đ§đ Building AI-Native Infrastructure for Developers with Erik Bernhardsson, Co-founder and CEO of Modal
Erik Bernhardsson is the Co-founder and CEO of Modal. We talk about building what is essentially a new cloud provider, created from the ground up, optimized for AI.
We also talk about what actually happened before and after the great GPU shortage, how Modal fixes the inference problem in AI, and why he thinks AI will lead to 10x more developers.
He also shares lessons on culture from joining Spotify as the 40th employee, treating hiring like a prediction problem, what most people get wrong working with early customers, why more people should start companies in their 30âs and 40âs, and reflections on fundraising in a hot market.
Shoutout to Timothy Chen at Essence Venture Capital and Erikâs Co-founder Akshat Bubna for their help brainstorming topics for this conversation!
Support this Episodeâs Sponsors
Meow: The lowest cost banking services for startups. Get free bookkeeping - click here and tell them Turner sent you.
Hanover Park: The modern, AI-native fund admin for institutional investment firms. Try it here.
To inquire about sponsoring future episodes, click here.
đ Stream on Spotify and Apple
Timestamps to jump in:
4:26 Modal: AI-native infrastructure
9:02 Why its so hard to get GPUâs
15:00 Hitting PMF with AI generated media
20:37 Competing in IOI competitions
23:09 40th employee at Spotify
27:17 Lessons from Spotify
31:17 Starting Better[dot]com
34:05 Treating hiring like a prediction problem
36:12 Erikâs favorite interview question
39:07 Sales, and common mistakes with design partners
42:02 Startups should solve hard problems
44:05 Evolution of Modalâs product over time
50:15 Rise in importance of inference in AI
52:07 AI development post-GPU scarcity
58:51 How to build a brand in dev tools
1:04:31 Fundraising from Seed to Series B
1:07:42 30-40 year oldâs should start more companies
1:10:00 Reducing developer tax, increasing productivity
1:20:37 Why Erikâs bullish and bearish on AI
1:26:17 Bubbles, downsides to inappropriate valuations
1:34:58 Why high CO2 levels make you dumb
1:37:38 Difference between US and European startups
Referenced:
Careers at Modal
Find Erik on X / Twitter, LinkedIn, and his blog.
Related Episodes
đ Stream on YouTube, Spotify, and Apple
Transcript
Find transcripts of all prior episodes here.
Turner Novak:
Erik, welcome to the show.
Erik Bernhardsson:
Thank you. Itâs great to be here.
Turner Novak:
Yeah, this will be fun. We were just talking about trying to learn English and different random words in different languages. I hope weâre going to learn a lot about just generally everything that youâre doing right now in your life throughout the conversation, but super quick, can you give us just like a super high level for someone who is not familiar with Modal on what it is?
Erik Bernhardsson:
Yeah, totally. The way I usually describe it is AI infrastructure, although thatâs a little bit of a simplification, but-
Turner Novak:
Thatâs good, though.
Erik Bernhardsson:
... yeah. Thatâs sort of condensed into two words.
Turner Novak:
Super simple and then get a little bit more complicated.
Erik Bernhardsson:
Exactly. Yeah, so if you think about a lot of traditional infrastructure like Kubernetes and Docker and things like that, a lot of it works really well for backend applications or traditional kind of products like if youâre building whatever retail online store or something like that. Thatâs what a lot of traditional infrastructure has been built for, but in this new world of AI, when you start to deploy models and start to deal with GPUs and scaling up and down and running all over the world and developers iterating, a lot of the infrastructures doesnât really work super well. That was the sort of impetus for the genesis for me starting to think about like, âCould we actually build a better infrastructure provider?â, back in 2020 or 2021 when I started Modal.
Kind of realized in order to do that, we had to throw out all the existing infrastructure, start over with a new platform, kind of rethink the entire developer experience, and rethink a lot of managing containers and managing large models and working with GPUs. Itâs been a couple years kind of rethinking that and building a lot of foundational infrastructure. Today, we have a platform. We basically you can think of it as almost like a cloud provider in itself. We manage many thousands of GPUs. We run a lot of different types of applications, generated media, large language models, vibe coding platforms, and training and inference and all kinds of other stuff. The core idea is we donât build the applications. We build the infrastructure below, and weâre building kind of a new infrastructure stack thatâs more well-suited for AI and other stuff sometimes.
Turner Novak:
Maybe this is either like a really dumb or a really insightful question, but is there something that was going on that made it so that you couldnât really use the traditional infrastructure for doing AI stuff? Are they designed differently? Are they too slow for processing things? What is it that makes them so much different? Are they more focused on storage or something?
Erik Bernhardsson:
No, I think thatâs a good question, and I donât know if I have an amazing answer, but I think that a lot of the existing stuff works for traditional applications. My experience, and by the way, this is kind of going into my background a little bit, but I spent a lot of time at Spotify for seven years, and most of the time I spent at Spotify, I built the music recommendation system. A lot of this like sort of the perception of infrastructure is broken was actually shaped by me building a lot of the music recommendation system at Spotify and realizing 90% of the work is just building infrastructure, not application.
What are the things that are broken? Thereâs a couple of things. First of all, a lot of traditional inference like not built for fast feedback loops, and it turns out when youâre building these AI applications, itâs kind of annoying to test locally and to then deploy it into production. You can either kind of split it in two different environments and you have like a local environment you test in and a cloud environment you deploy into. Then, instead, you replace it with a different problem where you have this problem of like going from one environment to the other. This is not as much of a problem in traditional application development because a lot of infrastructure is pretty good at mocking the cloud development when youâre developing locally.
In AI, it gets really hard because you have GPUs and you have very large models, too. The initialization step is very complex. You need to load all these models into GPUs. I think a lot of it comes back to like originally how I thought about it was really developer experience and fast feedback loops. I wanted something where I can just write code and then hit enter and just run it in the cloud immediately. Then it turns out thatâs actually really good also for just running applications in the cloud because it enables you to scale up and down very fast. Then, it turns out with these new applications, another thing you have within AI land that you donât have in backend land is GPU capacity problems.
Now, you have to start to think about, âOkay, you need to run multi-region and you need to aggregate capacity from all these different places in order to guarantee capacity.â Thereâs no region and no cloud provider has enough capacity to handle this dynamic flexible scaling. Most of our workloads are inference. Thereâs a number of things that sort of break with a lot of traditional backend infrastructure, which is why we decided to start over and build our own.
Turner Novak:
Then, whatâs the big issue with GPUs? You kind of mentioned that was really a pretty big deal for somebody whoâs never... I mean, maybe you can say like whatâs a GPU versus CPU? Maybe thatâs an interesting way to start it, but then why is it so hard to get them and whatâs the big problem going on there right now?
Erik Bernhardsson:
Yeah, I mean, GPUs, and hopefully most people have heard the term, they do a lot of computation in parallel, and it turns out itâs very good for AI applications because you need to do a lot of calculations or flops. Some of those people talk about floating-point operations. I honestly think a lot of it just comes back down to the fact that theyâre expensive. The fact that theyâre expensive means a couple of things. One of itâs like you want to really push utilization of those. Actually backend application, it doesnât matter if you have a lot of CPUs just sitting idle most of the time.
Itâs not like a crazy amount of costs. Actually, most of the cost is electricity anyway, so if you scale up and down, if the GPUs are just sitting idle, itâs actually kind of find, but GPUs, most of the cost is the GPU itself. One of the problems is just you want to keep them utilized. The other problem is because theyâre so expensive, you canât assume that thereâs an infinite amount of them in the cloud, kind of like CPUs. CPUs, actually, you could sort of assume a little bit more at least. Thereâs an infinite number of CPUs in the cloud, but GPUs pretty quickly run into capacity problems. I think thatâs a big difference.
Turner Novak:
Just because theyâre so expensive and there just hasnât been enough that have been created yet that are just out there for people to use?
Erik Bernhardsson:
Yeah, totally. Yeah. I mean, theyâre like 10X or a hundred X more expensive than CPUs at this point.
Turner Novak:
Iâve actually never really looked into this. This actually might be a dumb question. Why are there not more of them? Why havenât we made more? Are we like capacity-constrained because of NVIDIA is just not making them? Or-
Erik Bernhardsson:
Yeah, exactly. TSMC or whatever. We just canât make enough GPUs. Anyway, I think eventually itâs going to catch up, and itâs going to be interesting to see what that means for GPU process and for resource management and all these things. Maybe it goes the same way as CPUs. Maybe we wonât care about utilization. Maybe they come super cheap. I donât know what happens to AI at that point. I donât know. Itâs going to be really interesting to see. I think weâre going to get there eventually. It might take three years, it might take 20 years. I donât know.
Turner Novak:
Yeah, so this was sort of identified as a big problem sort of in end of â22, beginning of 2023 this lack of GPU supply. What kind of happened just over the next couple of years just in the industry, from your perspective, as things kind of started to evolve?
Erik Bernhardsson:
I mean, weâve definitely caught up a little bit in terms of capacity, and I donât really think people talk enough about this, but GPU price is actually down a lot, and if you look at a few years ago, there was this sort of almost like lore like, I donât know, like sometimes fueled by VCs unnecessarily.
Turner Novak:
VCs are the worst at fueling narratives.
Erik Bernhardsson:
Itâs just almost like narrative that like, âOh, youâre starting an AI startup? Well, you need to go buy a thousand GPUs. Otherwise, youâre not a real... By the way way, you need to raise a hundred million dollars in order to do that.â I almost felt like it became a self-fulfilling prophecy. It kind of reminded me of like the Great Toilet Paper Shortage during the pandemic. People just talking about shortage, then it becomes like, or I donât know, Silicon Valley Bank. They have these self-fulfilling problem where so everyone went out and bought all these GPUs. By the way, now itâs been like a year or two, we talked to a lot of these startups. Theyâre like, âYeah, weâre sitting on a thousand GPUs. We canât leverage them, all of them. We donât know what to do with them.â
I think this model we had a couple of years ago where people just bought a lot of long-term reservations, in many cases they got like three-year reservations. Itâs not good, and by the way, thatâs something Modal solves is everything with Modal is usage-based. You only pay for what youâre actually using. Itâs consumption-based, so you donât make upfront reservations or anything like that.
Turner Novak:
One of the issues you might run into as a smaller company, less cash to work with, smaller balance sheet, youâre going out and youâre buying these GPUs, but then youâre really not using them a pretty high percentage of the time versus if youâre using an option like Modal, you might pay a little bit more on a per usage basis, but youâre only using it a small amount of the time. Youâre only paying for what you actually use, so in theory, you actually save a lot of money.
Erik Bernhardsson:
Yeah.
Turner Novak:
Thatâs the pitch. Yeah, thatâs the customer pitch, the investor pitch.
Erik Bernhardsson:
Yeah. Yeah, and thereâs also like kind of the flip side is if your app goes viral on Hacker News and you get a million customers overnight, weâll scale for you, so you can sleep well knowing that we can scale up. If you need a thousand GPUs, we got you a thousand GPUs, like...
