with it. I did something that I felt passionate about and I felt really strongly about and enjoyed my day to day and it's exactly the same decision with starting a company. That's what I will feel fulfilled doing. I believe in this mission, I believe in building something that I think would be long lasting and important to me and sort of like the final financial outcome of this thing was always… secondary is an overstatement. It was really not important. I'm very glad to say that we've done really well and I think the eventual outcome hopefully would be great for me personally, but even now with the day-to-day, even now to this day we're five years in, PineCone is a raving success and so on. I would've done much better to this day. I mean put my stock aside, I would have done a lot better had I stayed at AWS right and it's still fine. I'm still very happy with it. I still think it was an amazing decision for me. So a hundred percent. When you're a startup founder, you really have to think about what's important to you, what you are willing to tolerate, what is a good outcome in your opinion. My suggestion is to not factor in financials into what a good outcome looks like.

Pablo Srugo (00:07:36):

So we jumped in because of me to almost the middle of the story, but maybe you can take just a few minutes, give us a little bit deeper just your background and how you'd gone to this place where you take this moment to take the leap of faith and start a startup in the first place.

Edo Liberty (00:07:50):

I'll start very early in my career I started my degree in physics and I knew absolutely nothing about computers or computer science and I just figured I'll be a very bad physicist if I don't know how to code obviously, and so I took a minor in computer science just because I felt like it's going to force me to become at least a decent engineer.

Pablo Srugo (00:08:14):

And where was this? Where were you going to school?

Edo Liberty (00:08:16):

I was in Tel Aviv University midway. I fell in love obviously with algorithms with math and the computer science side of it ended up doing my moving to do my PhD at Yale in computer science. Interestingly, I sort of bounced around a lot in computer science between different domains, so I've spent a bunch of time on data science, on functional analysis, on graph theory, on high performance computing on different domains of computer vision at some period. Only after two, two and a half years in my PhD I really fell in love with what is now called AI and then before that it was called ML and-

Pablo Srugo (00:09:03):

What year was this that you fell in love with it?

Edo Liberty (00:09:05):

So this is 2003-2004

Pablo Srugo (00:09:09):

Okay. A while ago

Edo Liberty (00:09:10):

I remember I was doing computer vision on hyperspectral images, so these are images. Images are usually taken with three spectra: RGB, red, green, blue, the special microscopes that actually take 150 or something different spectra. That's incredibly important because different chemicals have different protocol color and so you can do things like biopsies and all sorts of analysis on these images in a way that- of course the human eye can distinguish. Also, we only see three colors, but computer vision you can really, for example, segment an image way better if you can see 150 different colors rather than three, but back then our machines had a 512 megabytes of memory and three images were big data, and so that kind of forced us- this first wave of AI machine learning people to really think very, very, very deeply about the implementation of everything. Exactly how you make everything efficient so you can actually do something with a very limited set of resources that you have and that created, that's where I found my passion on marrying the machine learning aspects of it and the optimization functions you're trying to solve with a very limited resources at hand. So this is very, very deeply intellectual and so the engineering of it and the math of it and the use case intermingled in a very deep way and I really fell in love again with the foundations of machine learning, all the optimization of later neural nets as well and high dimensional geometry clusters

Pablo Srugo (00:10:54):

I was going to ask, was there deep learning happening at that time with people talking about, because I know high level it was not highly thought of until they became very, very important, but I just don't remember the timeline of when that kind of happened.

Edo Liberty (00:11:08):

Deep learning wasn't a big thing. Again, deep learning even now is antiquated. People don't even call it deep learning anymore. That's already ancient history, but the same algorithms that train those networks we're training, even linear classifiers we're training other kinds of simple networks, shallow networks and so on. The optimizations that we see today, a lot of them are rooted literally 20 years back. Specifically the part that I felt super excited about was this viewing data as this distribution of these points in high dimensional space, which is how all machine learning models work. Okay. All the clustering, all the classification, everything works in some kind of the data is this data point, this point of this is a collection of high dimensional vectors. Yeah, I mean I spent multiple years doing that. Then started my postdoc and applied math, created my first company and then in 2016 I moved to AWS to basically start a new organisation. I wasn't leading the organisation, it was three people at the time who said, okay, we're going to build AI services out of AWS,

Pablo Srugo (00:12:17):

And this is by the way, the idea of AWS at that point was people would build what? computer vision on top, like predictive analytics on top or something else? What was kind the main products?

Edo Liberty (00:12:27):

Correct. It was very- it was early days for machine learning, linear classification, recommendation systems, XG boost and model hosting and data prep and all that stuff that we've been doing for 20 years. The world was waking up to them and we said, Hey, AWS needs to have a set of services. Again. For me that was extremely exciting in two ways, A. because it was a big play and we could actually go do something very meaningful, but also because the whole business model around cloud services is how efficiently you can do something, right? If you can achieve the same outcome with a 10th of the hardware, then you can charge a 10th of the price and still be profitable. And so efficiency and speed and accuracy and all that stuff became not only something that I enjoyed geeking about and building, it became the business offer.

