Gustav Söderström - How Spotify Thinks - [Invest Like the Best, EP.424]

My guest today is Gustav Söderström. Gustav is the co-president, chief product officer, and chief technology officer at Spotify. Gustav lets us behind the scenes on how Spotify thinks about the future of audio and video and what leadership lessons he's learned from making mistakes and taking risks in a rapidly changing technology landscape. He shares fascinating insights on their synchronized team structure and how they position themselves as the R&D department for the entire music industry. We discuss their integration of AI, their unique bets board process for allocating resources, and how they've evolved from a music service into a multimedia platform with over 650 million users.

Maybe a fun place to begin is the obvious place. Everyone is facing this giant shift in technology. My friend Ravi Gupta calls this “imperative AI or die.”

I would love to hear how you and Spotify are thinking about this challenge. I know you've embraced it very quickly, and you were very early to using machine learning and data science all over the product. But this is a big shift, and you and Daniel and the team are some of the most thoughtful people about addressing shifts like this, and you've done it before. Walk us through in some detail how you first felt it, what you did about it, and what it's like to be at a big company and process something like this.

1. AI Becomes a Macro Wind

Yeah. It's a great question. I think it's the right description. At least in the longer term, I think it is AI or die, just like it was smartphone or die, and before that, internet or die, computer or die. This is one of those shifts where it's not your choice whether you adopt it or not. It's gonna happen to you.

It's the epitome of a macro wind. Usually, when these macro winds come, we have a saying internally that you can have the macro wind blow in your face. It's not gonna change its direction. So you basically need to reposition yourself so you get the wind at your back and can surf this macro wind, or some people call it a macro wave that you surf.

We've been through a few of these. The first one was really the smartphone when that came along. Spotify was really well-positioned for the internet before the smartphone, where we had a free tier on desktop, and that's where we'd acquired users. They built a playlist, and they retained themselves. Then the ability to listen on the go was a paid feature on Spotify, and that was fine when the majority was computers and the minority was smartphones.

Then when smartphones took off, we faced an existential crisis where there started to be consumers who didn't have a desktop. They only had a phone, so they had no free experience, and our entire model died. So that was one of those moments when we had to reposition the entire business model and figure out how to do a free tier on mobile that doesn't cannibalize the paid feature, which was mobility. We can talk about how we figured that out later, but that was one of those examples, and I think this is a similar one.

The big question to me is: Does this require a business model change, or is it, quote-unquote, “just a product change”? The other thing that is different, I think, about AI is that it's not gonna touch one thing. It touches the consumer product, but it also touches your productivity and competitiveness as a company. So there are lots of different angles to start.

But as you said, we were quite early with machine learning. The journey we had was that we saw users coming onto Spotify, and then they started playlisting, and that retained themselves. But it was only a certain number of people who were good at playlisting because you have to know the catalog in your head, the new releases, and the back catalog. Some people retained themselves really well.

Then we tried to scale that behavior by having editors create playlists for people who couldn't playlist that well, and we saw people using social to find inspiration. Eventually, machine learning started happening, and we saw this opportunity to build a music friend for everyone. So that's where we started. We started investing in that and got quite good at it.

2. Generative AI Rebuilds Consumer Products

I think some people say that AI is just machine learning. It's just a new word. And it's an interesting question: What is the difference? I think the difference between what people used to call machine learning and what we call generative AI is that statistical machine learning was an output mechanism, and I think the epitome of that age is the full-screen TikTok feed.

UIs shape themselves after the technology that powers them to maximize metrics, and I think that is the UI that maximizes the statistical explore-exploit paradigm of old-school machine learning. Even if technically they're both machine learning, I think of generative AI as the new age.

I think the big shift with generative AI is that you can take natural-language input. So what happens with generative AI is that it's two-way. If you think about Spotify, for example, this consumer product, the way it looks, it's almost like an old-school broadband connection.

The broadband where you have maybe 1 megabit downlink, but only 150 kilobit uplink. A lot of bandwidth down, but not a lot of feedback. This is what most consumer services look like. You have streaming video on the downlink, a lot of information per second, but the uplink is only a few clicks and swipes. It's a very, very narrow signal, and this is what the previous machine learning age focused on.

I think what changes in the age of generative AI is that the uplink can now be English language. It can be almost as rich as the downlink, and I think that requires all of us consumer companies to, in the limit, totally rethink the product.

So if you just do the deduction of what I said, if the full-screen TikTok feed is the epitome of the ML paradigm, the asymmetric downlink-uplink paradigm, what are the chances that that is also the epitome of this generative AI age? I don't think so.

I think consumer products are gonna change fundamentally. I can't predict exactly how. I think they're gonna be much more symmetric in terms of information you receive versus information you give. And I think if you fast-forward 5 to 10 years, almost all big consumer products are gonna be a conversation to some extent, rather than this service that you use.

So really the job for us on the product side is to try to figure out what is the next paradigm, and I don't know exactly what it is yet. We're experimenting, and if I did know, I probably wouldn't tell you right now. I'd sit on it for a bit. But this is where we are on the product side.

Mm.

Then we can talk a bit about the productivity side as well, where there are the obvious gains in terms of coding productivity, and we are using all the tools that everyone else is using. But as a big company, there are a few differences from the startups because, so far, generative AI in coding has had the most impact when you write net-new code, which is a lot of what you do as a startup and a tiny bit of what you do as a big company.

Most of it is just refactoring, et cetera. And I think I saw some statistic that in a big company, you basically code 1 out of every 8 hours in a day. So not only is coding only one-eighth of the time, but of that one-eighth of the time, net-new code is very small.

So I actually think the biggest impact is yet to come when it comes to coding, and that's two things. These models are getting big enough to understand really large and complex code bases like Spotify's, and we're not quite there yet, where these things can refactor our code base. It doesn't have quite a deep understanding, but it will, and that will be a big shift.

The other is doing automatic peer review. We're just on the verge of that working. It's not quite good enough that you can trust it, so a lot of developers sit and wait for their code to be in review and come back. So I think we're seeing that ramp. I think we're gonna see it ramp a lot in the next few years.

But then the really interesting side is these other 7 hours, what a developer does, which is a lot of communication, planning, working with designers, prototyping, and meetings. Those things, I think, will actually have as big or even more impact than the coding itself.

I want to start with this downlink-uplink part. What have you learned about consumers' willingness to put a lot of effort into the uplink? It seems like with the chat interfaces, the GPTs of the world, we know that people are willing to do a lot of back and forth because it's the native interface. You're going there expecting to write a lot of stuff. You're copy-pasting prompts from Twitter or whatever.