Turner Novak:
How do you get GPUs? Did you strategically play this out? Whatâs the story there where you have all this capacity for people?
Erik Bernhardsson:
Thereâs this guy in the alley I know and you have to go make this weird handshake and-
Turner Novak:
Heâs got a fringe coat, he opens it up.
Erik Bernhardsson:
... yeah, exactly, exactly. He has a truck, but he blindfolds you and then drives, I donât know where it is, somewhere in Jersey I think. No, we just spent a lot of time integrating with a lot of... I mean, we mostly use the hyperscalers, so we integrate with a lot of different hyperscalers, and it turns out they all have a bunch of spot capacity in different regions at different times. By aggregating, I think weâre using 80 different regions at this point. We can always find capacity. Weâre integrated with more and more cloud providers and more and more regions all the time, so by kind of putting all that together, we can get capacity pretty much, as much as you want anytime you want.
Turner Novak:
Interesting, so do they have this whole sort of instead of like a CPU rental business, they almost have this GPU rental business that theyâre also running?
Erik Bernhardsson:
Totally. Yeah. Yeah, exactly, and Amazon has had GPUs for a long time. I remember running deep learning applications on Amazon AWS course, I donât know, I think eight, nine years ago, so they started... Theyâve been around for a long time doing GPUs, too.
Turner Novak:
Then, you mentioned a lot of people are using Modal. Whatâs kind of like the most down-the-fairway and most common use cases you see people using it for?
Erik Bernhardsson:
Some of itâs generated media, so we found product-market fit in particular with like applications like Stable Diffusion. That was like the first mainstream application that we found that people started coming to us and running a lot of stuff. Then, it was fine-tuning. It was called DreamBooth. It was like one thing a few years ago.
Turner Novak:
Oh, I do not remember that one. It was like images, AI images?
Erik Bernhardsson:
You can upload a picture of yourself and then you could fine-tune a model based on that, and then generate realistic AI photos. The first PMF application, the killer app was like Stable Diffusion. I think the second one was these like AI photos. That was our second main applications, but then it kind of continued from that. We have customers over the next year or two. This is like mostly 2023, 2024 we saw a lot of influx that customers using us for video and music. Suno, for instance, is a big customer that used us for AI-generated music. You can put in like... generate a hip hop song about whatever, Kubernetes or cloud compute.
Turner Novak:
My daughters loves Suno.
Erik Bernhardsson:
Actually funny, my daughters, too. I have two daughters who also still love it, very scatological lyrics, but thereâs a-
Turner Novak:
Yeah. The lyrics are like, âMake a song about Daddy coming home from a trip to New York,â and thatâs literally the song. The lyrics are just like, âDad went to New York, he came home,â but they love it.
Erik Bernhardsson:
Yeah, my kids make way more immature songs, by the way.
Turner Novak:
Yeah, mine are very... mine are still pure. Weâre holding on as long as we can. Please just stay as innocent as possible.
Erik Bernhardsson:
Yeah thereâs a lot of flatulence in my kidsâ house.
Turner Novak:
Okay.
Erik Bernhardsson:
Thereâs like AI, stuff like that, but more recently we started seeing a lot of... I think it kind of reflects a little bit the adoption pattern of AI. Initially it was a lot of these pure kind of GenAI-native companies, but more recently youâre starting to see more kind of mid-stage/later-stage companies also adopting us. It tends to be more for like normal use cases, a lot of customer support or LM, like fine-tuning, stuff like that. Then, another couple of things weâve seen more recently, seen a lot of vibe coding platforms adopting Modal. We released itâs kind of a separate product in a way because itâs CPU-based, but we have a product for safe execution of third-party code, which means in most cases LM-generated code.
How these vibe coding platforms work is they ask an LLM to generate code, and then they need to execute that code somewhere to build that app. We have Lovable as a business customer. They use us for our sandboxing capability. Another use case has come up a lot in the last year or so is actually computational biotech. Weâre starting to see a lot of customers using this for protein folding, sequence alignment, like those kind of quantum chemistry. Itâs actually super cool, I find. I was talking to a customer the other day and theyâre using us to cure cancer. Iâm like, âNice, youâre not just face-swapping whatever.â Like-
Turner Novak:
Making fun songs.
Erik Bernhardsson:
... this is actually like... Exactly, yeah. Actually kind of fun. Itâs like the Sam Elfman comment. They didnât make a comment just the other day, something similar. Weâre like, âWe need the GPUs to both cure cancer and for like face-swapping or whatever.â
Turner Novak:
AI slop. Yeah.
Erik Bernhardsson:
Yeah, AI slop.
Turner Novak:
Wait, so you mentioned quantum chemistry. What is quantum chemistry? Iâve actually never even heard that word.
Erik Bernhardsson:
I didnât even know. I should know because I studied physics, but I donât know. Wave functions, I guess.
Turner Novak:
Okay, Iâm Googling it.
Erik Bernhardsson:
You have like probability distributions instead. Thereâs like Newtonian physics. This is my stupid trivialization. Yeah, itâs like Newtonian physics like sticks and balls. You have molecules, you have like these balls with sticks in them and you kind of just simulate it. Then, thereâs quantum stuff, and in that case itâs everything is like probability distributions instead.
Turner Novak:
Yeah, thatâs what it looks like this is. Just that it uses the principles of quantum mechanics to study atoms and molecules.
Erik Bernhardsson:
I think thatâs right. I should know.
Turner Novak:
Yeah, way over my head. I would fail that task. Actually, I think in chemistry in high school I got like a C plus or something and I was just like, âIâm done.â
Erik Bernhardsson:
I hated chemistry.
Turner Novak:
I probably should have, I donât know, just tried harder maybe. I was just not... I donât know. I was not a good student back in high school or college. Are there any surprising use cases? Maybe that is one. Is there anything else? Maybe like the most shocking way someone is using AI infrastructure to build like most shocking or most interesting product?
Erik Bernhardsson:
Last week I was talking to a customer who is using it for weather forecasting. Iâm like, âItâs kind cool.â I find it really interesting. It turns out the state-of-the-art weather models, you can run them on a single GPU. Kind of cool.
Turner Novak:
Oh. I feel like something like that could come into play like you... We were at a wedding a couple of weeks ago and the weather was really bad, and if you could pick, you probably would not want it on that day. That could be an interesting use case for just like long-term weather forecasting. If you can plan out a year in advance and being like, âThis day will be sunny, this is a good wedding day.â
Erik Bernhardsson:
I think thatâs just physics. You canât really predict that far out. I think now where we can reliably predict like seven days out. I think with better models we might be able to push that to like 10, 12, you know?
Turner Novak:
Yeah.
Erik Bernhardsson:
Thatâs going to be like take so much more compute, but itâs still 12 days out. Itâs like thatâs good.
Turner Novak:
Thatâs pretty good. Yeah, give someone 12 days notice on their rainy wedding. They can plan around it. Going back, I usually like to talk about when you first started the company, but I feel like the story kind of starts probably back in Sweden. I know we were talking before we started recording, you grew up in Sweden. What did you like to do as a kid back in Sweden?
Erik Bernhardsson:
Oh, I liked to code. I started coding â92, so my parents had an Apple Plus, Macintosh Plus. I started coding and it was like HyperCard. Do you know what that is? Itâs like a multimedia thing. Kind of like Visual Basic in a way, but like Mac. I figured out how to write code and it was very basic, simple, but yeah, so I actually think a lot about how grateful I am because back then, I had no idea that I could make money from it. I was thinking about my friends. They were trading hockey cards. That was their hobby. I was writing code, and look at me now, Iâm like making a lot of money from my hobby.
I actually am very, very grateful about that in a way. I donât know, I just kept coding throughout my childhood. I switched to JavaScript at some point and Perl around late â90s. Starting doing C++. I discovered the programming competitions when I was in high school and discovered I was actually pretty good at it.
Turner Novak:
What is a programming competition? Especially back in sounds like probably late â90s, early 2000s?
Erik Bernhardsson:
This was, yeah, early 2000s. Yeah, you typically had like five hours to solve a bunch of problems. Thereâs slightly different formats. Sometimes itâs solo teams, sometimes itâs like three people. Just a bunch of different formats, but the core idea is you have to solve, I donât know, half a dozen to a dozen problem in a few hours, or even sometimes just three problems. Itâs usually like very algorithmic problems. Theyâre very artificial, theyâre super silly. Itâs always like, âThere are N cows on a farm and thereâs a graph that describes their movements.â Itâs stupid, but theyâre fun. I mean, theyâre like fun puzzles, and I donât know, I discovered I was pretty good at it. I actually went to this international... Itâs called IOI, International Olympiad in Informatics. It was my first time to the U.S. I went to Wisconsin.
Turner Novak:
Wow, the pride of the U.S., like the-
Erik Bernhardsson:
Yeah, yeah. Yeah, yeah, yeah. Americaâs Dairyland.
Turner Novak:
Yeah, yeah, exactly, in the Midwest. Itâs like right in the middle there.
Erik Bernhardsson:
Yeah, itâs funny. That was the first thing I saw from the U.S. was Milwaukee and something like Kenosha. This is Chicago. Anyway, so I did pretty well in that, and I kept doing that throughout college, too. At university, there was a bunch of other people doing it, too, and then they joined this random startup called Spotify. I was like, âWhat are they doing? Like music? Music streaming? I donât think thatâs ever going to work.â
Turner Novak:
Yes, isnât it illegal? It was illegal at the time, right?
Erik Bernhardsson:
Yeah. I didnât think itâd make any sense, but all of my smart friends from school joined that company, so I was like, âOkay, I guess Iâll join it, too.â Obviously now, itâs like whatever, Europeâs biggest startup or something, but again, I think of that as like so random. I kind of ended up there when it was about 40 people, I think, and ended up staying for about seven years, so that was I lucked out.
Turner Novak:
Yeah, and you did your masterâs thesis on music recommendation. Was that like-
Erik Bernhardsson:
I actually did it at Spotify.
Turner Novak:
Oh, you did it at Spotify? Okay.
Erik Bernhardsson:
I did it at Spotify, yeah. Yeah, and by the way, I didnât know anything about music recommendations. I was able to convince them. Iâm like, âHey, can I just come and do some machine learning?â I didnât know anything about machine learning either, but theyâre like, âYeah, cool. You seem smart.â
Turner Novak:
Itâs just a math problem right at the end of the day.
Erik Bernhardsson:
Yeah, a lot of math problems. Itâs like a whole stack, 20 math problems together. I did that and I chose Spotify full-time early 2009. No wait, whatever, something like that, 2000... yeah. Initially, I kind of ended up working business intelligence, which is actually kind of in hindsight, Iâm actually very happy about because they were like, âYeah, music recommendations are cool, but now we have all these more pressing problems. Investors are asking for all these numbers. Actually, Erik, you know numbers. Can you do that?â I ended up spending a year just crunching numbers and just generating charts for investor decks and stuff like that, but that ended up being super valuable like 10 years later when I was running a startup.