(00:13:26):

It became the offering in the cloud. And so for the next four years, roughly three years, I spent building SageMaker and a bunch of other services, ended up managing roughly a hundred scientists and engineers working with every part of AWS and different storage search recommendations and so on. And throughout that time, vector search and what we now think about as what vector databases are was a fundamental piece of how almost every machine learning system was working, but we would keep building it into other products like if you're doing ad serving, you would have to have that capability inside your ad server if you're doing your recommendation engines. That's how your recommender engine worked, but you didn't think about yourself as building a vector label base. You thought about yourself as building a really high performant recommender system. And then 2017, I think late ‘17,  BERT comes out, which is probably - I'm pretty sure this is considered the first true public actual LLM that transformer models and so on. Very sort of quintessential technology product breakthrough. It's crude, it's hard to work with. It's kind of slow, it's unpredictable, it has many bad attributes, but it also does something qualitatively different and better and you're like, wait a second, there is a breakthrough here. There is something that we can do that we couldn't do before. Sure, it's very crude, but now we figured out we can do something we couldn't do before.

Pablo Srugo (00:15:08):

What was that something? What was the kind of key, the simple demo of the power of BERT?

Edo Liberty (00:15:15):

So people started scraping the top layers and creating embeddings with BERT, and suddenly semantic search started being a lot more powerful. You started being able to deal with language in a way that was not super brittle because again, before that, the keyword based search engines and so on, where you spend God knows how many hours tokenizing things and stemming words and all sorts of the journey of making a keyword search-based solution for search work well was months worth of data science and optimization and you had to be pretty deep into it and suddenly you could have not only all of that delegated to a model, but the results were just significantly better. You could just see that this thing sort of gets it in a way that none of the data pipelines that we've built before were able to do that, right? Again, it was far from being perfect. It was expensive and slow and clunky, but it did something qualitatively significantly better. And what happened in the market is suddenly embeddings and vectors and search and all that stuff started really popping up as something that engineers were talking about. It wasn't just the scientists and the PhDs sort of geeking out about this thing and the engineers saying, I dunno, I have no idea what you're talking about, but I sure hope it works.

Pablo Srugo (00:16:45):

Maybe for all the non-technical founders, what is a vector database? When is it used? What's it for?

Edo Liberty (00:16:51):

So the primary use case of that is really: when you represent complex data, say an image or text or whatever with a model, if you process it with an LLM or if you process it with a computer vision or image foundational model, what you get at the end is called an embedding. It's a numerical representation of that object that is literally a list of floating point numbers, right? It's the output of the layer in the network, which is just a list of numbers. That list of numbers tends to be incredibly rich and incredibly insightful about the content of this image. So if you process an image with a neural net, you can ask yourself, is there a person in it? Not necessarily that but whatever, what is the color of that pixel or if it's language, what does it mean? What does it align? So for example, a very simple example that shows that there's something semantically deep is that we would take networks that do machine learning translation.

Edo Liberty (00:18:00):

So you start in a sentence in Arabic and you end in a sentence in French. In the middle of that network, there is a numerical representation that knows nothing about the original sentence, but somehow it can translate it into French. So clearly the meaning is hidden inside that representation somehow. It might be opaque to you and me when we read this thing, but for models, it's everything. It has the actual meaning, it doesn't care about the tokens, frankly doesn't even care about the language. That is a breakthrough. You can suddenly search across all languages. Now what a Vector database does is let you search with those representation stores organisers and let you search and operate directly on those millions or billions of numeric representation of objects, and suddenly you can access and transact against really complex pieces of information that you just couldn't before. If it was, again, if it's multilingual search, if it's image search, if it's representing shopping preferences of somebody, whatever, and then now you can do a shop mirror recommendation. It opens the door to being able to access and deal with massive amounts of data that is unstructured or semi-structured or complex in a way that you wouldn't know how to code against it, but you can train a model to understand it and then you can work it with a vector database. The interesting thing is we talk about founder stories. I had at that point when I started Pinecone, I had built the internal workings of vector databases, God knows how many times, 3, 4, 5, 6 times in different products. I had myself experimented a lot and built my own libraries and so on. The term vector database didn't exist, it just didn't exist. Nobody talked about it this way. I had just figured that this is a wide enough use case. It's hard enough. So few people know how to do it well, and it needs to be done really well once and for all, so everybody can use a service.

(00:20:06):

But again, I didn't know how to call it a vector. I talked about vector search and about retrieval and about this semantic stuff and so on, but frankly, neither potential customers that I spoke with had any clue what the hell I was talking about nor investors, and this was probably the hardest part. I really had incredible difficulty raising money for this because I had just come out of AWS, I was known as the SageMaker guy back in the day, if you rewind back what is like six years? ML ops was the hottest shit in town. Everybody hosting models and Kubernetes automation and all that stuff, and everybody had assumed I was leaving AWS to build Pinecone and that pinecone would be an M and ops system. So every investor in the Bay Area was very happy to meet with me.

Pablo Srugo (00:21:03):

I would think you're incredibly fundable. Your profile reads a screaming, whatever this person's doing, write them a $5 million seat check and just wait and see.

Edo Liberty (00:21:14):

Correct. But then I would explain what I want to do.

Pablo Srugo (00:21:18):

Then you would start talking. <laughs>

Edo Liberty (00:21:20):

They’re like, I have no idea what this shit is. It's got nothing to do with what I thought you were going to do. Frankly, I don't get it.