In an app like Spotify, how willing are people to get not lazy and really descriptive about what they actually want? What have you learned about the nature of people's laziness versus their willingness to put a lot of work in to get the thing that they want via that richer uplink?

3. AI Playlist Reads Your Mind

Yeah, so that's probably the most exciting thing for us about this generative AI age and the dual-uplink paradigm. Previously, we mostly relied on some explicit input when you playlisted. That's high-value information. You are sitting there thinking, "This song goes really well with this song and that song." So if you think about it as labeling, even though you're playlisting for yourself, you're sort of labeling these tracks, at least in relation to each other, and you're putting a lot of effort into it.

That was and is our big advantage in music recommendations, even though generative recommendation systems are starting to take over from these more old-school collaborative systems. We had some really strong signal like that, where you quite seldomly invested a lot of time in producing a data set that described you by playlisting. But most of the time, we just had skips, and the challenge for us is that the phone is in the pocket.

So even if we had a thumbs-up or thumbs-down, you're not going to take out the phone every time and say, "I didn't like this because of that." Even a thumbs-up or thumbs-down requires you to pick up your phone, unlock it, and open Spotify. What you can do from your earphones is skip. So we have the skip signal, but that is a very blunt signal.

We play you a song, and you skip it. That could be because you absolutely hated it. It could be because you love it, but it's the hundredth time, so you're tired of it. It could be that you love it, you're not tired of it, but you're at the gym, so jazz is not the right thing. All of those just look like a skip to us.

We have a lot of that signal, but it's blunt, and you will never get to perfect personalization through that. Now what we find with generative AI—one of the first services that we've launched, which is live now in 40 countries—is something called AI Playlist. We literally use an LLM that is trained on your listening data, world knowledge, and so forth, and you can literally tell us in English what kind of playlist you want.

Previously, you could playlist songs and maybe put a title on it, and we could guess this is probably a running playlist. So we could do something. Now you can say, "I want a running playlist that is EDM. I want big drops. I want it to be 160 BPM." Then you get a suggestion from the LLM, and you can keep a few tracks and say, "These were good. These were not good. Now refine it. I don't like these artists, but I want more of that."

For us, it's the first time that we get that kind of fidelity about what is actually in the user's mind. One way to think about Spotify is that we always try to reproduce a small part of your neural cortex on our servers, which is very hard with a clickstream of skips. Now when you tell us what's in your mind, it gets easier to approximate you as a person. This is really the first time that we have that signal.

One way I like to think about it is that when we do user research, we do two things. We do quantitative testing, A/B tests, but before that, we do qualitative testing. We interview a few people deeply to understand the need, then we build a product, and then we A/B test to see if we're right. The promise of generative AI is really deep, ongoing qualitative user research with almost 700 million users all the time. It sounds big, but if you squint at it, that's kind of what it is.

It's interesting how many different ways you could take the product with this new technology. I would be really curious to know the apparatus inside of Spotify—the leadership team, the product team—and literally how you run the process of deciding what to do with your—You have a big team, obviously, but no matter what, you have limited effort, limited units of energy—

Yeah. There's never enough people.

—you can apply. So there's a huge space of stuff you could do with this technology and the exciting advantage that you have of 700 million users. What does the literal meeting-by-meeting process, the setup process, look like?

The reason I'm asking this question is that so many companies face this same challenge. It's exciting, but also scary, that they need to get the innovation before somebody else does and disrupts them. What is the background process for how you arrive at the things you might try?

4. Spotify Ranks Its Bets

There are really 2 things. We have a very structured process that I want to talk through—how it works—but there are also some concepts that we use. Over the years, I've introduced some strategic frameworks to the company, which I know you're passionate about, like Seven Powers from—

Yeah.

—Hamilton Helmer, and Shishir Mehrotra's bundling framework—

Shishir, yeah.

—who wrote it.

He's on the board, right?

He's on the board.

Yeah. Secret weapon.

Exactly. Very good secret weapon. I also found Better, Simpler Strategy by Felix Oberholzer-Gee to be very good.

What's that one? I don't know that one.

It's a concept of the value stick, where you have willingness to pay. It's very important for us. The way we measure value is your willingness to pay. But what it introduces is also the willingness to sell. If you think about your staff, what is their willingness to sell their services to you?

Everyone focuses on increasing the willingness to pay, but you can also increase—or, depending on how you think about it, decrease—the willingness to sell. It turns out the best companies in the world are not necessarily the ones who actually pay the most. They're the ones with the most interesting mission, the best culture, et cetera.

Because we're a bundled service where we try to just give users more and more value all the time, we put in lots of music—that's value. Then we put in more podcasts—that's value. Now we put in books—it's more value. This framework of willingness to pay and willingness to sell fits our business really well.

The job for us is to keep the willingness to pay quite far from the actual price. That gap is how much consumer surplus you're giving. Our goal as a service is to make sure that Spotify is just an amazing deal. You're always going to feel that the willingness to pay—the actual value you perceive—is way over the price that we have.

So we use that framework quite a lot. Introducing these frameworks not just in the business organization, but also in the product and technology organization, makes people think in more structured ways. It gives people a vocabulary. We can talk about network effects, amortization, brand power, and all of these things.

I spend a lot of time getting the teams to use these frameworks so that we have structured strategic thinking. The other thing I've tried to push as a thesis is that I'm a big fan of Socratic debate. I'm amazed, like many other people, how far the Greeks and the Romans came with just discussion.

Mm.

Even though they didn't have science, just reasoning. Strong reasoning is very useful. So I try to push this provocative line of, "Talk is cheap, so we should do a lot of it," as a counter to "moving fast and breaking things."

Yeah.

It's so cheap to talk, so we should actually do a bit more of it. It took the Greeks to the concept of the atom. So we do a lot of talking and ideation that is quite structured, and I do that with my leadership team. Often, it's the leadership team plus one, which means the VP layer and the director-plus layer.

I have a lot of time just for discussing concepts. Going back to one of my heroes in life, David Deutsch, his books The Beginning of Infinity and The Fabric of Reality shaped me quite a lot. He talks about something called good explanations, and he has a list of what a good explanation is.

It obviously needs to be falsifiable and so forth, but it also needs to have reach. It needs to scale. An okay explanation explains this phenomenon, but it doesn't scale to other phenomena. It doesn't scale up and down. A really good explanation scales from explaining how the Earth works to the solar system to the planets.

But it's also very hard to vary, which I think is often underestimated. If you have an explanation and you can switch it out for another explanation that explains the same thing—if you explain the weather using gods, you can switch out this god for that god—it's probably not a good explanation. It needs to be very hard to vary. If you vary it, it doesn't explain it anymore.