Turner Novak:
Oh, Iâve done this before. Yeah.
Erik Bernhardsson:
Yeah, so itâs kind of fun. Anyways, but eventually at Spotify, ended up going back to music recommendation systems and built up a team around that.
Turner Novak:
Whatâs the most interesting thing or surprising thing about music recommendation systems that people might not know? When I think of Spotify, I think of the Discover playlist. Sometimes I find new stuff from them. How do they work? What would people want to know about that?
Erik Bernhardsson:
I think that maybe people get a little bit surprised, maybe not, is that the recommendation systems have no idea what theyâre crunching. They just look at a bunch of correlations. Basically, I ended up using a lot of language models, but not in the way you might think. The language models know nothing about music or lyrics or anything like that. You just think of individual songs as words and then you just look at the sequence of words basically, sequence of songs that people listen to. Theyâre just trying to find patterns in that. Essentially it was like a next token prediction problem where the words are like individual songs, but the songs are just IDs. Theyâre just in a big matrix. Theyâre just like entries in a big matrix essentially. So itâs called collaborative filtering. Basically this idea that youâre just mining this data set for correlations. And then basically how it works in practice, itâs like you create a big matrix and you factorize its matrix and then you end up with vectors. So you get vectors for the songs and vectors for users and playlists and artists and other stuff.
And then I ended up, I actually built a vector database at Spotify too, open source it for a while. It was actually quite widely used. Briefly used by Twitter and all these other companies. Since then you have to do vector search, but thatâs the core idea. You put it in a matrix, you factorize that matrix, then you get vectors, and then you do nearest neighbor search in that high dimensional vector space.
Turner Novak:
I think I read a blog post about how Spotifyâs music rec worked a while back, and maybe it was a different version of this, but it was something like... It basically just looks at who do you have a similar listening history or habit to, and then what are they listening to?
Erik Bernhardsson:
Thatâs roughly right. But you do it in a more statistical, less direct way. It doesnât exactly do that, but as a mental model roughly.
Turner Novak:
And then what did you learn about just building a culture at Spotify?
Erik Bernhardsson:
I think it was super formative. Spotify in the early days, there was zero management. At least for me. I didnât even know who my manager was. I think that was actually bad.
Turner Novak:
Oh, geez. That could be bad, yeah.
Erik Bernhardsson:
For two years no one told me who was my manager. That was my first job, so I didnât know what to expect. But I think it was actually, it ended up working out because there was a lot of smart people just working on hard problems and people knew what we needed to build and we didnât need a lot of structure. So I always took that with me. I think most startups, if you just have smart people who are pretty autonomous and have a mission, you need a lot less structure than you think. You can get by. People should just come in every morning and just ask themselves, âWhat can I do for the business today?â Thatâs my platonic ideal for what a startup should be. This is a little naive maybe, but aspirational maybe. You can get pretty close to that actually.
Turner Novak:
Yeah. Well, itâs kind of like the whole, âWhat did you get done this week,â thing. Itâs just like, âDid you actually accomplish anything? Go back, look. Did anything change? Did you materially make a difference on anything? If not, maybe revisit how youâre spending your time.â
Erik Bernhardsson:
Except with that phrasing feels like kind of negative. Youâre like, âWhat did you do today?â It almost seems like youâre expecting, âDid you actually do anything?â
Turner Novak:
So what do you think is the correct phrasing? Itâs just go in and get shit done. Is that the right way?
Erik Bernhardsson:
What can I do for the business today? What are you waiting for? Just do it.
Turner Novak:
Have you learned anything then from the Spotify days related to Modal?
Erik Bernhardsson:
A million things. How to build software, how to hire, how to... I think another thing Spotify did really well, itâs kind of random, but I think everything with Spotify did everything. But it had high taste. I feel like the design, the websites, how we talked about the product, itâs a certain level of quality. I remember the first time you tried the product, you clicked on a track and youâre like, âWait a minute, this is in the cloud.â Thatâs a magic experience of like, âWhoa, this is actually...â Thatâs actually something you get in Modal too. Like youâre running the code. Youâre like, âWait a minute. This code is running in the cloud?â And having that kind of magic moment and then the whole product and onboarding design around getting people to that magic moment and keeping the bar high for what the product looked like and the design.
Sometimes Spotify lost that a little bit, but they eventually came back to it. Spotify was postured as a bigger company than they were, I felt like. And I think thatâs quite important. Visual language and how you talk about it and the identity and your voice, you got to seem like youâre a thousand person startup even though youâre 10.
Turner Novak:
Yeah, the design aspect reminds me of... Did you ever use Limewire or any of those-
Erik Bernhardsson:
Yeah, all of them.
Turner Novak:
... file download? You search a song and thereâd be a hundred versions of it and youâd download probably a couple and see which one sounded the best to keep. Versus with Spotify, itâs like you just know youâre getting the song and itâs going to be great. And thereâs not going to be some guy made a mixtape and he has his intro credits rolled over it, or sometimes itâs a different song.
Erik Bernhardsson:
I mean, people would put fake songs sometimes. Sometimes I download it and I listen to an album, but it turned out two of the songs were actually the wrong song. And then when you switch to Spotify, youâre like, âWait a minute, is that actually the real song?â That happened to me a lot because people would just pollute. Maybe the labels hit it just to undermine streaming or downloading.
Turner Novak:
Just to get people to keep buying CDs for an extra year or two. I think you moved over to the US. At some point while youâre at Spotify and then you helped start this company called Better.com.
Erik Bernhardsson:
Thatâs right, yeah.
Turner Novak:
Yeah. Real quick for people who donât know what Better.com is, what was that one?
Erik Bernhardsson:
It was a mortgage lending startup. I think at some point... I was at Spotify for seven years. I was just burnt out and wanted to try something new and I was like, âIâm going to try something completely different.â So I ended up starting as part of the founding team at Better, I think we were like 10 people when we started. But that was a total roller coaster. Weâre like 10,000 people when it left, six years, five years later, and then the company basically collapsed.
Turner Novak:
Itâs public now? They got it public at some point?
Erik Bernhardsson:
To the credit of the founder of Vishal, they were able to take it public a few years later and randomly, actually a couple weeks ago turned into almost like a meme stock. Stock went up 3x
Turner Novak:
Yeah. I was looking. It looks like Open Door and Better are both going vertical.
Erik Bernhardsson:
No, itâs so bizarre. Itâs so bizarre. So Iâm actually bullish on the company in a way. I do feel like thereâs a space for innovation in the mortgage lending. I think with interest rates high, itâs tough and thatâs ultimately a lot of what made the company almost collapse. But I think itâs still worthwhile. Thereâs a lot of stuff in the housing industry thatâs just fundamentally extremely broken in the US.
Turner Novak:
And itâs a very large percentage of GDP.
Erik Bernhardsson:
Itâs enormous. Itâs crazy.
Turner Novak:
And for most people, itâs the biggest purchase that theyâll ever make for 90, I donât know, 98%, 99% of Americans in most countries. Your house is your biggest asset. You work for a decade to save up, to make a down payment. You put it in a house, you pay off a loan, and then thatâs your retirement. It is everyoneâs fundamental. Just the value of their life thatâs accrued is the value of the home.
Erik Bernhardsson:
No, itâs crazy. Itâs so large. And so thereâs a lot of good reasons to think about this problem. Ultimately I think tech is just one part of it. And I know I was the CTO there for five years. I built a tech team. I took it from basically one person to 300 or something like that. I learned a lot about management and hiring and all this stuff. Took a lot of stuff with me to Modal. I donât know. Now that Iâm in Modal, Iâm like, âThis is what I want to do. This is so much more fun.â Better is fun too.
Turner Novak:
It seems like based on what I know about you, this might be a slightly more interesting problem for you to be working on, but still both important. And you mentioned you learned a bunch of stuff. Iâm assuming youâve interviewed thousands of people, hired hundreds. Maybe youâve hired a thousand people at this point.
Erik Bernhardsson:
Not a thousand. I think Iâve hired maybe 400, 300, 400 probably something like that.
Turner Novak:
Wow. Okay. So just thinking back over the past, I donât know, decade or two of this, what have been the biggest predictors of on the job success from hiring... Or sorry, from an interview? I know itâs extremely difficult, but you have a blog post on your blog thatâs basically you need to treat this as almost like a prediction problem of figuring out probabilities of what actually leads to success.
Erik Bernhardsson:
Exactly. Thatâs at its core of what it is, right? Youâre trying to predict their future performance. So I always feel like people are like... When theyâre criticizing interview processes, theyâre like, âWhy are you doing these weird brain teasers, whatever?â Who cares? Whatever predicts future performance, maybe brain teasers donât do that. And thatâs a valid criticism, but donât throw them out on the grounds that theyâre not realistic work, because thatâs actually besides the point.
So the real thing here is to predict future job performance. And so what actually predicts future job performance? I donât know. I think itâs important to be humble about it. The more I interview actually... I think thereâs a false sense of confidence in the early days of hiring people. I thought I knew, I was like, âYeah, this person is great. This person is not good.â Iâve just realized itâs very noisy. Itâs very noisy. Itâs very hard to predict.
But the things Iâve learned, honestly, the biggest thing is actually intelligence. And itâs boring, but itâs hard of assess. But when you work with a person for a long time, you generally observe that smartest people, theyâll always just do better. So thatâs a boring answer. I think beyond that, artificial intelligence, what else predicts it? I would say at startups at least, Iâve seen high agency commercial autonomous, like autonomy that tends to be quite important.
Turner Novak:
It seems like it would relate to your earlier point about just what can you do for the business to move it forward kind of a thing, find an opportunity, close a deal, make money, save money, insert whatever. Moving business forward, commercial objective.
Erik Bernhardsson:
Being a little impatient. I like people who want to just get stuff done. Theyâre high agency and they just get stuff done. So I think thatâs pretty important. Thereâs a lot of stuff like that. People are fundamentally positive. I hate when I interview people and theyâre complaining about stuff. I think thatâs a bad, negative trait. I donât know. Thereâs a lot of stuff like that. I often ask just almost seemingly random questions to people and trying to see how well it holds up.
Turner Novak:
So whatâs the point? What would be an example of that? Ask me one.
Erik Bernhardsson:
Whatâs the hardest thing you ever did? Tell me something you did in the last year that youâre proud of?
Turner Novak:
Me personally?
Erik Bernhardsson:
Yeah. Weâre role playing here.
Turner Novak:
I mean, I figured out how to start making money from this podcast, which is not easy.
Erik Bernhardsson:
And how do you make money?
Turner Novak:
Sell sponsorships in the podcast. I mean, itâs super simple, but also-
Erik Bernhardsson:
How do you sell sponsorships? How have you learned about sourcing sponsorships and winning those deals?