Pablo Srugo (00:21:28):

I'm actually super curious. How did you at that point try to explain what you were doing or why it mattered? How did you approach that?

Edo Liberty (00:21:40):

I'm going to tell you what I did. I'm not advocating, I'm not recommending it as an approach. It was a necessity. Again, think about the fact that I was dedicating my life to build something that I talked to my potential customers and have no idea what I'm saying. That's not a good sign, let alone want to buy it. What we decided to do was basically, okay, let's just build the base system and let's put it out there. Let's just make it self-serve so people who actually know- because I knew a lot of engineers and a lot of people who build great systems need this. I should just tell them that it exists as a service and enough of them will figure it out and we will start getting more signal from the market. Sure enough, that's what we did. The very first system, we called it V1 , was a very, very good core engine that I built myself as a part of fundraising for this thing, but frankly, a pretty rickety distributed system. We had no billing and metering, we had no autoscaling, we had almost nothing, but the core performance was really, really good, and so a very, very early set of developers started picking this up. I remember in the early days, we literally would- we would have a Slack channel alerting us on people who access the console. It's not even people who sign up or start paying us. Even accessing the console was like, Hey, somebody's actually using this thing. This is amazing,

Pablo Srugo (00:23:07):

And who's by the way- you say “we”? Who's working on it at that point?

Edo Liberty (00:23:10):

So now we at that point are a very small team. We are essentially one person who's a good friend of mine who basically picked up all of the non-engineering stuff, basically called himself chief of everything else. Well, a handful of engineers, like five, six engineers

Pablo Srugo (00:23:27):

and how are you paying them?

Edo Liberty (00:23:29):

We had small friends and family initial funding, but we had not had a seed fund yet. So it took us -

Pablo Srugo (00:23:36):

You raised what, like a million or two or so? 

Edo Liberty (00:23:38):

We raised a little bit of money. Yeah exactly a small amount of money to fund the team and run some stuff. Okay. We went to the people who use our system. Again, we're talking about very small numbers. I'm talking about maybe 10, 20 teams who are using our service, and we just went to them and we asked them, what do you call this thing? We told you what the API is, but we never told you what this thing is called. <laughs>

Pablo Srugo (00:24:04):

This thing is so technical, it doesn't even have a name.

Edo Liberty (00:24:08):

No, we tried. It's not that we didn't try to name it. We tried a bunch of shit, but it didn't make sense. People didn't know anything and then we literally asked them, Hey, you internally in your team, you talk about using this Pinecone thing. What do you think it is? Well, what do you call this internally, 7/10 would say, we just call it a vector database. It's like that's where we have our vectors, we have our embeddings. You guys are the database and we just call it that. And he said, okay, then if that's what everybody thinks it is, and that's what it is.

Pablo Srugo (00:24:41):

I have to admit. I've done maybe what? A hundred, 150 episodes. This is a PMF show first that, I mean, you know what your thing is, but you don't really know what your product is. You have to go to your users to let them tell you.

Edo Liberty (00:24:56):

And then of course we started calling it the vector database. We start explaining what it is

Pablo Srugo (00:25:01):

And people are building. those first teams. What are they building with what they think of as a vector database?

Edo Liberty (00:25:08):

They're building search engines, they're building recommendation systems, they're building deduplication, anomaly detection kind of search stuff. All the LLMs and RAG and agents have not been created yet. This doesn't exist yet. Interestingly enough, talking about a founder story I had at some point I figured we really don't know how to talk about this thing. We're doing an absolutely abysmal job educating the market about what this thing is. One of the very first people I hired was a person called Greg Cogan who came in as a VP of marketing and his job was just to explain to people what this thing even is. Go try to find a way to get people to even know about our existence. We tried, I'm telling you, we tried everything explaining how to use it and what it's for and what the business value is and what the added benefit is. We got so much bad advice.

(00:26:05):

It's insane, insane. Every investor, every business leader, every marketing guru, everybody gave us the absolute worst advice. And I remember a very important meeting with Greg. We were just kind of sitting there wrecking our brains about what do we do? Nobody gets any of this. And then we had the epiphany of like, wait a second. If you don't know what a vector database is, and when we tell you vector database and you don't know what it, you can't already guess that this is the thing you're looking for because you have embeddings, you're doing search and you're doing vector search and you're trying to scale and it's painful and hard and so on. If you don't already experience that, you're not our customer.

Pablo Srugo (00:26:47):

That's right.

Edo Liberty (00:26:48):

It doesn't matter. It just doesn't matter. I remember we just said, why don't we try to just scrape the website of anything that has to do with value? It's not what you're going to build. It's not why it's important for your business. It's not how it's going to propel your application forward. It's not about the time you'll save. It's not about your developers, it's nothing. It's an API. It's a vector database. Here's the API, here's how much it's going to cost you. Here's the performance. If you like it, great. If you don't like it, go to hell. Amazing. All the metrics, all the adoption, everything just completely was like a discontinuity. It went from essentially nothing to something that looks like adoption.

Pablo Srugo (00:27:32):

What do you think was happening? Because I understand the point that, hey, if you're not an ICP right now, you don't matter. That makes sense to me. But for your ICPs, what was happening before? they were looking at your website and even they were confused and not pushing through, or why do you think that led your ICPs to convert better?