The last thing he says is that explanations should not just be predictive. That's not an explanation; that's a model. An explanation needs to explain why. So I try to push my teams—even if something works in an A/B test, I tend to say, "I don't want to launch it until you have a good theory of why it works."

If you figure out the why, it's the difference between pattern recognition and actually understanding something. Pattern recognition is useful. That's called seniority. I love people with good pattern recognition, but if they can explain why it works, it scales to the entire organization. Other people can use that knowledge.

Mm.

It's much, much more valuable. So those are some of the concepts that I've tried to put into the organization over time.

So I think that's important because that shapes the culture. Then we have the structured process, which is that we execute for 6 months at a time. We have something called a bets process where all the VPs—about 14 of them—participate. One of the benefits of Spotify is that we're so small that all the VPs, the entire company, can fit in one room.

We meet for 3 hours every Tuesday, so the company is completely synchronized, for good and bad, and we can talk about that later. Every 6 months, these VPs literally pitch, as if we were a VC and they were a startup, the bets that they think the company should pursue and why. It's very much like a startup process. You don't get to use the fact that Gustav or Alex or Daniel may like you. This is like a VC meeting. You have to convince us.

Then the other co-president, Alex Norström, and I decide, based on these pitches, on a global stack rank. This time we have 44 bets. That's usually between 30 and maybe 50. We stack-rank them from 1 to 44, then we go out to the org and say, “Now try to resource this.” They start from the top, and then maybe they get to 30 and say, “This is what we can do in the next 6 months.” Then they commit to those things, and we start executing.

It's a good mix of bottom-up innovation, where you leverage not just Daniel, not just Alex and me, but all the VPs and the layers below to come up with good ideas because they're closest to the user. But then there is global synchronization. We stack-rank them, make sure that they fit a single strategy, and then it's back to the org and they commit. As I think you know, you're going to be much better at delivering something if you were the one who said, “I can do this,” than—

Yeah.

…if your boss said you can do this. So that's the process. But leading up to that, we have something called a prototyping phase. During the previous 6 months, we prototype, in a combination of Figma and increasingly generative AI tools, what Spotify should look like after the next 6 months, or could look like.

This prototype also helps synchronize the entire company. What I found previously was that when people submitted these bets, everyone had in their mind what their great feature would be. You start building, and then down the line you realize that you were not actually aligned. Then you get a lot of fighting toward the end of the cycle, where this thing doesn't work with that thing, and things don't work out so well.

What I've tried to do now, together with Alex Norström, is synchronize the entire company. Alex and I don't have our direct reports in the team. We meet as a single team for 3 hours every Tuesday, and we try to use the fact that we're small as an advantage instead of as a disadvantage versus our competitors, who are very, very large companies.

We prototype everything up front. So all the so-called, quote-unquote, “fighting” happens before you actually commit to doing something, and you have something you can hold in your hand and say, “This is what Spotify would look like if we pull this off.” That's a combination of cultural input and then a very structured process for actually making it work.

I have so many questions about process. The first is how that 3-hour meeting on Tuesday works. What is the structure of that meeting?

5. Synchronization Prevents Execution Drift

It's called the E-Team, or execution team, so it's very focused on execution of the company. The idea is that if you have 5-day working weeks, on average there's never more than 2½ days before you can escalate to me, Alex, and all the other VPs if you're blocked on something. You should never be blocked for more than 2½ days at most.

Because we run this synchronized ship, if you're blocked, it gets very expensive because everyone else is downstream of you. If you're running a synchronized operation the way we're doing, escalation processes are very important, and resolution is very important. So a big part of that meeting is people saying, “We're off track here. I'm dependent on this man or woman over there who hasn't done what they said.”

The beautiful thing about being able to have all the VPs in the same room is that, in so many meetings that I'm sure you've been in, people say, “Okay, we'll take that offline. I'll talk later.”

Mm-hmm. Mm.

And what we said is you're not allowed to say the word “offline”—

Can't leave. Yeah.

…or “later,” because that person is in the room.

Yeah.

I'm dependent on maybe Anna over there for this, but then Anna is actually there, and Anna can say, “Okay, I didn't know that,” or, “I'm going to solve that.” So it's real-time resolution. Very simple in theory, but incredibly powerful in practice.

Mm.

Most companies don't do it. This notion of not taking it offline or taking it later—real-time resolution—is why it's 3 hours. That's one thing about this meeting.

Another principle we have in that meeting, besides nothing going offline, is that you actually can't bring your direct reports. The idea is that if you bring in a lot of direct reports, two things are going to happen. One is that the VP isn't going to be forced to know the details as much. So I'm trying to literally force the VPs to solve it themselves, because I want them to be in the details.

You're not allowed to bring anyone else in to explain your thing. You have to be on top of it enough to explain it yourself. The other benefit is that, over time, these groups get very tight because you don't switch people out all the time. So you build strong rapport. People can be honest. No one is afraid. It's a very strong and high-functioning team.

That's a lot of what we do. The other part is strategy and looking forward. Let's say something we've been working on for some time now that is public: We wanted to introduce music videos. A team goes off and says, “What does that take in terms of licensing and product? What are the cost implications for the company and for the P&L?” They come and present to the team: “This is what we want to do. This is how long we think it should take.”

So it's a combination of keeping the engine running and never stopping, and also planning for the future. But we don't really plan in that meeting. It's too big to have detailed planning. That happens in focus rooms with smaller groups and experts, and then they come and present to that team.

We're not the only company that does this. I know Airbnb does something similar. I spoke a lot to Brian Chesky about it. I think Netflix may have had something similar at a time, but they're now divided into content, the business, and product.

What I think is important about this is that it's both the business and the product side, and we talk a lot about product there. So the business people at Spotify know an awful lot—

Mm.

…about AI and about what a monorepo is. They're there for the prioritization of technology. But on the flip side, all my product people and engineers know exactly what the P&L looks like. They know our goals. They know what gross margin is. They know everything.

So that's quite unique, and that gives them the CEO perspective that I think disappears in many companies because we put people in these roles: you're a product person, so you're not supposed to understand finance.

Right.

That's not true. If you're the CEO, you have to understand all of it.

If people are listening and are curious about this bets board process, where you can submit projects and it seems like a really elegant way to allocate capital, what advice would you give them about the pros and cons of this process? I know you've been doing it a long time. How has it changed over time to reflect the learnings of what makes it work or fail?

The concept itself is actually really straightforward. It comes from the Kanban board. It comes from the developer community, which eventually comes from car manufacturing.

Oh, interesting.

It's really the concept of stack-ranking, which is very easy in theory and very hard in practice. Very few people manage to say, “This is actually more important than that.” They just say, “These things are very important, both of them.” When you press them, they say, “No, they're equally important,” but then they're not ranked.