Turner Novak:
Itâs actually a lot easier than I thought it would be. Itâs just a volume thing. You just got to be like, âHey, I have this podcast a lot of people listen to. Hereâs the data.â
Erik Bernhardsson:
Okay. Whatâs the conversion rate when you reach out to people? Whatâs the percentage of sponsors? Anyway, my point is I just ask questions like that and I just go super deep.
Turner Novak:
Some people like that though. Some people hate it.
Erik Bernhardsson:
The reason why I like that question is smart people, theyâre able to go almost arbitrarily deep. But when people are starting to like, âThis didnât really hold up,â thatâs when I get a little hesitant. It turns out actually they donât know what they built themselves or... Thereâs a lot of stuff like that you can do. I donât know. But fundamentally interviewing is hard.
Turner Novak:
Yeah. I actually want to finish answering that question because youâll see why in a second. What I found is that the people that are most likely to convert on sponsorships in the podcast are people who follow me on Twitter or listen to the podcast. They already know. When I bring it up to them, theyâre like, âOh yeah, Iâm in.â Thereâs not really a big back and forth and you have to sell them pictures just like theyâve been listening to it or theyâve been following you. So if youâre listening right now and you want to sponsor the podcast, respond to the email, comment, DM me on Twitter, whatever you need to do.
Erik Bernhardsson:
By the way, your answer was a great interview question answer. It made me seem like youâve thought deeply about the problem and you observe patterns and you come to realize something, right?
Turner Novak:
Yeah. I think just generally sometimes just trying to go out of your comfort zone is so hard sometimes. For me personally, I just never done it before and thereâs somewhat of a fear of failure and you just have to do it. And then once you do it, youâre like, âOh, this is actually way easier than I thought.â I emailed someone. We get on a call and they give me 20 grand. This is pretty cool. That feels pretty good. It was an extremely high ROI.
So then it motivates you to do it more. So I think itâs like for any hard thing, itâs figuring out what the dopamine unlock is or what the part of it that you like, and just figuring out how to get to that point. You have to do the hard stuff. I mean, thereâs every job that you have. One of my first bosses told me this. I worked at a private equity firm and as an intern and my job was literally to cold call investment banks and be like, âHey, do you have any deal flow for us?â
It was the stupidest thing. It goes the 0% conversion rate. People would be like, âWhy is this 20-year-old kid calling me asking for companies to invest in? This is not how it works.â
Erik Bernhardsson:
Sales is like that for me, by the way. I have to do a lot of it. By the way, thatâs why Mormons are so good at sales. I love Mormons because they have to go for three years to Europe.
Turner Novak:
And just try to convert you to their religion. Thatâs a little different than most other religions.
Erik Bernhardsson:
I admire that so much. So many people saying no to them and they just learn to like, okay.
Turner Novak:
Yeah. I feel like thatâs the thing is when youâre selling anything, nothing is like a hard no. And if youâre just getting hard nos, youâre almost doing it wrong. You think of it more of a discovery process or like, âWhatâs their problem? Can you solve it?â And in some cases they might just not have your problem. If Iâm a restaurant owner and itâs like some one-off, non-shame, I sell Indian food, and you come to me and try to sell me Modal, Iâll be like, âNo, you shouldnât even be talking to me.â Itâs just about figuring out who actually has the problem and then the bigger the problem is and the more relevant, thatâs who you should be fishing for.
Erik Bernhardsson:
I think itâs like youâre bringing up a really interesting point about starting companies by the way. I think a lot of founders make this mistake. They go and talk to a lot of potential customers and then try to land those as design partners and people are like, âYeah, this is not quite what we need, but if you do this thing, then maybe we would do it.â If you support on-prem, we would consider it.
And then people are like, âOh, on-prem. Okay, so we need that as our MVP.â And people end up with this ridiculous MVP. Like you should disqualify a lot of people when youâre talking to potential customers.
Turner Novak:
Really. So when youâre citizen, maybe like a good transition into Modal, but so when you were starting it, did you get that a lot?
Erik Bernhardsson:
No. I mean, I actually took a terrible approach in hindsight. I decided Iâm just going to take my own experience and just sit for two years and build a product that I wanted to have. Thatâs very different approach.
Turner Novak:
But you probably tuned out a lot of noise of somebody mightâve given you feedback like, âOh, we need on-prem.â And youâre like, âItâs not... We donât need that.â
Erik Bernhardsson:
Screw that.
Turner Novak:
So itâs almost like you were your own customer really at the end of the day. You were building a product?
Erik Bernhardsson:
Exactly right. And I think thatâs maybe different because I maybe started my company a little bit later in my career. I felt like at that point Iâd already... I spent enough years on the application side that I knew what I wanted from the infra side. But I think for most people itâs probably better to spend a bunch of times talking to potential customers.
Turner Novak:
So then I know you described it as you were just heads down building for a year or two. What was that like? Were you trying to find market pull or were you literally 16 hours a day, donât talk to anyone, just building it and then you got to the end of the tunnel?
Erik Bernhardsson:
Maybe not 16 because I have kids in a wife and stuff, but-
Turner Novak:
Probably like 12, right?
Erik Bernhardsson:
Yeah, something like that, right? Letâs also remember, this was during the ZIRP era, late pandemic. So we took a little bit of seed money and people were kind of like, âYeah, it was cool. Youâre building some hard stuff.â But that being said, I do feel like thereâs a little bit of sometimes disconnect where I think if youâre starting a startup, you should pick a hard problem. You should almost pick something that takes a year to two... I donât know. Thatâs the challenge I have with when I look at a lot of YC startups. Theyâre pushing them so hard to find PMF.
Turner Novak:
Instantly. Like within a week or two.
Erik Bernhardsson:
Iâm like, âThat doesnât happen for both startups.â And so I think it leads to a certain amount of startups, certain type of startups, but thatâs not the startup that I wanted to start. I wanted to build something that takes a couple of years to build, because then I know thereâs a little bit of a moat also. Thereâs something more durable there.
Turner Novak:
Because if you built Modal like, âOh, that seems cool. Iâm going to build it too.â Iâm going to do it, vibe code it in a night and then start stealing your customers. So you got a year or two of just building this thing and it was stable diffusion. It was the first thing where it really started to take off.
Erik Bernhardsson:
Thatâs right. We targeted a kind of... We always have this very general purpose approach. Weâre like, âWeâre going to build something that works for everyone that builds stuff with machine learning data and AI.â And we werenât really sure. âWell, whatâs the killer app?â And that turned out to be Stable Diffusion, which I think was somewhat serendipitous because Stable Diffusion wasnât around when we started the company. So we ended up being a little lucky in that sense.
I think we wouldâve found something else, but I donât think revenue wouldâve been where itâs today if it hadnât been for all the gen AI stuff. So we were a little bit lucky in that sense.
Turner Novak:
So then how has the product evolved over time? Was there a specific first thing that you built that people were using, or how did you light her up?
Erik Bernhardsson:
We started out with this very general purpose product. Like I said, we didnât really know what was the killer app. And then Stable Diffusion ended up being that, and then that pulled us super hard in the direction of inference and specifically gen AI inference. But in a way, weâve gone full circle because I think in 2025 roughly, we came back to the idea of, âNo, this is a general purpose platform.â We also have way more engineers now so we can-
Turner Novak:
Build more stuff.
Erik Bernhardsson:
Yeah. I mean, when you have five engineers and you find product market fit, you donât have a choice. Just chase that shit. Just go for it, right? Now we have, I donât know, 45 engineers or something like that. Now we can actually afford to also build a little bit more speculative stuff. So we are very much going back to this original idea of like, âNo, weâre trying to build a general purpose platform. Weâre trying to build an infrastructure layer.â And thereâs so many other parts of that. Weâre focusing a lot on training. We released a notebooks product recently, scan boxes, like I said.
Turner Novak:
Which notebooks?
Erik Bernhardsson:
Basically the idea of like Google Colab or Jupiter or things like that. Itâs more like you have a graphical instant feedback thing in your web browser where you can just write some code and plot stuff and run code, and get more instantaneous feedback from inside the web browser. And you can share that with other people. Itâs not really necessarily meant for production use cases, but itâs meant for more interactive exploratory stuff.
Turner Novak:
Like testing things?
Erik Bernhardsson:
Like playing around and trying out quick and dirty prototypes and plotting stuff.
Turner Novak:
Interesting. So then as you were building the product, I know you mentioned earlier making it super simple in some cases, but also infrastructure is hard. You also mentioned all this stuff is hard. How did you decide when to make the product super simple and abstract things away, but then also kind of keep it complex when it needs to be complex? Was there any kind of secret to that?
Erik Bernhardsson:
I donât like the way simple. I also donât like the word democratize. Thereâs a bunch of words I try to avoid when I talk about the product because I think a challenge with a lot of infra products is they take away a lot of the possibilities. Basically, you want to reduce the burden on the developer by 90%. You want to collapse the amount of effort it takes to build these things, but if youâre also removing the ability to build stuff by the same amount of percentage, then thereâs no-
Turner Novak:
Itâs not worth it almost.
Erik Bernhardsson:
Yeah. So I think you want to retain, I donât know, 90% of the possibility space while reducing the complexity by 90%. So I donât necessarily think of it as simplification. I think of it as finding the right abstractions that enable people to build things faster. We always focus very intentionally on... I think it was high code use cases. We go to the power users, the machine learning engineers, the people running the models, training the models, running really complex stuff. And we look at what theyâre doing and we want all of that to run at Modal, but we want it also to take a lot less effort.
So itâs hard to find those abstractions. I think in a way, by sacrificing a lot of the backend workloads, that helped us focus a lot more on AI. I think itâs very important that, at least for me or how we think about Modal is to pick a persona and a problem space and think about, âOkay, what exists in that problem space and builds for that.â And when you pick a problem space, then I think it gets a lot easier to make those simplifications. I donât like the word. This is, by the way, a challenge with the cloud providers. The cloud providers, by the way, the hyperscalers, theyâre trying to build for everyone at the same time. And if you build for everyone at the same time, itâs actually really hard to optimize for any particular audience.
Turner Novak:
So it sounds kind of like to your point, when youâre at Spotify, itâs like you spent so much time just setting things up and building the infrastructure. So itâs making it so they donât have to spend set up time to customize an environment or a production run or a training run, whatever the thing is.
Erik Bernhardsson:
Exactly.
Turner Novak:
So they can still do specific unique, hard to do things, but just abstracting away the manual boring, less creative aspect of it or something like that.
Erik Bernhardsson:
Keep the power, but reduce all the garbage.
Turner Novak:
So then was there certain points on the roadmap? Are you getting a lot of feedback from customers? How do you decide when to listen to them, what to build? Has that been a big thing youâve had to figure out, or is it still just like you guys know?
Erik Bernhardsson:
Now we talk about it. Once we emerged from that cave, then we started talking to customers all the time. And we have a big Slack community.
Turner Novak:
How do you decide what to prioritize?
Erik Bernhardsson:
Itâs very hard. I think when you talk to customers, you should take their pain points extremely seriously. But when they suggest solutions, you should almost ignore that.