Edo Liberty (00:27:48):

Correct, because when you're an engineer and you're looking for a a vector database, you don't want to hear about how much it's going to streamline your- whatever,

Pablo Srugo (00:28:07):

All the marketing

Edo Liberty (00:28:08):

streamline the relationship between engineers and data scientists. You're not connecting your whatever. You don't care about it, all of that already. You are already trying to build a goddamn thing. You have a pretty good conviction that what you're doing is important. You just need the tool that solves the problem that you have. And so if we waste your energy or time or distract you with traditional marketing stuff, it's a waste of energy and it alienates you in waste, frankly confuses you. So you're like, oh, what is this thing? Then it's a platform for this. It's a dev tool for, I don't know what this shit is. Oh, it's a database and you do vector search and you do that at scale and this is how much your workload is going to look like. That's exactly what I'm looking for.

Pablo Srugo (00:28:57):

So I'm obviously not a technical founder, but my kind of simple mind looks at this and thinks to myself, the kind of analogy that comes to my mind is you go to Home Depot and I'm also not a carpenter or tools person by any means, but there is so many different tools, literally tools, you mentioned an engineering tool, but tools that somebody that wants to build stuff needs and some of them are like a hammer and a screwdriver, super simple, but some of them are very, very specific. If you need to cut wood in this specific way for this specific thing, this is the thing that you need. That maker of that tool doesn't try to simplify or explain to the rest of the world why somebody would need this thing. To your point, the person who needs that thinks knows exactly what they need and why, and when they see it, it's immediately obvious and anything else would kind of be a waste of time and maybe add more confusion than anything else.

Edo Liberty (00:29:51):

Yeah, it's a perfect analogy and frankly, I am a hobbyist carpenter. I used to build a lot of our own furniture, and you're absolutely right. I frankly actually use this exact example with some of my marketing folks sometimes. It's exactly that. When you go and look at a mitre saw and you look at the specs and what it is and how it's promoted, if somebody wrote to you that building chairs is fun, you'd look like a fool. You'd say, I'm never going to buy this shit because it’s clearly not- What is this? Of course not. I just need to know whatever it has, the features that I need, if it fits in my workshop, whatever, can I afford it? Whatever, all that stuff, right? So yes, I mean that was a pivotal moment like, wait a second. Our buyers and our ICP and the people we speak with are not what have application builders or scientists or any of those things or engineers. They're engineers who build big, hard complex systems and they need a professional grade tool.

Pablo Srugo (00:31:01):

And what year is it, by the way, when you make this change and start to see that traction?

Edo Liberty (00:31:05):

So this is 2022 roughly, so two and a half years 

Pablo Srugo (00:31:11):

You love this show, you don't want to miss the next episode. Why would you so hit that follow button? Trust me, it's in your own best interest. 

What about on the fundraising front? Had you been able to communicate the story to VCs that you already raised like a big seed round or series A whatever? Where were you at?

Edo Liberty (00:31:27):

Interestingly, again, this is our fundraising or seed fundraising spend was sort of- this epiphany came in roughly towards the end of the seed fundraising round. And so I have at that point spoken with every VC in the Bay area and have gotten a brutal rejection from each and every one of them. I swear to God there's not a third rate VC in the Bay Area that I have not pitched to with everything I got and got the door slammed in my face. I remember very much at the end of this journey and we had a pitch deck and this and that we worked on and really fine tuned the hell out of it and thought it was as good as it was going to get. And I remember towards the end of the journey I was, I was given up. We had six more months of runway.

(00:32:23):

We had effectively failed to raise the seed round. We're like, okay, now what do you do? Do you just stop and try a hail Mary again in four months when hopefully something changes with two months left of runway and have your back to the wall or whatnot? I had, there was one more meeting where I was going to cancel it to be honest. I said, okay, fine. It's not going to work anyway. It's stupid. We have tried like god knows, 40 times now, it's stupid. Whatever we got doesn't work. So I remember the meeting is the next day, it's the first day of my son's first grade in Covid, so I have to manage that. My wife is working in the Covid ward at Stanford. I'm alone. I remember 10:00 PM the day before. I'm like, screw this deck that we worked on that we tried failed 40 times.

I'm going to just rewrite the whole thing in a freaking whatever, Google Docs. I'm just going to do the most bare bones thing, but explain it in a completely different way. And I went on the call and I said, Hey, forget everything that I told you. We're just building a different kind of database. That's it. It's a different kind of database. This is what it does really well. This is what it doesn't do. I think it's big and here's why that completely changed the story. Suddenly this clicked, and so within literally a few weeks we had secured our seed funding from Peter Wagner who leads Wing, who had before that invested in Snowflake and others who really knows databases, knows the data infrastructure space really well and sort of figured there's something there

Pablo Srugo (00:34:04):

That was the $10 million kind of seed,

Edo Liberty (00:34:07):

Correct. So this completed the $10 million seed, which gave us enough time to go grow.

Pablo Srugo (00:34:13):

One thing I do have to ask is, it's funny, this happens a lot when something finally clicks in a sense when you tell me now, well then we just went in and just said, we're building a different database and it resonated. I'm like, well, what the hell are we doing the other 40 times? But it makes so much sense afterwards. But I'm curious before that, that deck that you destroyed, what was that story?