The real secret is to stack-rank and say, “You have your 2 darlings, but if you have to kill 1 of them, which do you kill first, in reverse order?” It's very easy, but hard to do across the entire company and agree on that.

But once you have it, it gives so much clarity to the org. What happens when you say, “These 3 things are equally important,” but they're not really—they never are, because you're going to have to choose—is that you just push the decision down the org. This VP, who was on the hook for that thing, is going to start fighting this VP, who was on the hook for the other thing.

If you as a leader don't bring clarity, you're going to set your org up for fighting, and people are very nice. They're going to think that they don't like each other. So just the stack-ranking and being completely transparent across the entire company means that if I come to you and say, “I need you to do this,” and you say, “Yeah, but I'm doing this,” we look at the board and say, “Oh, right. We should do this.”

It's very simple, but very effective. When you do that, there are a bunch of theoretical questions you run into. Once you have this bet board done, do you resource it globally? Do you go through every developer and say, “Let's just try to get as far as we can”?

That planning process is hell if everyone is up for grabs. None of my VPs have any estimate of what resources they will have. You basically disempower your entire VPs. It's effective in a sense because you do perfect resourcing, but it's incredibly inefficient to do the planning.

So then the question is, how do you divide it into blocks? The structure we have is a platform organization. They work with GCP and the cloud, developer tools and so forth, security, and all of that. Then we have an experience organization that’s responsible for the entire consumer product across mobile, car, desktop, et cetera. Then we have a personalization organization, because that’s so important to us, that does all the AI and the recommendations and balances between books, music, podcasts, video, et cetera. And then we have 3 business verticals: music, podcasts, and books.

They have their own resources, and what we do is start by asking them to resource as far as they can with the resources they have without stealing from each other. Then we get as far as we can because you need to give them predictability for them to be able to plan their own work. At the end of that process, you may move some people around globally to make sure that you don’t have something really important with 2 people missing. That’s not optimal for the company. So you may move some people around, but largely, we try to let people keep the resourcing.

There are lots of those problems that you run into. But I would say the biggest risk with this model—it sounds nice if you’re perfectly synchronized—is that the planning is very expensive. So you have to be really good at planning, and we’ve had to build our own tooling. We tried some external tooling for planning, but that wasn’t good enough.

If the planning doesn’t work, the overhead just grows very quickly versus execution. We execute for 6 months in order for the overhead not to get too big, but we can’t go to a year; then you can’t react. A quarter is too short. It’s too much planning overhead versus execution. So the planning is the thing that you have to get really good at, and I’m not going to say we’re really good, but we’re getting better all the time. It’s the thing that I care the most about, making sure that the planning is reasonably good.

If you can do it, for us, it’s critical because the whole of Spotify’s product strategy is that we have large distribution, closing in on 700 million MAUs for a single application. Our entire strategy is basically that we decided this many years ago, before it was popular, but you saw the Chinese starting to build super apps, whereas the Western world built 1 app per use case. We’ve adopted the Chinese super-app idea and said the hardest thing is going to be to get installs. You could see the average number of installs from the app store dropping below 1 on average, so distribution became the most important thing.

Then we chose, when we did podcasts and later books and videos, that we were going to build them in the same application because then we could leverage our own distribution. But that has drawbacks because everything is dependent on each other. You’re going to ship 1 app to the app store, and everyone is a stakeholder, so you cannot divide and conquer. You cannot say, “Well, the book team, you can run ahead,” or, “The music team, you do this.” No, everyone has to wait for everyone.

So because of our consumer strategy, the company needed to be synchronized, and because it needed to be synchronized, we needed a really strong planning process. It’s an outcome of our consumer strategy. What I would say is, it’s not the right one. It’s the right one for us. We’re good at doing global changes, like changing the entire UI, because we’re synchronized, but we’re probably much slower than other companies at—

Small change, yeah.

Trying something because it needs to go through a lot of planning. I don’t think you can win in planning. The best you can hope for is to be quite good at the important things and not so good at the less important things.

Hmm. I want to come back to something very interesting you said around the adoption of some of the tooling that’s at the most cutting edge. So let’s take Cursor as an example—

Yeah.

—of a company that now everyone’s familiar with, with a $10 billion valuation. It seems like every software engineer is using Cursor to make themselves better.

Yeah.

But the way you framed it was so cool: sure, but that’s new code primarily. That’s a fraction, 1/8 of their time. In the pie chart, it’s a very small sliver that’s being addressed by Cursor at big companies. Can you describe how you think this will play out? Because it feels like the public markets especially, I guess private markets too, are very curious about how AI companies and products and tools will address this much bigger part of the pie that sounds like it really hasn’t been hit too directly yet.

Yeah. There are a couple of things that are interesting that I don’t think are super obvious. One is, it used to be that every developer started using Cursor, but now I’m starting to see a lot more non-developers using Cursor. That’s partially because the industry is starting to agree on this protocol called MCP—

Mm.

—Model Context Protocol, which means that if you take your internal services and wrap them in an MCP, you can speak English to your infrastructure.

Mm.

If you’re a developer now, or if you’re a designer, for example, or a product person, let’s say you want to prototype a feature in Spotify. One workflow is that you take the existing Spotify, double-click and screenshot it, upload that into Cursor, and say, “Wire this up,” clickable in HTML. Then, if your services are wrapped in MCP, you could theoretically say, “Now wire this up to my songs feed or something.”

You can prototype even though you’re not a developer because the infrastructure is wrapped in English now through MCP. I think that’s an important thing. So I think you’re going to see many more people using Cursor than just developers. I’m starting to see that. I had 1 of my PMs in Sweden—she’s from New Zealand and doesn’t speak Swedish—do her taxes in Cursor. She managed to wrap the Swedish Tax Agency in an MCP. She’s not a developer.

So I think it’s going to grow outside of developers, but I think this points to what is actually happening in many of these big companies, which is why the startups can move faster. A company like Spotify has tons of infrastructure. We have the database with play history going 15 years back. You have who is in the family plan—that’s 1 server. There’s 1 dataset somewhere. Your taste graph is a dataset, and so forth.

Now, here come the big AI companies, and they give you this reasoning engine. Some of them are open source, so basically, for free, you get what’s getting close to AGI. So now you have this thing that you thought would be incredibly expensive, and you get it almost for free. It’s a gift. You start using it. What’s the first problem you run into? You say, “How has my music listening changed over the last year?” That’s not exposed as an API.

Because in the previous machine-learning world, that data—the listening data from 15 years back—is on cold storage somewhere, and an engineer would have had to do an SQL job that may have taken a week to pull it up. Then you would have trained a model, and then you put it back in cold storage. Now you want to be able to reason over that in real time. You need to expose all your data as APIs in real time. So actually, my biggest job to enable AI is not AI engineering; it’s old-school engineering.