Turner Novak:
Iâve heard that before, yeah.
Erik Bernhardsson:
Because customers will be like, âOh yeah, can you do this feature for us? Weâd really want that.â And then you have to ask them, âWhat is the problem here?â And then go back a few steps. And then you identify, âOkay. Theyâre having an issue with this.â And then you have to think for yourself. Maybe thereâs a much better, more general solution actually. And then the second problem is prioritization. By the way, prioritization is just incredibly hard.
I think thereâs no way around it because we have 45 engineers right now. Of course we can do more stuff, but by the way, if I had a thousand engineers, I would still struggle with... I would want to build a moon rocket and whatever, space laser and all these other things. It just never ends. So you just have to be super ruthless about picking the two to three things that are existential for the company and just that is probably where you should spend time at any point in time.
Turner Novak:
And you mentioned that just the rise of the importance of inference changed a lot. What kind of happened there? What were people starting to use it for, and then how did that impact the company?
Erik Bernhardsson:
Yeah. So all these generative media models and other things like large language models and things, and I think a lot of the focus traditionally on AI, at least in the last five years or so, has been on training. But as people put these models into production, they need inference. Inference is a little bit harder because unlike training where you actually find like having a fixed amount of compute resources, like a static amount of GPUs, but when youâre doing inference, you canât fundamentally predict how much. I mean, you can estimate, but you have very volatile demand.
So one of the value proposition at Modal, part of the reason why I think we found a good market is that in inference is that we solved that for you. And so you donât necessarily have to think about capacity. And there was a sort of realization, thereâs a fast-growing market here in inference, which is why we invested so much in the last few years in inference. We still invest, by the way. Thereâs so much still to be done. But thatâs why I think inference has been a good space for us.
Turner Novak:
And is this a good way to think about inference? Training is just like your brain and then inference is like using your brain or something, or like itâs doing things?
Erik Bernhardsson:
I think thatâs right, yeah.
Turner Novak:
Okay, cool. Iâm always trying to get a better understanding. Plus, also, sometimes people will reply to the emails I send. Then theyâll be like, âWhat does this thing mean?â And it was a basic concept in the conversation. I was like, âOh, fuck, I probably should have explained that.â So to try it as we go and be like, âHey, this is actually what inference is, by the way, for people who donât know.â And so then when you just think about the future of Modal and AI development and all this stuff, where is it going over the next couple years, in 10 years, 20 years? I donât know if thatâs even possible to predict, but how do you think about how this stuff is? Yeah, howâs it all evolving?
Erik Bernhardsson:
I always thought of Modal as my 20-year project, by the way. So for me, I always make ... I have a lot of ideas. I donât know how itâs all going to unfold. I think of there are so many things around AI development or more broadly working with just compute resource-intensive problems that kind of goes beyond just training and inference. Thereâs entire life cycle.
And when you start to think about that, thereâs many other things. Thereâs data pre-processing, which we actually have some customers doing. Thereâs training, thereâs inference, but thereâs also other things. A lot of observability. Thereâs features, scores, thereâs real-time streaming data pipelines. Eventually, if you want to start working on more like data stuff, thereâs a lot of things around query engines, working with more databases, more like primitives for structured data.
So I think when I look down the road three to five years, my goal is if youâre building an AI application, you can do the entire thing on Modal, like the entire thing. We have all the software tools you need, and we do it, by the way, much better than anyone else because we control the underlying infrastructure and itâs all super well integrated too. So thatâs my hand-wavy addition for what I want to get to.
Turner Novak:
So itâs sort of like traditionally, if youâre not using Modal, thereâs almost like multiple tools you have to stitch together and thereâs trade-offs between how things sync versus Modal, youâre almost like the full stack of it. So, you will get a better, more succinct, faster experience. Are there any ways you might fall short with that approach that someone can provide a different or a product that might be considered better if somebody wants everything, like cost or something like that?
Erik Bernhardsson:
I mean, I think cost is always going to be ... We try not to compete too hard on cost. And thatâs always fundamentally, I think if people want to run something at absurd scale and theyâre willing to do all the hard work that it takes to get the absolutely lowest cost, probably theyâre going to be able to go buy their own data center and whatever, buy their own power plant. If people are willing to spend an enormous amount of effort, they can always lower cost.
So we try to strike a happy medium where most companies are probably way better off using Modal because they can just iterate so much faster. And I think ultimately, thatâs what people care about. Maybe it gets 30% more expensive with Modal, but maybe thatâs fine, because the engineers are way more happy, they can ship things to production much faster. In a way, weâve been in this weird paradigm in the last 5, 10 years with AI, which I donât love where I call it the carbon to silicon cost ratio. The chips are more expensive than the humans. The humans are the carbon and the chips are the silicon.
Turner Novak:
Okay, yeah, that makes sense.
Erik Bernhardsson:
Right? Right now, thatâs the challenge. So what it means when the humans are more ... The chips are more expensive than the humans. Thereâs no consideration paid to how much does it take an engineer. But when itâs the other way around, when the labor is more expensive than the silicon, then yeah, take more chips, whatever, we can run things at a little bit less efficiency, letâs use a higher level language, whatever in order to make engineers move faster.
And I think weâve been a little bit too far to the extreme of optimizing for GPU efficiency in the last few years. And I think as we make engineers more important again, I think thereâs going to be a lot of premium people are going to want to pay for just the ability to move fast and have good developer experience and shipping stuff into production.
Turner Novak:
Yeah, Iâm definitely curious on your opinion on that. But I feel like one thing I learned, I took a business strategy class, it was my last semester of college or whatever, and my biggest takeaway was you basically have to figure out, do you compete on a different, unique, better product? Or do you compete on cost? Are you the lowest cost provider? Itâs probably the shittiest product honestly, maybe not. But are you the cheapest, or are you the most expensive because you are the best?
And itâs like if you get caught in the middle, you might say something like Walmart would be low cost. Itâs like you go in, you know what youâre going to get, itâs not fancy, but itâs cheap. And on the high end, itâs like CanadaGoose or something. Theyâre the $1,000 jackets, right? Itâs like just premium, the most expensive, but itâs so warm, so comfy. And then in the middle is JCPenney or Kohlâs where itâs like they kind of have fancy things, but they donât have the most scale, so theyâre not really the lowest cost and it just kind of sucks and no one goes there, so you die basically.
Erik Bernhardsson:
Yeah, yeah. But look at another market. Look at Android versus iOS.
Turner Novak:
Yeah, exactly.
Erik Bernhardsson:
Android has kind of found a niche in low cost, but Apple is conquering everything else and making many more money, by the way, than Android. So I donât know, I think depending on which market you look at, sometimes thereâs 25 players in retail, but in technology it has to be one or two players. And I donât think itâs necessarily evident that being low cost or being high cost is the best model. I think in iPhones, that actually goes pretty far down the spectrum and they make way more money. But of course, you really, really want to save money, youâre going to go buy an Android.
Turner Novak:
Yeah. Well, and I think too, if you look at iOS, I feel like they almost continue to move up market in the sense of the phone just continues to get more expensive, but then also I feel like theyâre slowly creeping down into the lower end of just more and more people, just as a percentage of humans are buying iOS devices. And then also you have tablets, you have the watches, you also have the laptops like MacBooks.
So I think I saw that. This is from a couple of years ago. I think they have over a billion devices active around the world. So theyâre probably just the top 10%, 20% of North America, Europe, Australia, Japan, South Korea, those most developed highest GDP per capita nations. And those people have money to spend on games and in-app purchases and Apple cards. However, Apple is making money, itâs from all those top 20% of income earners.
So then how did you think when you were building the brand for Modal, I know that was a pretty important sort of intentional thing youâve done over the past couple of years, but how do you just generally think about as a startup, maybe specifically as a developer platform? How do you think about building a brand and your go-to market?
Erik Bernhardsson:
In certain ways, it was very intentional. In certain ways, it was not at all. We havenât had a marketing team or a sales team until very recently. So a lot of basically our early customer acquisition was me tweeting share and talking about the product. Itâs like founder-led marketing I call it, and it sort of works. And I think we ended up being very PLG, almost like unintentionally without me necessarily thinking deeply about it, because I always built consumer products.
I think today, there are such a lost opportunity for so many companies to have just a boring brand, like an enterprise infrastructure and SaaS and stuff like that. My fear is the era of boring brands is over. People want to buy a product that makes them feel cool, that makes them feel like theyâre working with something that the cool kids use. So I think itâs such a lost opportunity to not be out there and have a voice, and have an identity, and be a little quirky, and carve out your own niche, and how you talk about the product.
I donât know, we always try to have ... Weâre trying to be approachable but also try to be a little cool. And I think itâs important as we build out a brand to really think about that. I donât know, thereâs so many products out there, especially in the enterprise SaaS that are so boring. You go to the website, itâs download the white paper.
Turner Novak:
Book a demo with a sales team.
Erik Bernhardsson:
And I think itâs not the way that these products are ... I mean, and I donât think weâre the only one to notice that, I think there are some really exciting brands out there. Vercel is a great one. PlanetScale, I love what theyâre doing, very crazy website. There are these new generation of products that are just taking a very different approach to branding and marketing and I love that.
Turner Novak:
Oh yeah, Iâm on the PlanetScale website. It almost looks like it borrows a little bit from MS-DOS.
Erik Bernhardsson:
Exactly, but thatâs the point, right? Theyâre speaking to developers. Itâs very clear when you go to that website, itâs like, âOh, cool, this is different.â I donât know. I mean, itâs weird, but I think it appeals to developers. Developers are very finicky. Itâs hard to sell to developers, and maybe thatâs also a different buying pattern. Maybe 10 years ago it was more like CTOs and CISOs and whatever buying the products. But today, developers paying for the product and they want the cool product and they want to work with the product that makes them feel cool. And so I think itâs a different way to sell software today.
Turner Novak:
I feel like developers are just smart too. And if your product sucks, theyâre not going to use it, because they have a choice. They can be like, âIâm not going to use this shitty one that I canât even integrate properly. Just give me the one that just works, because I got shit to do.â
Erik Bernhardsson:
Yeah, exactly. Exactly. And I think to see through marketing fluff, I always push ... When we write up content, high performance, blah-blah-blah. Iâm like, âWhat does the high performance mean?â As a developer, youâre like, unless thereâs a very specific meaning, donât write the word because it just sounds like fluffy.
Turner Novak:
So then if Iâm just getting started on, I donât know, building in public, tweeting things out, building my marketing strategy for PLG, bottoms up like that, how would you recommend getting started? I know you kind of blogged for a while and tweeted for a while before starting the company. Is that probably the best way to go, like just start today, start five years before you start the company?
Erik Bernhardsson:
Yeah. I mean, I think so. I think everyone has to find their own journey. I think everyoneâs go-to-market journey. Every company is going to have their own idiosyncratic thing. And when you listen to podcasts of founders saying, âOh, we discovered that influencers or we discovered whatever.â You always have to discount that a little bit, because I think everyone has to find their own kind of journey.