Edo Liberty (00:34:34):

We were trying to explain the value, so we looked at the applications that you build with vector databases today, which are still mostly, I mean now that with LLM there's much more you can do, but back then it's the kinds of applications you build with this thing and we try to explain that you need this specialised system to be able to build all these different applications, but then it's very confusing because it's not clear what you're building. Are you building a platform of solutions that you can do recommender systems on that you can do semantic search on what is this thing? It's not clear. 

Pablo Srugo (00:35:17):

It's not simple enough and it's not pointed enough. And I'll say one thing, VCs, I think we're tech VCs. We pride ourselves on understanding technology, but the reality is, and I was speaking for myself and most, some are truly technical, but the vast majority have a surface level understanding of this stuff. If you start going deep and technical with these guys, they usually can't keep up. And so in a sense, what you had to do was really boil it down and simplify it. And that's why sometimes again, analogies work, they really work when it comes to fundraising. So that's why you get this Uber 4x kind of mania throughout the 2010s and then it gets overdone. But in your case it was like, listen, our database is big. Yeah, database is big. We're building a new database and there's going to be a world that needs it. Here's why. It's like, okay, that kind of makes sense. You know what I mean? All of a sudden you're talking their language. So it makes sense to me that it would resonate

Edo Liberty (00:36:07):

In hindsight, it all makes perfect sense and I remember by the way, even if after we made the first PR about our fundraise from Wing and we said, Hey, we’re Pinecone, we're building a vector database, people called me concerned like, Hey, who cares about a new vector? Who cares about what's a vector? What do you want from us? Who needs a new database? Are you insane? Databases take a decade to build. Before they even produce Jack shit, what are you doing? At that point, I had the very good fortune of having Tim Tully, who is a partner at Menlo, be a friend, and he used to be the CTO of Splunk, and before that we used to work together at Yahoo. We ran a lot of the data platforms at Yahoo. So we worked together, together with Ram, our CTO, who's still our CTO at PineCone.

(00:37:05):

So all three of us worked together. Interestingly enough, he had tried to hire me into Splunk before that from AWS, and I decided to go start Pinecone, but a couple of years later he had left Splunk and became an investor. He knew both me and ROM and knew the technology because we have built this inside Yahoo, full God knows how many applications, so he knew exactly what we're doing. It was clear as daylight. I didn't have to explain anything. I just said, Hey, do you want to go do this? He’s like, hell yes,

Pablo Srugo (00:37:37):

Different conversation.

Edo Liberty (00:37:39):

There's no way I'm not leading this round. We just have to go do this. So I had the very good fortune of having somebody like Tim lead the A round,

Pablo Srugo (00:37:49):

Which was what, a year or so after your seed?

Edo Liberty (00:37:52):

That was probably a year after.

Pablo Srugo (00:37:53):

Okay, and what had happened through that year? Where was the business at? What was the business model? Even at this point?

Edo Liberty (00:37:58):

We had always been exactly the same thing. We had a fully managed cloud service vector base with the understanding that it's hard to do, it's hard to manage, it's hard to get the efficiencies in. The only thing people could do at the time was really spin up all sorts of open source libraries and primodial versions of what a vector database would end up being. Also containers and so on, but it was incredibly hard to manage for you to use anything. You had to really become really well versed in algorithms and how to optimise whatever floating point representations of whatnot and choose exactly the hyper parameters of everything. It was a mess. Nobody wanted to do that. You said, we need to do this really well. Out of the gate, you bring your workload and everything should just work. Mind you, by the way, there were many other complications of simultaneous reads and writes and scaling and horizontal and vertical. Well, it was clear that to do this in a way that was actually convenient and cost effective and scalable, you had to manage a service. Okay, there's no way-

Pablo Srugo (00:39:07):

And you're charging by what? you're charging per API call or what's kind of the-

Edo Liberty (00:39:11):

So back then, we charged by something we called a pod, which is just a fraction, essentially like a small node, like a fraction of a node. So when Tim comes in, the business is essentially not really existent yet.

Pablo Srugo (00:39:26):

You're still building, you're still kind of building the product.

Edo Liberty (00:39:28):

We're still building. We have start having very little traction. We are- I forget what the revenue was, but it was tiny. 

Pablo Srugo (00:39:38):

100K sub 100K?

Edo Liberty (00:39:40):

Yeah, something like that. Say 100k, ballpark. I mean it doesn't matter. I forget-

Pablo Srugo (00:39:45):

It is insignificant. 

Edo Liberty (00:39:46):

Yeah, could even be less than that to be honest. I don't remember exactly, but the usage started picking up. You could start seeing customers and people come in, adopt, start spinning up, use it, it started working again, the marketing, the product was stable enough and the workloads were growing. So again, when we started a well sized vector search use case was, I don't know, like a hundred thousand vectors was great. It was like a good sized workload, right? A million. It was like a lot. If you try to do 10 million, that was like a flex. You're just trying to show off that you can do it, but you started seeing that grow as well, so it was clear the market had demand for it. Clearly we had not landed the leadership of it. We have not cemented our whatever dominance in the space. A lot of different things were really sort of very, very nascent. But again, I had the very, very good fortune of having the investor who knew me, who knew my CTO and who knew the product so well that that was enough for them. That was like, okay, fine. This is going to be the winner and I'm going to go help with that.