Mm.

Exposing all this data that we have so that you can have a reasoning engine reason for you as a product person, or actually for me as a consumer, potentially over my own data in real time. So I think that’s what’s happening. The combination of there now being a standard, MCP, that you can wrap these APIs in, and companies like us trying to expose all of this data means that a businessperson or lawyer, a product person, or a designer will be able to use Cursor without having to code. They can actually at least prototype or talk to real services. That, I think, is the journey that we’re on.

Mm. Mm.

It started with developers, but I think as you expose the infrastructure and wrap it in APIs, I think it’s going to have to go outside.

Is there a way to say that and lose a lot, but summarize it as: what we’ve seen happen with developers is going to happen with other similar tools? I’m surprised that it’s Cursor that they’re using. That’s quite interesting. But with other similar tools, more of our work is going to feel like we’re working with a team, speaking to a team, using natural language to prototype things and to try things. That will diffuse slowly through the entire— not only the hours of the software developer, but the hours of each of the other functional areas.

Yeah, I think so. It’s hard to see where it’s going to land because you’re somewhere right now, but we’re pretty certain that that somewhere is on this curve. So you can be pretty certain that the workflows you see right now are not going to be the same, and that’s actually one of the problems. How much are we going to build for what we see right now when you know the models are going to be more capable and there’s going to be different tooling very soon? So you don’t want to overfit too much to the moment.

A reasonable view of a modern company is that all of its data is exposed in real time, and you have some tool on top, like Cursor or something else—maybe different tools for different skills. The licensing team at Spotify may have a different tool to reason over all the contracts and quickly say, “Do we think we can do this in that market, and what do we need to license to do this?”

But also, the product team could ask that licensing engine. We have 15 years of contracts, both current and previous, so this AI has a lot of insight into what music licensing looks like—more than any single person in Spotify, if you train it that way.

There will probably be slightly custom interfaces for different skills. I'm not sure which is going to win out, but I think it's going to look something like that.

Mm.

Right now, what we see people doing is sharing examples of prompts they used for the workflows and then prototypes that they've used, and that feels very much like a point in time. It's kind of hacky and involves different things.

If you were to calibrate the world out there, so few people have the inside view that you do, where you're excited by this technology and trying to embrace it. You're only able to embrace it so fast in the ways that we've described. On a 1-to-10 scale, what score would you give how much this is impacting you so far and how crazy this might get?

People are very excited that this is going to literally change everything, and there are some people who are actually worried about how much and how powerful it might be. From a practical, real-world standpoint, can you calibrate us a little bit as one of the few people who is both excited about it and also faces reality on a daily basis?

6. AI Impact Starts Small

If you want to be as realistic as possible about it, you take the developer use case. I've seen studies from other big companies that, if you actually measure it against a developer's time, the speedup is 7% or something.

Coding is a small part, net new is a small part of that, and so forth. So I think right now it's a bit overhyped in terms of actual impact, at least for these big companies. But I think it's going to turn into the opposite. Right now, people are overexcited relative to the actual impact, but I think the opposite is going to happen. I think it's going to have a tremendous impact over the longer term.

What I see people doing right now depends. I use it a lot personally. I see a lot of my developers, product people, and designers use it all the time for productivity purposes: putting things into an engine, asking it for a summary, and so forth. Those things happen all the time.

It's hard for me to estimate how much that speeds them up already. It certainly does. But I think the really big impact comes as you reshape these companies around this technology. Right now, we're just tacking it on top. But as I said, you have to reshape and rebuild the company for this work, where a reasoning engine can reason in real time over the entire company's data. That requires a lot of retooling. That's why startups are ahead: they don't have to rebuild.

There's a data—

They don't have 15 years of data. So they're probably in the future, which is why they feel like, "No, no. Gustav is wrong. The impact is really big already." And I think it is for a startup. I think it takes a bit longer for big companies.

Big companies like us have to shape up and accelerate in order not to fall behind.

I'm sure they would all like to have 700 million monthly active users to experiment with, though.

On that topic, you mentioned going through mobile and the experience that not only was everything changing as a result of mobile, but the business model also needed to change. We've really talked about product so far, and there's more to ask about product, but talk about the business model.

What would be the world in which, as a result of this technology, Spotify's whole business model needs to change? How do you go about evaluating something like that?

7. AI Changes Business Economics

We've seen a few examples of those business models, and I tend to tell my product teams that everyone says the world is disrupted and changed by technology. I think that's true in the sense that the underlying force is technology itself, and it's this gift that keeps on giving. It gives you computers, the internet, smartphones, ML, AI, quantum computing, and these gifts keep coming almost on a schedule. They actually come closer and closer.

Previously, technology companies were not called technology companies, as a side note. They were called car companies, but they were technology companies, or pharmaceutical companies. That was the state-of-the-art technology right then. But because these macro waves came so far apart, they called themselves a car company. They never became ubiquitous technology companies; they overfit to that.

I think somewhere in the 1990s, around Google, Amazon, et cetera, these macro waves started coming so fast that people tried to pin them down. Amazon is a books company, and they were like, "No, not really. We're doing books, but here's other stuff we're selling." And then, "Okay, you're the everything-store company." It's like, "No, not really. Now we're selling Amazon Web Services over here."

So I think these companies are the first set of companies to have technology as the strategy. The previous ones took one wave as the strategy, and IBM comes along and does computers as a strategy, or first mainframes and so forth.

I think this is the first wave of general technology companies, which, interestingly, might mean that there could be—I mean, companies almost always die after a while. These could be the first companies that never die because they're ubiquitous technology companies. Whatever the technology gift is, just try to have a company that can quickly wrap around it, figure out the product and business model. So I think that's interesting, and that's how I think about Spotify.

Yes, we're a music company, then a podcast company, then a book company, and a video company, but it's really about trying to anticipate technology, figure out what it can do, and then adapt the product and often the business model.

I said that mobile was one of these things where we needed to change the business model. I think what happens when one of these technology gifts comes along is that there is a big change when the technology happens. Piracy caused big havoc, but the real change happens when someone also figures out the business model.

So I tell my product teams, "Technology can do good things, and a new business model can really change the world." But without a business model, there's seldom large-scale change. You can destroy a lot of things, but you never really create value.

Mobile was the first, where we needed to figure out the free tier on mobile without cannibalizing our paid tier. What we did there was look at our data and see that 50% of premium users were listening in shuffle mode. So we said, "What if we take shuffle as a feature and give that away for free?"