For us, it just ended up being leaning into the stuff weâre good at, which is I already had a distribution channel through my blogging and tweeting. So if youâre asking for yourself ... I mean, if you have a lot of Twitter followers, I would absolutely do that. If you have a lot of following online, you should absolutely lean into that. If you donât, I mean, thereâs other stuff. Some other companies have found very different pathways to user acquisition.
Turner Novak:
Well, itâs probably also what are you good at. If youâre a great public speaker, just go speak at some conferences and wow the crowd and get them to sign up. Or if you like podcasting, start a podcast, go on some podcasts. If you hate podcasting, donât do that. Just whatever youâre good at, itâs probably the way to think about it, because itâs super hard. I think thatâs the main thing if youâre thinking about starting something. Itâs like itâs probably going to be 100 times harder than you even can even conceptualize. So you better make sure that you enjoy it and that youâre good at it, because it just increases your chances of success if you can stick with it and you pick stuff that youâre actually, you have an advantage in versus everybody else.
Erik Bernhardsson:
Yeah. And especially if you enjoy it. Iâm always like shit posting anyway, so I might as well write about Modal.
Turner Novak:
And you mentioned you guys had raised some money. I think you mentioned it was ZURB, sort of. I think that was probably the first round. What sort of fundraising journey ... I know you announced a series B by the time this comes out, maybe a couple weeks, a month before this episode came out, but whatâs fundraising been like? And any just general advice for people thinking about raising money?
Erik Bernhardsson:
We raised the seed round. Actually, they raised two seed rounds back in 2022 and then the A round. The seed rounds was primarily from Amplify. Then the A round was from Redpoint, that was 2023, and then we didnât race for almost two years. We did a small safe round in between, but then we did a series B recently with Alexis leading.
I donât know, me personally, I like to keep things lean. I donât like the overhang of a lot of capital. I actually feel like I work better with little capital and forcing myself to really use that capital well to get to the next round. So Iâm not a big fan of these mega rounds that have happened recently. And I will say though, to some extent, itâs also function of the market as the market is heating up, we felt that at some point you just have to play the game the market is playing.
Turner Novak:
Yeah. You need some amount of capital if everyone else has hundreds of millions or billions. Yeah.
Erik Bernhardsson:
Exactly. Yeah. So to some extent, itâs like you canât choose your game. You can play the game by its own rules and you have to figure out the rules and play it by them. But at the end of the day, if competitors are out there raising a lot of money, you kind of have to do that too. Itâs just the market heating up. You go from this blue ocean to red ocean or whatever, things just heat up. Itâs a different game. Early in the creation of a company, typically if youâre building something new, there is no category and donât raise a ton of money.
Turner Novak:
Is it harder though to raise if something doesnât exist? Youâre like, âHey, weâre building this thing.â And thereâs no comps for people to compare it to, so you kind of got to sell the story to them about why this is going to be a massive, $100 billion company in the future.
Erik Bernhardsson:
Probably, but I mean today, you can just say itâs AI, right? Yes, we probably had the benefit of getting credibility as a team. I think in the early days, Modal raised money. Yes, we didnât go and raise money without any sort of experience. I donât know, I was in my late 30s when I raised money and I think people looked at my experience and, yeah, it seems good.
Turner Novak:
Yeah, why not? It seems like he might know what heâs talking about.
Erik Bernhardsson:
There is a little bit of that. My point is itâs probably easier for me to say, but sure. I think for most people, maybe you need to have a little bit more traction. And maybe if youâre in your early 20s and have zero experience, maybe YC is good for you because it forces you to find product market fit very early. And thatâs always easier in a way to raise money based on. When youâre building something hard, that takes several years. I think investors need a little bit more of a belief in the team and the technology in order to put in money.
Turner Novak:
Yeah. I mean, I think a very under-discussed reason of startup failure is that you just kind of lose faith or give up. Youâre saying, âI donât really want to work on this anymore.â And if something super hard and it takes you two, three, four years, thatâs a hard thing to underwrite for the investors of will they even stick with it to even build the thing? And then it actually starts of building the company around the thing that they build.
Erik Bernhardsson:
Yeah, itâs also hard to find a problem thatâs hard but not impossible and knowing that.
Turner Novak:
Well, I think something you mentioned the other day when we were talking, you said that you think probably too many people start companies in their 20s and they donât start them enough in their 30s or 40s. I donât know if I got that quote right.
Erik Bernhardsson:
I think something like that. I donât know, maybe itâs just me being sad that I didnât start a company earlier in my life, but Iâm very happy I started a company, but I was like 37 when I started Modal. But on the other hand, starting a company when I was 37, in a way, itâs been a lot easier, because I know how to hire, I know what good looks like, I know how to build products, I know how to manage tech teams and how to ...
A lot of stuff is so much easier. I donât necessarily need to learn those things on the job. Some stuff I have to learn on the job. I donât know, my feeling is thereâs a lot of people out there kind of like me when I was in my late mid to late 30s. Theyâre just really smart and really good at what theyâre doing and they should just quit their jobs and start a company.
And the only reason why I did that was for a while, my previous company had a tiny bit of an exit. I was able to sell some stock and take a year off and jam on stuff. So I really wish, I think thereâs a lot of people out there in their 30s who if they had a little bit of money, maybe they would take the risk and do that. And I think that would be awesome for society because I think thereâs so many smart people out there with great ideas and we just need to push them a little bit.
Turner Novak:
Is it harder though when youâre married with kids? I donât know how many kids you said you have, but yeah. Because personally-
Erik Bernhardsson:
I think it wouldâve been hard to start company and my kids were newborn. So when I started Modal, my youngest had just turned two or three and she just started sleeping, and I think that was a massive difference. My IQ went back up.
Turner Novak:
Just like 10x increase, yeah, productivity.
Erik Bernhardsson:
Yeah, no, it was terrible for a while. They werenât sleeping, but now, itâs good.
Turner Novak:
Yeah. Well, speaking about being more productive, I know you have a really strong opinion on how to make developers more productive. Whatâs your general thinking on developer productivity?
Erik Bernhardsson:
I think itâs super hard to measure. Although actually, one thing I think is lines of code is actually not a terrible metric. Thatâs something I need to write more about.
Turner Novak:
That was my favorite thing from the Twitter takeover with Elon. Heâs just like, âPrint off the code.â And people are making fun of it, but itâs like how big is the stack? Are you actually writing code, or is there nothing there?
Erik Bernhardsson:
I think thatâs right, but I also think the reason why it worked is like suspected median lines of code was zero. I think thereâs a lot of people exactly who actually didnât write code at all. By the way, thatâs actually something I think culture has shifted a little bit. Sorry, weâre getting on a side track here, but that managers write code. I think for the last maybe five years ago, that wasnât necessarily evident, but now I think itâs shifted back, and thereâs more of an expectation that, yeah, managers should also write code. I still write code by the way at Modal.
But anyway, going back to develop productivity, like I said, itâs hard to measure. The only thing I think gets close to it is ... Or the thing that Iâve empirically learned and gets to it is, how fast are your feedback loops? And by that, I mean if you have an idea ... Or maybe even a better example is if you know that thereâs a bug on your website and itâs an obvious fix, how long does that take to fix?
And I think itâs incredibly dysfunctional if you canât get it out in a couple of minutes or an hour. Active time, if you know what the fix is, of course, just fix it and just deploy it. And sure, maybe it takes half an hour to run a CI, CD test suite or whatever, get it out. But the way a lot of companies operate, itâs like you write up a linear ticket or whatever, Jira, and then a product manager has to prioritize it and then it goes back to some other engineer, and then they have to write something, and then they send over to some other one for code review and then it takes a day or two for that person to review the code and then it goes back to me and then I merge it. And then three days later, thereâs a deploy. And by now, youâre talking two weeks of work. Thatâs absurd. Maybe not two weeks of active work, but adding up all those steps.
Turner Novak:
Time. Yeah, all the different passes off between people. I mean, you just mentioned thereâs three or four different people that get involved to fix a bug.
Erik Bernhardsson:
Itâs insane. So all these handover points, it just adds tax to all these steps and thatâs especially for small features that tax becomes now like 1000%. And so I think itâs very important to reduce that tax on every step and make it possible for engineers to get obvious code out in production as fast as possible.
Turner Novak:
So one thing you mentioned was code reviews. I was talking to your co-founder, Akshat. Is that how you say it?
Erik Bernhardsson:
Akshat, yes, right.
Turner Novak:
Akshat, yeah. Heâs like, âYou got to ask him about code reviews.â I have no idea what youâre going to say, but whatâs your opinion on code reviews based on what you just said?
Erik Bernhardsson:
Weâre not anti-code reviews. First of all, I do want to point out a very common misconception is that people somehow think that code reviews are mandated by a bunch of compliance standards. That is not true. People think this, and thatâs something people like internalize. Theyâre like, âWell, obviously, we do code reviews, we have to because of compliance.â That is not true. You donât need to. As long as you have a well-defined business reason for every line of code that goes into production, as long as you have an audit log or why did you push this code, thatâs fine.
I think there are some benefits to code review, spreading knowledge, stuff like that. But what I think is an anti-pattern is people sometimes I think just use code reviews to absolve themselves of the responsibility if something breaks. Theyâre like, âOh, I donât want to merge this into production myself, because what if something breaks? Then I want to point this other person also.â That thing is super bad. And I actually think in practice, a lot of why people like code review is they feel like by doing the code review, they donât have full responsibility anymore.
Turner Novak:
And if you have multiple people involved, if thereâs three, four, five, six, the more people, the less individual unit of responsibility you have, and everyone can just say.
Erik Bernhardsson:
Yeah, I think thatâs very dangerous when people just use it as a way to get to that. I mean thereâs some benefits. Sometimes I go and change some code in a code base that I donât fully understand and I kind of want someone to take a look at it just to make sure I didnât screw something up. Or if you have an intern, theyâre writing a bunch of code, maybe the intern hasnât thought about all the security implications or whatever, you should probably have a senior person taking a look. I donât know, thereâs a lot of reasons why I think a lot of code review it gets internalized as this cargo cult process and I think thatâs quite bad.
Turner Novak:
I actually have, his name is Daksh Gupta at Greptile. I donât know if youâre familiar with Greptile. Theyâre an AI code review platform. Yeah, Iâm actually having him on the podcast, meet him in two weeks. So that will be coming out I think probably a week or two after this one does. So thatâll be interesting. Iâm interested to get his take on code review. Obviously, thatâs the whole business, but it kind of lies into the point of how AI is automating coding and obviously that case it might be a little bit more necessary to have some code review with AI. I donât know, whatâs your opinion on how thatâs kind of all changing?
Erik Bernhardsson:
Like AI coding?
Turner Novak:
Yeah, I mean, it sounds like a lot of stuff is getting automated.