Pablo Srugo (00:40:53):

And how big was the series A?

Edo Liberty (00:40:55):

Series A? We brought in 28 more.

Pablo Srugo (00:41:01):

Okay. And so he must have had- I think for him, obviously there's not enough traction for this to be a traction driven series A, so this is a bet on you, the team, but also he must have had, like you said, a pretty good understanding of the solution and a serious thesis that, because this is still, we're talking early 22, this is pre kind of the ChatGPT moment. There's still obviously a lot of stuff happening in GPT but he must have had, not necessarily known that that was going to happen, but at least a very big conviction on where AI was going and how important all this stuff was going to be. 

Edo Liberty (00:41:33):

A hundred percent. Interestingly, I mean just in an amazing timing of the market for us, we had built this service that was frankly too advanced and too good scalable for what most people were trying to use it for. And then essentially ChatGPT happens, the foundation models start kicking in. ChatGPT happens

Pablo Srugo (00:42:02):

Inside of Pine Cone, what was it like when ChatGPT, end of ‘22 and you see that happen? 

Edo Liberty (00:42:09):

It's frankly insanity. So this is, again, we're talking about a very small team. What happens then is something frankly I think that very founders or very few founders actually experienced, which is it's a scaling of our service in a way that is frankly unsustainable. We cannot even meet the demand of what people want of us. At some point, again, I dunno if you remember the early days of that, but pretty much anybody that knew how to write any Python, the dentist who did the computer science at whatever high school, whatever, and a member of how to write Python now wanted to build a little agent, try to figure out how to build RAG and all that stuff, and everybody was basically, we were the only game in town, and so everybody would just write in the docs, go get an API key from Pinecone and go spin up this and now your Jupyter Notebook or whatever your Python app is going to start working with a vector database. We started getting thousands of signups a day. We would spin up more and more and more machines. We got to the point where we would saturate whole regions in AWS and GCP. We just couldn't meet the demand. It was insane. We started spending insane amounts of money on the free tier.

Pablo Srugo (00:43:42):

ChatGPT is at the end of ‘22, but at first people are just using- I mean the mainstream people are just using the chatGPT and just shocked with the power of it. But what's happening on the technical side, is it right away that people are kind of using the APIs and whatever and trying to build apps with it?

Edo Liberty (00:43:58):

Immediately? It actually started picking up even before, because ChatGPT was only the chat interface like the-

Pablo Srugo (00:44:04):

That's right,

Edo Liberty (00:44:05):

GPTs and LLMs and started being already pretty good and much more. Well, that improvement started happening even before chat g pt, so you saw that pick up, but chatGPT really put everything on the front stage and it became the main conversation everywhere, and so not only was the capability there and we were the only game in town, it became a top down directive from every CEO and both of directors to go build stuff and everybody was super-, all the startups already was like the firing shot of like, Hey, let's go build the next what have you that's LLM based. RAG, which is a retrieval augmented generation, which is pretty much the only way to add your own data and your own context and bring the value of the data that you have and expose it and make it sort of useful. Bring the knowledge that you have into the model that it was pre-trained on. Of course not your data. It didn't train the model. That became the only true, fairly well understood, easy enough, good way to add knowledge to AI applications, and so that started taking off like insane. Like I told you, literally thousands of signups and conversions to paid customers a day, we’re spinning up an infinite amount of hardware. Again, it was sort of insane. We didn't sleep for three months. We would wake up in the morning and revenue's up 5%. It was like, okay, what happened? We don't know. Let's go look at the logs.

Pablo Srugo (00:45:35):

Give me a sense, we don't need to get into revenue today though. Whatever you share is great, but at that time we were talking about by the time of the series A early ‘22, you’re insignificant revenue sub 100k, when do you hit a million? When do you hit two? Are you doubling every month at that point? You know what I mean? What is that curve like in those early days? 

Edo Liberty (00:45:58):

Yeah we hit- So our series, so this kind of fits almost perfectly into the series B, which was eventually led by Andreessen Horowitz. We finished that year with roughly 2 million in revenue and roughly-

Pablo Srugo (00:46:12):

‘22? the year of GBT or the year after?

Edo Liberty (00:46:15):

Yeah, ‘22.

Pablo Srugo (00:46:16):

‘22 with 2 million. Yeah.

Edo Liberty (00:46:17):

We finished with roughly 2 million in revenue with a few hundred customers and an insane growth. All the graphs are just not normal to the point that it's just not sustainable. Clearly, something insane is happening here. The interesting thing is, this is kind of funny also, founding story, this is the series B is the exact polar opposite of my seed funding. You cannot imagine a more polar opposite. It was just that we are the hottest company in the Bay Area. Everybody of course knows who we are, what we do and why it's important and not a single person. I didn't have to pitch anything, I didn't have to say anything. It was I was invited for VCs to pitch to me why I would be, I swear to God, I remember it as a cultural shock. VCs would send me their pitch deck on Pinecone about why it's a great investment.