It should be 50% of premium consumption, so very valuable, but it's not going to be 100% of anyone's premium consumption, so there would be no cannibalization. We managed to create a tier where you could playlist all your favorite songs in a playlist, press play, put the phone in your pocket, and listen forever for free in the background. So that was a business-model innovation along with technology.

The most recent one was audiobooks, where there were audiobooks in the US à la carte. Sure, we did some nice innovation around being able to stream that book, but the real thing is not streaming a book. You've been able to stream audio for a long time. The real innovation there was the business model: being able to bundle audiobooks into Spotify Premium.

It's almost like music. Music was also à la carte and quite niche, and once we made it an access model with no marginal cost, it got much larger. That's what we think about with audiobooks as well. So we've seen a few of those and managed to adopt them.

To your question, is AI going to do that? Do we need to change the business model? I'm not sure. I think there is one glaring thing that is different, which is that the previous technology model, coming all the way back from chips and silicon, was that you make a big upfront investment, then you amortize it, and you get to almost zero marginal cost. That's how software worked. That's not how AI works. The marginal cost is high, and you need to cover it.

You could say that should change everyone's business model. You're going to need to somehow either monetize very effectively through ads or charge users. You see OpenAI being a subscription product, and I think you're going to see more of those. So the marginal cost is a net-new thing.

For Spotify, it's interesting because we're the one technology company that always had a marginal cost. One more stream had a marginal cost to labels. So we grew up in a world where, if we were too successful on the free tier, we could go bankrupt overnight, which was never true for Twitter or Facebook, and that's why venture capitalists said, "Just go crazy. Worry about monetization later."

Spotify could never do that because we could go bankrupt overnight. So we always had to worry about monetization and the balance between free-tier and paid-tier conversion and free-tier monetization. The good thing for us is that we're fairly used to marginal cost in our business model.

I think you're going to see those things. It's very likely that some consumers are going to want tons and tons of inference, and because that's a marginal cost, you're probably going to have to pay somehow for that as a consumer. So I think you're going to see more tiering of consumer products based on how much inference you want. But for us, that's not that new. We've had several tiers already.

I'm curious because I'm an investor in a company called Etched that's going to be one of these companies that pushes down that inference cost. Like the history of compute, you're going to see this incredible consumer surplus and consumer benefit that comes from cheaper and cheaper unit-by-unit inference costs.

But the countervailing force is that we would just use more of it—more reasoning tokens, more whatever. So it makes me wonder how much more you can imagine better models being useful to you. It seems like if we just froze reasoning and model capabilities today, we'd probably still have a decade-plus of digestion to do around how we could use these models to make better products, better features, and whatever.

Can you imagine another couple of orders of magnitude of better models opening up lots of features that you can’t currently do? Is that a thing, or do you think we have what we need, and therefore inference could be really cheap?

I both subscribe to the product-overhang idea—that there’s a huge product overhang—and if we froze, I think we would see products ship that look amazing for several years before we exhausted what we have. So I subscribe to that, but I also subscribe to the idea that there is no limit for compute. Eventually, you get to computronium, but if you look at just the physics of computronium—

What’s computronium?

It’s the most computation a universe could theoretically do.

Yeah.

We’re very far from that limit, so I think we’re going to go all the way there before we stop, and I think we’re going to be very inventive. There is a nice analogy—I don’t know who came up with it, but I think Ben Evans talks about it quite often. You know, when the spreadsheet came along, the idea was, “Now all accountants are going to go out of business.”

What happened was we could just not imagine, if calculation cost went to zero, what was going to happen. You could imagine that the value of doing that was going to go to zero because there were so many accountants in the world. What happened was we just started doing massively more accounting. When there’s no cost to spreadsheeting, you’re going to start to do models to predict the futures of this asset or good or something into the future forever.

We just came up with so much more spreadsheeting that you could do, and it’s bigger than ever. I think from a financial point of view, when the cost of something drops, the demand usually increases more than the drop, and I think that’s bound to happen with intelligence. It is the ultimate thing, and to say, “No, I have enough intelligence,” is not interesting.

I think we’re going to be ashamed of how mundane the things are that we spend inference on. Could my coffee be one degree warmer tomorrow? If it’s truly no cost to ask the questions—

They’ll do it.

I think people will.

Maybe now is the time to ask you about sitting in the back garden with David Deutsch and talking to him about this concept of The Beginning of Infinity. Computronium made me think of your interest in this topic and your answer there that there’s no endpoint here. It’s just going to keep going. We’re going to keep learning and keep deploying our new technology. Can you talk about that book, why it influenced you, and your conversation with him?

Yeah. David Deutsch has been a hero of mine since I read The Beginning of Infinity, and then he wrote another book called The Fabric of Reality. He’s considered the father of quantum computing. Obviously, quantum computing is one of these gifts that technology is going to give us, and it’s about to get very real, I think, very soon.

I’ve always been interested because quantum mechanics is the most insane thing on this planet. We live in what we consider this reality, but if you go to the bottom layer, this is not reality; it’s just some three-dimensional projection that we live in. The Fabric of Reality had a big impact on me. I’m an Everettian. He believes in a many-worlds scenario. I read that book and it blew my mind.

Then The Beginning of Infinity is maybe his most famous book, whereas The Fabric of Reality is really about quantum computing and how a quantum computer works. The Beginning of Infinity is very philosophical, and he has a bunch of big ideas there.

He’s a very positive person, and now, at 70-plus, I finally got to interview him in his garden in Oxford. He’s not in great health, so it had to be outdoors and distanced, and everyone is very negative on the future. There are so many problems: AI could go wrong, all things could go wrong, climate change. He’s actually very positive about the future.

He’s clear that there are risks, but he sees us going out there into the stars. I asked him, “Where do you think we are in a million years?” And he’s like, “Well, maybe we’re this far outside of the solar system, but not quite there.” He’s very certain we’re going to get there.

He’s a very positive person, and when I asked him about his life, he’s very content with his life. He’s very happy. So he’s just an inspiring person still at this age. I wish I would be like him at that age.

But this book has a few concepts that I’ve tried to apply at Spotify, and one of them is that he talks about the power of explanations. He thinks the human mind is infinitely scalable. He does not think there’s a limit to what we can understand because of explanations.

I think this is something that a lot of people—I agree with that, but a lot of people disagree. Certainly, there are things we could never understand. His view is, no, there is no limit to what we can understand. We are the only species that broke that barrier because we have explanations.

Other species have pattern recognition. They can do things and learn the pattern that this works. There’s some cultural transfer, maybe, of looking at someone else doing that pattern. A bird can see another bird. Some species can teach their kids, but they never produce explanations.