Erik Bernhardsson:
A lot of stuff is getting automated. I donât know, I use Cursor to write code and itâs like it works well for certain stuff, not so well for other stuff. I still get Rust, but thanks to Cursor, I can actually write okay Rust. Because anytime I run into compiler errors, it just fixes all the borrowing stuff and Iâm like, âOkay great.â
Itâs great for weird SDKs. Like the other day, I needed to pull some data from Azureâs SDK and asked Cursor and it wrote up a bunch of code and it worked. I removed 25, try, accept, because it injects all this stupid error handling. Sometimes it works for mechanical refactorings, stuff like that, but sometimes it also fails spectacularly. So I donât know, man, I donât know. I feel like for most day-to-day stuff, itâs not the sort of 10x ... I donât know, I find that AI coding makes me, I donât know, 10%, 20%, 30% more productive, but thatâs it. By the way, thatâs not necessarily that much more than coffee.
Turner Novak:
True. Yeah. And thereâs a reason coffee is like ... What is Starbucks? Starbucks is worth a couple hundred bucks.
Erik Bernhardsson:
Yeah, yeah. I was spitting something about that. Yeah, exactly, right?
Turner Novak:
I mean, I got my coffee right here.
Erik Bernhardsson:
Python, switching from, I donât know, Perl to Python or whatever, this may be a bad example, but I think there are new tools coming out if you look 30 years ago to now. There have been new tools coming out every 5 to 10 years and they all make engineers 10%, 20%, 30%, 50% more productive and it all adds up. Actually, it all multiplies up, itâs all adds up but in logarithmic scale. So engineers today I think are 10x more productive than they were in the â90s or even more.
And I think AI represents a new tool and itâs actually really good, but I think to me, itâs not necessarily fundamentally different than whatâs already been happening over the last 30 years with engineers being more and more productive. And by the way, if you look at the last 30 years, employment of engineers just keeps going up. Thereâs more and more engineers in society. I donât know, I sort of expect that trend to continue for another few decades at least. I think thereâs so much software that has to get built, so many bad websites and so much bad stuff out there. I think weâre going to need a lot more engineers over the next decades.
Turner Novak:
Yeah. Well, I had Aaron Levie, the founder of Box on the podcast, and he was like, âIf youâre just a rational CEO and you realize that these engineers are getting better and more productive, why would you just lay them off and cut your spending? You should be hiring more and building more, because the more productive they get, any rational capital allocators, theyâre looking at the ROI, theyâre going to be investing more there. Itâs almost like, if it actually works, youâre going to hire more, thereâs going to be more engineers.
Erik Bernhardsson:
I think whatâs going on is there was a lot of bloat that happened pre-pandemic and during the pandemic, and a lot of companies were quite bloated. Then there was a lot of layoffs and people look at Twitter and theyâre like, âElon laid off 90% and itâs still operating,â and thereâs some truth to that. I think a lot of companies got bloated, but thatâs actually pre-AI. Thatâs unrelated. I think thereâs been a little bit of a bump in engineering employment, but I think itâs going to continue.
Turner Novak:
Yeah, Iâm sure people have seen the subreddits, people working multiple jobs. Iâve always read those and these people trolling, like this guy brags about having five jobs at once and all the keyboards.
Erik Bernhardsson:
It was like a few weeks ago, people started on Twitter, theyâre like, âYeah, so we hired him and then we have to fire him.â Then 25 other people piled in, theyâre like, âYeah.â
Turner Novak:
Yeah. I actually met a guy. We host some board game nights just with some friends, I met a guy, heâs just casually telling me he has two jobs. Iâm like, âWhoa, Iâve never actually met one of these people before in person,â and heâs just like, âYeah, they donât know. Neither one knows, I get the work done.â You make double the money. You just stack, you make 200K here, 200K here, whatever extra... Make 400 grand a year, whatever your salary is, but itâs interesting, because you think the first job, it just covers all your living expenses or whatever, and that second job, itâs just free money. Itâs not like youâre paying double the rent, double the car, car insurance, food expenses. Itâs like 10X more enticing to get that next job, because itâs just all cashflow to the bottom line. Itâs pretty high margin as an individual.
Erik Bernhardsson:
That to me, though... Great for that person, by the way, but that, to me ,signals to me that thereâs a lot of still bad management of software engineers. The fact that those two companies still havenât realized this guy is actually not contributing at the level of a software engineer, to me, thatâs insane, right?
Turner Novak:
Yeah. One of them was big tech, one of them was the government.
Erik Bernhardsson:
Yeah, yeah. It doesnât surprise me at all.
Turner Novak:
They barely even have the internet connected and half their stuff, plus their software probably doesnât even work properly. You mentioned something that I thought was pretty interesting, you said youâre both really bullish on AI, but youâre also bearish at the same time. I feel like itâs maybe a common sentiment with a lot of people. How do you rationalize whatâs going on? What are you most excited about? What is giving you pause right now in the industry?
Erik Bernhardsson:
The more I talk to other people, I actually think, like you said, itâs not necessarily a uncommon view. I think there are so many cool applications out there, and the other thing that Iâve been so shocked by is how much revenue there is in everything. If you asked me a year ago, âWhatâs the revenue in vibe coding or APIs for speech transcription?â Or whatever-
Turner Novak:
A couple million bucks, I donât know.
Erik Bernhardsson:
The fact that thereâs probably combined revenue of $500 million a year today in vibe coding, itâs mindblowing. These are people, theyâre paying for value theyâre getting. Another example is like Suno. I wouldâve never anticipated that the opportunity for AI-generated music, which I actually have no idea what theyâre making. What makes me bullish is those people they donât care about if thereâs an AI bubble or not, theyâre paying because they get value from a product, right? On the other hand, the things that make me a little bit bearish is, when I see the enormous amount of capital going into AI startups and AI infrastructure, I feel like, is that revenue actually going to match... Sorry, is that investment going to match the revenue? That, to me, brings me some pause. When weâre talking about, whatever, half a trillion dollars going into data centers today, is that actually going to materialize in terms of revenue in the next 10 years? That makes me a little nervous, that there might be an over investment in certain parts of this AI economy.
Turner Novak:
Yeah, I think I saw a stat. It was like OpenAI and Anthropic have actually added more net new ARR, or something like that, I donât know how you define ARR, like revenue, whatever, than every publicly-traded SaaS company. I might be getting the stat wrong, but it was a pretty insane statistic, that just those two companies are literally... Theyâve grown the same amount of total new dollars of revenue as the rest of the publicly-traded software industry.
Erik Bernhardsson:
The revenueâs insane, but on the other hand, yeah, itâs crazy how much money they make. Hats off, theyâre doing a fantastic job. Are they going to make $100 billion a year in five years?
Turner Novak:
Yeah, actually, they maybe might. Thatâs the thing.
Erik Bernhardsson:
To me, it feels like incredible. I donât know, I could be wrong. I probably wouldnât have expected them to make it billions a year either. I donât know, maybe Iâm continuously on the wrong side of history. I just feel like, okay, great, theyâre making, whatever, $10 billion a year right now. Are they going to get to 100? Are they going to get to 500? That, to me, makes me a little bit nervous. At some point, I feel like the law of gravity is going to take over and...
Turner Novak:
Well, and I know this is a little bit of the VC content marketing machine Kool-Aid, or whatever, that Iâm sure weâve all been drinking, but part of the rationality on this is AI just... It automates things. It does these low-level tasks for you, itâs replacing labor. If you just think of most companies, what do they spend money on? I donât know what the average is, but itâs a good chunk of revenue. Depending on your business, you might be spending 10 to 50% of your revenue on just labor costs, and whether thatâs in manufacturing, whether thatâs in software, whatever. If youâre just switching labor salaries into software revenue, I think itâs hard to predict what the ramp will actually look like, really. Itâs pretty big, itâs tens of trillions of dollars that could eventually move over, but then itâs a question of, does that happen in one year, 10 years, 100 years? Itâll probably happen within 100 years.
Erik Bernhardsson:
A few thoughts on that. First of all, if you look at historical technology shifts, I actually donât think a lot of money has been made in saving stuff. It usually comes from coming up with a bunch of new applications. I donât know if you look at trains or something like that, rails, we probably saved a little bit of money getting rid of horses, but I feel like it actually turns out we open up 10X more newer market. Itâs not necessarily saying order magnitude is wrong, I actually think, on the savings side, âYeah, weâre going to save so much money automating stuff,â I donât know if the value capture is going to be there.
Turner Novak:
I think itâs just on making people more productive. It should get more done.
Erik Bernhardsson:
I think that thereâs always more money in creating new markets and itâs not making people more productive. I think thereâs going to be a bunch of new jobs no oneâs even thought about, and thatâs going to be where the money gets made.
Turner Novak:
Youâve probably seen the memes of people... Tweets of âThese are the five jobs that are going to exist in the future,â and itâs like AI researcher, vibe coder, insert a couple more funny things. Itâs true, though. Probably will be a lot of vibe coders in 10 years.
Erik Bernhardsson:
I think internet is a good example. I donât know if internet displaced that many jobs, but it created a fuck ton of new jobs, and thatâs where most of the value of internet was created, not saving money for librarians, or whatever. Yeah, sure. It just turns out everyone can now find information faster, and that makes everyone more productive. Now thereâs all these new jobs, like making websites and building startups, and all this other stuff. I donât know.
Turner Novak:
To your point about the downsides of raising too much money, just for someone who hasnât thought through this before, typically, if youâre thinking about getting an exit on some investment, whether youâre a founder or an investor, when someone puts capital into a business, letâs just say I invest $100 million in Modal right now today, letâs say youâd raised no other money before. Typically, the terms you agree to is that the investor that puts their money in gets their return first. That $100 million you put in, thereâs like a 1X liquidation preference. If I put in $100 million, I get paid first. Letâs just say, whatever, you raised a bunch of money, multiples are really high, your companyâs worth 100 times revenue. You raise $1 billion. At $10 billion post-money valuation, things look great.
Fast-forward two years, multiples come down, youâre no longer worth 100 times revenue. Letâs say your multiple falls and youâre now worth 10 times revenue. Your valuation right there is down 90%. Well, letâs say you raised $1 billion at $10 billion post. Youâre now worth $1 billion on paper, youâre down 90% and you raised a billion. As the founders and the employees of the business, you get no capital back. Letâs say, also, maybe the companyâs only worth $500 million or $100 million or zero, the investors also donât necessarily get paid back. Itâs an interesting dilemma of like, hey, you do want the capital, itâs good to have money, but the more that you raise, it sets that minimum bar of failure, which itâs just good to consider. Donât just assume.
Erik Bernhardsson:
I 100% agree. Iâm going to throw in another point about that, that people often forget about. In Texas, when you raise at a very high valuation, you also screw up your 49A and you make it a lot more expensive for your employees to exercise your options. Thereâs all these tricks you can do if you avoid raising, especially when youâre near the QSPS threshold. Maybe you do some safe rounds in order to get around that and keep the 490A low. Thereâs all these things you can also do. I think thereâs many benefits to trying to keep the common valuation down, but 100% agree with what you said, right?