(00:47:15):

Their entire research, they would send me 50 page documents on all the market research and exactly how we're the leader and how we're going to win and what we're going to do and why they think we're the best company to invest in. It was just nuts. I'm telling you. It was just not normal. The interesting thing is we got a very high good valuation in our B round. Maybe for this reason. Obviously we didn't try to raise so much, but you know how these things go. No, exactly. It was like super, super tonne of pressure to come in.

Pablo Srugo (00:47:46):

Well, everybody's coming at you and the valuations make sense. Yeah.

Edo Liberty (00:47:49):

I think some people felt like there's a handful of VCs that got priced out. They say, oh, fine, this price is too high, blah, blah, blah. We closed a round between when we signed the term sheet and when we were closing, we 5x’ed the number of our paying customers and this is the year. This is the year after. Yes. This is beginning of ‘23

Pablo Srugo (00:48:15):

And this is a year after, so you hit 2 million out of ‘22, 5x in a month, then a month or two. That's how long it takes to close. Insane. 

Edo Liberty (00:48:22):

of course, it's the number of customers, many of them are pretty small, whatever. It's a managed service. There's a very wide distribution of sizes of customers, but still people who converted to paying customers and are paying you some number would've some amount of money, usually very small numbers obviously. Again, most use cases were pretty small. A lot of people are learning and so on. The most interesting thing in that journey, and then this is again 2023, we start seeing for the first time, so we start understanding two main things. A.People use us as the system of record and really rely on us as a fundamental building block in their stack. The second we have a glitch on an outage or anything, Twitter starts exploding. It's like, holy shit. Pinecone having an outage and like holy shit, this is, I didn't-

Pablo Srugo (00:49:21):

Makes sense though, right? As a database that you would see-

Edo Liberty (00:49:23):

It does, it does, but it sounds insane. Of course now it's obvious. We are a database. We are foundational infrastructure layer in a very large application. We just, frankly, we were naive enough to frankly not even understand how critical we were in their infrastructure. This was even whatever, degradation in whatever, right throughput in a small region became a big issue. I'm like, holy shit, okay, fine. We have to really up our game A and B. We started seeing probably a year and a half before the rest of the market what scale the workloads in production truly look like because we're the managed service. We track everything, all the skews of data and the workload sizes and shapes and throughputs and acceptable latencies and all the contentions between writes and reads and all that stuff. And we could also see the prices, sorry, the costs of running the system and we start to figure out, wait a second, if we need to do this at 100x scale of where we are today, we are going to have to build something very, very, very different.

(00:50:38):

something that separates rights and reads that has multi-tenancy sort of built in, that the storage is all BLOB (Binary large object). You cannot afford to use SSDs anymore. It becomes too expensive. There's so many core tenancy you had to change in the system for it to be able to work at those sizes and we embarked on building what is now Pinecone. The main product for Pinecone is a serverless product, fully managed serverless product. We basically redesigned the whole DB, the core indices. The core algorithms of course were reused and improved and so on, but we had to rebuild the entire system, which is a huge bet for a company, for a database to go rebuild what it is mid-flight and keep serving all these workloads was a massive strategic bet that we sunk in probably more than a year of the company to almost stop improving the old product must build the future,

Pablo Srugo (00:51:40):

And that launched when?

Edo Liberty (00:51:41):

That launched at the beginning of 2024. Amazingly enough, we could not only reduce our own operating costs significantly and roll on the prices with a price decrease for customers, so now we could run, we can offer the service to be 10x cheaper. So people who used to pay us 3-4,000 a month started paying 200 bucks a month and it became a no brainer. So everybody started scaling because now they could, everybody for whom price consideration was an issue or scale was an issue, started scaling a lot more. We now have more than a year of operating this new system. It's a heck of a lot better. It's a heck of a lot faster, cheaper, more performant, more stable and so on, and we're still advancing on it significantly. In fact, we have a major version coming out very shortly that we're not talking about yet, but it will be yet another significant step forward on scale and production readiness of these systems that really extreme ends of the space and so a lot of bets, a lot of early bets really paid out in a massive way.

Edo Liberty (00:52:56):

The fact that we were a system, the fact that we could manage the workloads, the fact that we could see what people are trying to do with our system and hitting limits and the fact that we could actually improve the software for them and have the applications just continue to run without a hitch, that was massive. There are very few other companies that I know about can do that. to this day. We ship very significant improvements to our system and completely change the serving layer and applications that do literally hundreds or sometimes even more queries per second just continue as if nothing happened and run in production even though whatever we change the engine mid flight.

Pablo Srugo (00:53:41):

Let me stop it there. And normally we finish with two questions and the first one's usually like, when did you feel like you had product market fit? I feel like you kind of answered that. Certainly that GPT and certainly the chatGPT moment, I think were big ones. Let me ask a different question, which is you built vector databases when it was kind of so early you didn't even know what to call them and everybody thought that what you did make no sense and then you were right there when the world effectively changed from the outside looking in, looking like zero to one, but obviously there was an undercurrent, but my question is what does it feel like to be right, to be right in a way where everybody else had no idea and you just were so right? What's that feel like as a founder?

Edo Liberty (00:54:29):

I mean obviously it's very validating. I wish I could tell you I had a round of victory telling everybody that I told them so but it didn't happen yet. Yeah, you're so busy executing and building and serving customers and solving problems that you don't, I don't think about that.