He has a definition of a good explanation. He’s very inspired by Karl Popper as a philosopher. Karl Popper is his house god. So he takes a bit from Popper, and he takes a bit from science. He says, obviously, that a good explanation has to be falsifiable, but he says a few other things that I think are obvious in retrospect, but not before.

He says that a good explanation has to scale. It has to have reach. What does he mean by that? He says that some explanations explain something quite locally. You can have an explanation about, for example, the sun revolving around the Earth, which explains a bunch of stuff, but it doesn’t scale to other planets.

A better explanation is to have the Earth revolving around the sun. It just scales better to different scales. So a good explanation has to scale up and down. A good explanation has to be compatible with all previous explanations.

But most interestingly, he says that a good explanation has to be hard to vary. I find this very obvious, but also very non-obvious to people. So what does he mean when he says a good explanation has to be hard to vary?

He means that, for example, if your explanation for the weather on the planet is that Thor is angry, so there’s thunder, it’s an explanation, but it’s too easy to vary. You can say, “Well, now someone else is angry. They also had a hammer.” It’s too easy to vary and get the same result.

A good explanation is one where, if you move one of the parameters, the entire thing is not predictive anymore. Then you’re probably close to the truth. I think this is so interesting because the problem with most conspiracy theories that people love is that they’re so easy to vary.

Mm.

You can just exchange that character for some other crazy person who did something crazy, and it’s still going to produce the same thing. So if it’s too easy to change people in a conspiracy theory, it’s probably not true.

I think that’s something very powerful: good explanations need to be very hard to vary. This is something I’ve tried to instill in my organization.

I think there’s an interesting meta point here, which is that people ask me, as a product person, how much of product development is magic and how much is science. I try to be provocative in saying, “I think it’s exactly 100% science and 0% magic.”

People get provoked because it implies that there’s no skill. I say it to provoke. What I mean is that certainly people are going to have pattern recognition in this neural network. They’ve seen a lot of examples. That’s what we call seniority.

People who have seen a lot of things are going to get instinctively to the right conclusion faster than others. So that is valuable, and I want lots of seniority. I don’t discard seniority, and it brings you a lot of value. You can save a lot of time and a lot of mistakes.

But the reason you call it magic is because that person can’t explain it. It isn’t actually magic. It’s just science. It’s just that you are not smart enough to explain yourself.

If you could think even further and explain it, and come up with an explanation for what you see the way David Deutsch does, it’s so much more valuable for the company. If you have a theory, instead of saying, “No, Patrick, my intuition is this. You’re not smart enough to understand it, so I’m not going to tell you. Just do what I say,” maybe I’m right and maybe I’m wrong, but it’s not very helpful for you.

When I leave the company, you’re going to take over and you’re like, “I have no idea why they did that.” You have to develop your own intuition and your own pattern recognition.

But if I come up with an explanation based on the psychological behavior of people—it’s like Kahneman’s loss aversion or prospect theory—I think people value losing something 1.5 times the value of getting it. Therefore, we should not just launch a feature and test it because it’s 1.5 times harder and more expensive to remove it.

Then you have a theory, and it can spread across the company in a week. Now everyone has that. So I really want to force people in my company—even if we see something working in an A/B test—I try to tell them that I don’t want to launch it until we at least have a theory of why it works, even if it’s super clear.

There’s a lot of pressure to launch it because there’s engagement value and monetization, but I want you to at least have a theory. Then, over time, the company builds up a consumer theory. If you have a strong consumer theory, then you can predict things that were very unlikely.

What David Deutsch also says is that pattern recognition will iteratively get you more on the same path, but it’s never going to jump all the way from the geocentric to the heliocentric model. Only an explanation can take you to quantum physics, which is entirely unintuitive.

No pattern recognition gets you to “maybe it’s a wave and a particle at the same time.”

What’s an example internally of a great explanation that then led to the geocentric-to-heliocentric type of jump? What’s an example of how that actually played out?

I think a good example that is public is the free tier that I told you about. We only had a paid mobile tier. You actually paid to get mobile access to Spotify. Now, smartphones are scaling. Users don’t have a computer, so we need a free tier.

The competition was YouTube. They were foreground, on-demand with video. The pattern recognition—the obvious thing—would’ve been to say, “Let’s do that. It’s proven.” But what we did instead, and this was specifically attributed to a person named Charlie Hellman, was to reason around it from first principles and say, “Okay, let’s look at our usage of Spotify.”

If we limited our license to the same thing, it only works in the foreground. As soon as you lock the screen, the music stops. How much of the listening is in the foreground? It turns out that back then, it was 9% or something. So you have 91% of the use case being in the background. We probably want to get something else. The user need there is probably background listening.

Then you look at the App Store. Is there a way to listen to music for free in the background in the App Store? The closest thing was Pandora, but that was radio. You could not listen to your favorite songs. So then we said, “We would like a consumer product where you can listen to your favorite songs with your phone in your pocket forever for free.”

The problem with that is that’s almost a premium use case. If we just launch that, it’s going to cannibalize our premium tier. So what do we do? Then we looked at premium usage, and we saw that premium users, about 50% of the time, were shuffling their playlists. They were using on-demand features—searching, clicking, and playing specific songs—50% of the time, but they were shuffling playlists 50% of the time.

Then we thought, “What if we take this? It seems to be something that, even when you have on-demand, you voluntarily shuffle.” It’s a big use case. We give that away for free. That should mean that none of the premium users convert back to free because they still want their 50% on-demand, but you’re giving a lot of value away for free.

So we tried to model a consumer need, reason around it, and came up with this shuffle-background tier that was very, very, very unintuitive. Even the people inside the company said, “That’s a terrible idea.” But we trusted the data, and I was even skeptical of it myself. I was like, “Look at this on-demand. Shouldn’t we try time caps?” A lot of people just wanted us to try long free trials.

But the problem with a free trial is, even if Nokia Comes with Music tried a year-long free trial, the user knew that if they started investing in playlists now, a year from now their playlist investment was going to disappear, so they never started investing. So we went with this shuffled tier, and this is what made growth explode. To this day, that’s our differentiation against the other services. It’s the only way to listen to music for free forever with your phone in your pocket.

Fascinating.

So that’s an example of theorizing and explaining rather than pattern recognition.

I’d love to talk about the evolution of the relationship with the music industry. It’s a company that has unquestionably wholesale changed music, which is so interesting and so cool. You’ve been here a long time. Thinking back to the early days, it’s amazing, the impact that it’s had.

From an investor’s perspective, one of the things that many were always keyed in on is just the gross margin of the business. How much of a transfer-pricing problem are you always going to have, that no matter how big you get, the music industry that owns the IP is always going to take the same cut of the pie? Talk about how you’ve thought about that change over time. It seems like it’s been both a good relationship for them, but also a very patient path for Spotify. Maybe just give us some insight into how it’s worked and how you think about it.