The more you raise, the more of a bar you create to exceed for the next round, and it gets harder and harder. I actually think VCs and founders have quite good alignment for most part, but the one clear misalignment I think I see a little bit in VC versus founders is VCs will always advise you to take more money than you need, and theyâre going to make all these reasons. Theyâre like, âWell, itâs always good to have money. Itâs good to raise from a position of strength. You never want to raise when you need the money. Who knows what happens in the market?â Which I always think is the most silly reason, because if you expect the market to go down, why are you then putting in money?
Turner Novak:
Yeah.
Erik Bernhardsson:
I donât know. I think VCs will sometimes make up all these reasons to want to foie gras their founders. I would be a little bit skeptical about some of those reasons.
Turner Novak:
Yeah. Ultimately, VCs make most of their money on average just from raising their next fund, because you get those management fees that are locked in. Itâs like a SaaS contract. You raise $100 million, you get 2% a year for 10 years, so you raise $100 million, you just locked in $20 million of management fee revenue for the next 10 years. Itâs just a massive enterprise contract, really. How do you raise your next fund? Well, I have an investor in Modal, I tell you, raise a big round at a super high price, so that, on paper, I look good and I go raise my next fund. Then itâs almost like who gives a what happens to Modal? Because thatâs in the rear-view mirror. Iâm doing new deals, Iâm deploying more capital, because I got to get the next fundraise. The system does work, itâs worked really well at creating a lot of value, but itâs definitely something to just be cognizant about if youâre a founder. The incentives are not always fully aligned.
Erik Bernhardsson:
Yeah. Especially when you think about how long the economic cycles, the macro cycles last. You have to think about the return profile over an entire economic cycle, and I think that the truth is most startup investors have not lived through a whole cycle, so they donât necessarily think about that. They think more about, âWhatâs the markup I can get on the next round?â Because thatâs going to determine my promotion.
Turner Novak:
Yeah.
Erik Bernhardsson:
I think most startup investors are more interested in, âI put in money in this valuation and then the next round better be a higher valuation.â They donât care about going public, because thatâs like 10 years later. Theyâre like, âYeah, whatever.â
Turner Novak:
Well, weâll all be dead by then. Hopefully not. Yeah, for me, my first fund, I started investing in Q1 of 2021, you get slapped in the face pretty fast. Within two years, it was like, âHoly shit, the valuation does matter.â Good businesses can be overvalued. Those are two separate things. Also, shitty businesses can be undervalued, too. thatâs called value investing.
Erik Bernhardsson:
I donât think thereâs any value investing in startups, by the way.
Turner Novak:
Yeah, thereâs not, thereâs not. Yeah, it was a good lesson I got slapped with pretty quickly. Was just like, âHey, the price does matter.â
Erik Bernhardsson:
You mean in the mini downturn in summer of 2022?
Turner Novak:
Well, yeah. You get in and youâre like, âHoly shit, you cannot be investing at 100 times ARR if the company is not actually growing 10X in a year.â Letâs say you value the company at 100X ARR, the company grows 10X over the next year. To hit the same price, you need the company to be valued at 10X ARR. Hopefully, your next round is like 20 or 30X ARR if you were 100X before, but multiple compression is a thing that is hard to control. A lot of itâs just market. Some of the issues, too, is letâs say Iâm another founder, I compete with Modal and I say, âModal raised it,â whatever ARR price you raised at. Thatâs all I know. I donât know, was there a debt in there? What are your margins like?
I might not even know what your growth rate is, honestly, and I just say, âModal is this, I should also be valued at this.â You get this game of everyone comparing themselves, and at the end of the day, only one of those is a good asset and thereâs probably a lot of subpar lower quality assets. Maybe theyâre not given quite the same premium on the multiple, theyâre slightly worse, but if one of them is worth 100X, one of them was worth 90, one of them is worth 80X ARR, the 90 and 80 should probably be at nine and eight.
Erik Bernhardsson:
I would say, though, that multiples today are healthy. I actually feel like multiples a few years ago were more out of whack.
Turner Novak:
Yeah. Yeah, theyâre pretty fair. If you look at OpenAI, you can argue that... There was that one viral clip about Bessemer explaining, âIgnoring the negative gross margins, these are actually pretty reasonably valued businesses on a growth rate.â
Erik Bernhardsson:
That, again, brings me some comfort. If thereâs an AI bubble, I think itâs going to fizzle out. Itâs not going to be as spectacular as like dot-com, when you were raising money based on eyeballs, whatever. Thereâs real substantive value being created here.
Turner Novak:
Yeah. To your point, people are using them. I donât know if you remember Web3 and crypto. The interesting thing was the spreadsheets looked good, there was revenue, but they were not... It wasnât real revenue. It was literally day trading gambling versus, with some of the AI stuff, itâs like, âWe made a website that someone uses to make commercial transactions on.â Thereâs underlying commercial purposes to it.
Erik Bernhardsson:
Although the Web3, I donât know in hindsight, was that a bubble that popped? I donât know, look at Coinbaseâs valuation, itâs still pretty good. I personally donât like crypto, I feel like thereâs an enormous amount of resources that weâre draining on society into whatâs basically gambling. For that reason, I despise crypto, but if you look at the valuations, theyâre still pretty high in crypto then.
Turner Novak:
Itâs all function. Iâm just like, âPeople have too much capital, theyâre looking for returns. Theyâre deploying it into things that they think will keep going up.â Really, at the end of the day, thatâs all it is.
Erik Bernhardsson:
Totally, exactly. Thatâs actually why I think in the way... I actually expect a little bit more inflation and slightly higher interest rates. This may be good, because it actually creates a higher hurdle rate, which means now people have to go and look for even more productive stuff to invest in, which in theory, it should drain crypto, because it has a 0% interest rate. I donât know, man, it hasnât really turned out, but in theory, I think thereâs an argument for it.
Turner Novak:
I got a question for you. Akshat said I got to ask you about how you keep the temperature and CO2 levels in the office.
Erik Bernhardsson:
Oh, boy.
Turner Novak:
Yeah. Whatâs the story there? You guys are pretty...
Erik Bernhardsson:
Iâm not a health nut, Iâm certainly not an RFK. I think thereâs certain things that actually do matter, though, and one of the things Iâve been radicalized on is CO2 levels. Thereâs actually very clear relationships. CO2 levels versus productivity and cognitive performance, itâs a very clear relationship. A lot of indoor air, a lot of office... Schools also. I actually think one of the most cost-effective thing we can do in this country, in this world, is just better indoor airs in schools, which... When Iâm a billionaire, I want to just fund this stupid thing, which is send CO2 monitors to all the schools in the world. If itâs too high, crack a window. Thatâs it. Yeah, I bought a bunch of CO2 monitors for our office, I go and look at them. If itâs bad, I open a window. Honestly, I think it matters.
Turner Novak:
Itâs that simple. You literally just open a window and it clears out the CO2.
Erik Bernhardsson:
Yeah. I donât know, thatâs the only way effectively to... Sometimes if you have a good ventilation in an office, it just automatically triggers. I donât know. Weâre in Soho, this building is 100 years old, I think the only way to really get fresh air in it is to open a window.
Turner Novak:
Interesting, okay. What does the data show about CO2 levels? It just makes your brain think slower?
Erik Bernhardsson:
Yeah, yeah. Thereâs all these studies. You look at problem-solving ability, it goes down very quickly. Normal indoor air is, I think, 500 PPM, parts per million. Thatâs good, thatâs outdoor air roughly, or something. Maybe three, 400. But often, in offices, it hits 1,000, 2,000, 3,000. 5,000, by the way, is brain damage. In airplanes, it often hits like 2,500, 300. I used to wonder, âWhy do I always get so drowsy during takeoff?â Itâs because CO2 levels go up inside the airplanes. I always used to pass out.
Turner Novak:
I do feel less productive on planes, I feel like my motivation level... Because I usually try to work on flights, but thereâs definitely some cases where Iâm like, âFuck it,â Iâm just going to sleep or listen to music, or something.
Erik Bernhardsson:
Itâs primarily during takeoff, that CO2 levels peak. It gets a little bit better once youâre in the air and it starts circulating.
Turner Novak:
Yeah, I could actually see that. Yeah, because usually, it takes me a little longer to switch into pull my laptop out type of mode after takeoff. One nice thing I wanted to ask you about, so youâve worked in Europe and youâve worked in New York, in seeing the differences between maybe just Sweden, maybe just Swedish tech, but Europe versus the US. Whatâs the difference in the tech ecosystems, in your opinion?
Erik Bernhardsson:
First of all, I think Europe has incredible tech talent. Iâm actually not a permit bear in Europe, I do think thereâs a lot of structural challenges with Europe, but I look at, actually Sweden I think itâs also doing okay versus just Europe, but I think thereâs tremendous talent out there. I think itâs just not being taken advantage of, or Europe is really not... I donât know. US has way more availability of capital, it also has much better options. Taxation, I think that makes a huge difference. I think when you look at these startup cultures, a lot of it just comes down to having smart people, which I think Europe has. Then the other thing is just having a history of success and just having a culture of like, âYeah, my buddy worked at that other startup, and he made so much money. Oh, my God, maybe I should also work at a startup.â
That latter part I donât think has existed quite yet, and then maybe a third thing I feel like hasnât quite existed in Europe for a long time is this... I donât know what to call it, but reality distortion, field-type founders. Daniel Ek was actually really good at it, the founder of Spotify, but I feel like, at least in Stockholm, thereâs been a lack of big vision, crazy megalomaniac founders until quite recently. Now thereâs a few of them, and I think that makes also another huge difference. Itâs just having these big visions. Iâm so tired of European startups building... Theyâre building a SaaS for grocery stores, whatever. Iâm like, âBuild something crazy, something big,â and I feel like thatâs starting to happen a little bit more. At least in Sweden. Iâm bullish, but I donât know.
Turner Novak:
Do you get that a lot more than in the US, just like bigger vision opportunity?
Erik Bernhardsson:
Yeah, itâs too big. You got to SF. Everyoneâs trying to build the next OpenAI or the next whatever. Iâm like, âChill. Youâre like 21.â
Turner Novak:
âWeâre going to cure cancer.â
Erik Bernhardsson:
Yeah, yeah, exactly. Itâs good. I think we need more people on average to have these big, bold visions. I think itâs good, I think itâs good.
Turner Novak:
It feels like the sweet spot is solve a specific unique problem today with the vision of, âWeâre going to cure cancer,â but weâre starting with this and you bridge the, âIn 20 years, 30 years, this is how we get there.â
Erik Bernhardsson:
Yeah, yeah.
Turner Novak:
Well, cool. Well, this is a lot of fun. Thanks for coming on the show.
Erik Bernhardsson:
No, this is great. This was awesome. It was amazing. Thank you so much for having me
Stream the full episode on YouTube, Spotify, or Apple.
Find transcripts of all other episodes here.