Pablo Srugo (00:54:47):

I'm going to build an app. It's going to be called the told you so app, and every time a VC rejects you just add their email there and then one day you click a button and they all get told.

Edo Liberty (00:54:57):

 I'll tell you one funny anecdote about this. I remember meeting an investor in the B round where we were the hottest shit in town that had rejected me. My scene in the meeting, he said, Hey, I looked up my notes and I now understand that what you were pitching that you were going to do, you actually did. You actually did the thing you said you were going to do, and I said, yeah, I mean isn't that how it usually works? He was like, no!

Pablo Srugo (00:55:25):

That's awesome, man.

Edo Liberty (00:55:28):

So yeah, you had moments like that, but frankly I think the more exciting, oh, not the more but thing that is just as exciting to me now, maybe as a scientist, maybe as somebody who's a big part of my cycles are dedicated to what does the world and technology look like in three to five to 10 years? We have just as much conviction about where we're going as the amount of conviction that we had five years ago. If you look at agents and what they're trying to do now, how conversations, we have people building conversational agents on top of Pinecone, efficiencies inside big companies, support, personal assistance, you name it.

(00:56:22):

The vision of making knowledge and information regardless of where it comes from, whether it's the text file or an image or a JIRA ticket or a legal contract or whatever, a sales call making that organised and summarised and indexed and ready to be actionable in real time as insights to those agents and to those systems and those search platforms and so on. That is still happening and that is still hard and a vector database is a big part of that is probably the biggest part of making that production ready at scale. But there's so much more and we see that the arc of the company and the journey that we are on really has just begun, to be honest. I mean we in the sense that in the same way that BERT was sort of like a spark and we know something is happening here. We can do stuff we couldn't do before, but it's still raw.

Edo Liberty (00:57:24):

This is where we are today with the spectral knowledge. We now understand RAG. people do RAG with Pinecone all the time. It became a very well understood pattern. People now synonymize us with that. People start powering agents with us and so on. It's doing qualitatively better than anything else you can do, but it's still hard. It's still clunky, it's still complicated. People still need to figure it out. We are now with knowledge where we were with models five years ago. If you talk to a big company building applications in house or you talk to a startup or a company that has a lot of data that specialises in say, support or some what have you, conversations and agents and whatnot, we are still very far from them taking it for granted that all the data that they have, all the information they've gathered, they can easily put into Pinecone, and that is obviously perfectly accessible and understood and organised and serviced at the right moment to the right application.

(00:58:28):

And that obviously works perfectly. That's the journey we're on. Okay. you are going to- I'm telling you, in five years the thought that maybe the conversational agent that you are talking to inside your company hasn't listened to all the sales calls that ever, ever happened. The company knows everything about your product and can talk about it intelligently. The thought that that's not the case is going to be weird to you. Obviously it knows everything we'll get there, but there are at least five years worth of innovation and building in front of us to do that in production, at scale with accuracy, with all the bells and whistles and all the models and all the hooks and everything for everything to work perfectly. So I have no doubt that this is happening and I have as much confidence that this is going to be the future in five years in the same way that I had confidence would be here with Pinecone five years ago.

Pablo Srugo (00:59:24):

And then the last question is, what would be your number one piece or a top piece of advice that you would give an early stage founder today?

Edo Liberty (00:59:33):

That's so hard. so I work with a lot of founders, I really work with a lot of early stage funders because I love it. I think it's such a hard job and it's such an ill-defined journey you're embarking on that every person needs a slightly different kind of coaching and support. And I had the very good fortune of being surrounded with somebody like Bob Muglia who eventually joined my board, and he's one of those people who always can zoom out and tell you, okay, where things are going and why and why you should be worried about and why you shouldn't, so on. I think the one thing that founders tend to misjudge is how much control they have over anything in both directions. Sometimes you feel like you're constrained by your team or by your customer or by this or by that, and you feel super stressed and boxed in about a bunch of stuff where somebody needs to come to you and say, yeah, you can just, it's your company.

You can just decide how things work and tell people what it is and just if they don't like it, they can go to help. At the same time, you also overestimate your ability to influence the market and the timing. One of the things that we have had amazing good fortune with is ChatGPT and AI shooting up to the sky. That's a market thing. We didn't do that. We had the very good fortune of being in the right place at the right time, but we didn't build OpenAi. OpenAI happened on its own and frankly, we are now going through another phase like that where models are changing and all the dynamics and whatever, everything that's happening in the maturing of the market, that takes time. That's going to take three, four years. It doesn't matter how good your product is, it just doesn't matter. You just have to wait. You just have to be there, right? I think the main advice is really, really try to be very open-eyed about what you control and what you don't control, and sort of go with it. It's sometimes even very critical things for your business outside of your control, and you just have to just stick it out and hope for the best. And sometimes you have to go and change big things even though it's scary.

Pablo Srugo (01:01:51):

Edo, thanks so much for jumping on the show, man. It's been awesome.

Edo Liberty (01:01:54):

Thank you, man. Appreciate it.

Pablo Srugo (01:01:56):

85% of people who listen to the show just started listening to it this year. You're probably one of those people. In fact, the odds are 85%. Guess what? You need to go back. There are over a hundred other episodes that you need to check out that are just as good, if not better than this one. Don't miss out.