Yeah, for sure. I grew up, and Spotify grew up, in the era of piracy in Sweden, which was the worst market. There’s this famous quote from a UK label executive to a Swedish label executive around the early 2000s, where the Swedish label executive showed the P&L of one of these Swedish companies and said, “That’s not a business. That’s a hobby.” That’s how broken it was.

That’s actually why Spotify could happen, because the music industry was prepared to take risks in Sweden. I want to give a lot of credit to the music industry. They took a lot of risks with Spotify. Spotify took an enormous amount of risk—an enormous amount of capital risk. We MG’d a lot. We ate a lot of the risk, but certainly they took a lot of risk. I think the music industry certainly deserves the success, as does Spotify.

I joined in 2008. Somewhere around 2012 or something, I started saying that my team, the R&D team, and all of Spotify—we are the R&D department of the music industry. At first, people were like, “What do you mean?” And I’m like, “Well, look at it. It’s an entire industry that doesn’t have an R&D department.”

Mobile phones have an R&D department. It’s called, you know, Apple or Google. Everyone has a lot of R&D, but there’s no R&D spend in the music industry. I think that’s turned out to be true.

If you look at the trajectory, this year is the first year of profitability for Spotify since its founding. People say that there’s a lot of talk about Spotify not sharing enough of the revenue. We share about 70%. But the truth is, the other 30%, we haven’t kept. We’ve invested all of that in the music industry, and then more.

We were unprofitable for 15 years. We just invested, invested, and invested, so there was a ton of patience. At the same time, the music industry has been profitable and Spotify has been unprofitable. So I think it’s fair to say we are literally the R&D department of the music industry.

We invested and had losses for 15 years, and the music industry has been gaining profit. Now, that is not sustainable forever. We needed to get profitable. We can’t be the R&D department of the music industry unless we can have the best machine-learning engineers, the best product people, developers, et cetera. For that, you need to be profitable.

It turns out these people are expensive because they’re sought after. We are a very, very patient and long-term company, and we invested for a long time, but it was just time. About 2 years ago, we decided, “Now it’s time for us to become profitable,” to take control of our own fate in terms of being able to invest in ourselves.

So yes, we are profitable, but we’re investing almost all of that back into more people, more product, and more AI. Now we just have our own investment vehicle instead of having to ask private investors initially, or the street, for more money.

So that’s how I think about it: really, as the R&D department of the music industry. I think we’ve done a good job. This year, we paid out over 10 billion, and that’s up from 1 billion, I think, almost 10 years ago. It’s just steadily increased. The music industry is bigger than it’s ever been.

People still talk about the heyday of music—

CDs?

—the CD era.

Yeah.

The truth is, the music business is bigger than it was back then. So this is the best it’s ever been. More money than ever. The pie is both bigger and higher, but it’s also getting sliced up. That’s because more people take a shot, and it feels very wrong for us to say, “No, the creators up until 2020, they were good, but no one should be able to try after 2020.” New creators should be able to try to do music.

So that’s the dynamic. I think a way to think about this is that people talk about the per-stream payouts and so forth a lot, and Spotify should share more per stream. There are 2 things that are happening when other companies say that they share more per stream.

The industry doesn’t pay per stream. They pay per subscriber. We have more than twice the engagement of our competing services. So if you take the same $10 and listen twice as much on Spotify, the per-stream rate is half.

These other companies have higher per-stream rates because they have a worse product. We’ve learned from the labels that we have twice the engagement and half the churn of competing services. That’s a curse of the per-stream model: the better we are as a product, the lower the per-stream rate is going to look.

But we’re looking at the aggregate number, and we’re leading everyone else there, where the vast majority of these payouts are. So I think if you look overall, the model is working. We took a lot of investment, and now the industry is getting a huge return. Spotify is also profitable now.

The way to grow this pie is that we are closing in on 300 million paid subscribers and closing in on 700 million MAUs. There are about 500 million paid subscribers, I think, in the world. We’re almost 300 million of those. But that’s half a billion out of the world’s population.

If you look at markets like Sweden, on average—we can just look at the public numbers—we convert about 40%. But if you look at the mature markets, I won’t give you the exact number, but it is much higher. If you look at the emerging markets, it’s lower. So the average is 40%. That’s not the average across the world. That’s a blend of low- and high-converting markets.

So far throughout our history, everything starts to look more and more like Sweden the more time passes.

Mm.

So the solution to this is just to scale it faster: a better free tier that gets more people on and converts them to premium. We think there should be billions of people paying for music, and that’s how you make the pie truly bigger.

The rev share is actually a red herring. So let's say that we share 70% today-ish, or let's say two-thirds to make it easier. Even if we were a charity and we paid out 100%, that would only be 1.5x what you get today. So if you think x pennies per stream is too little, even if we were a charity, it would be 1.5x.

The solution is not the rev share or giving away the vast majority. The solution is to quickly scale the number of people paying for music. If you just look at the numbers, you have to keep going, and it's going to get to billions of users paying, and then several billions. Then the music industry is absolutely massive. I just think the music industry is undervalued. Ultimately, it's going to be much bigger than it looks.

I'm curious how you're thinking about the podcasting world. This is something that we're sitting here doing right now, and I've been doing it for a long time. Now it seems we've entered this interesting new era where, when I started doing this, I remember it was quite—I would call it low-status. When I told people about it in 2016, they either didn't know what it was or thought it was silly.

And now, especially in the US, with what happened around the election and the importance of podcasts in the election, it seems as though it's hit some tipping point where basically anybody that might make sense to have a podcast now has one or is launching one. The corporate marketing strategy is to get a podcast, and the communication strategy is to go on them. So it's really exploded in importance and visibility.

What role has Spotify played, and will Spotify play, in all this? And what do you think about podcasting and its importance?

8. Spotify Builds a Multimedia Bundle

The reason we went into podcasting—one thing that I'm very precious about when it comes to Spotify, and so is Daniel, and the other co-president, Alex, who's my closest partner—is that there are many ways we could go as a company, and I think your business model, to some extent, steers you.

If you're an advertising business model, mostly you're going to be steered towards any additional engagement. Fortunately for us, we're mostly a subscription-based business model, so we focus more on retention, and you're going to vote with your wallet every month if you want to keep paying for us. We're not as steered towards engagement at any cost.

Having been at Spotify for a long time when this happened, one of the things that made me feel very good about Spotify was that when people used it, when they lost an hour on Spotify, they felt very good about it. If you lost an hour on music, you came out feeling that was a good hour.

One of the reasons I pushed quite hard for podcasts in the company was that I was using them myself, and a lot of our developers were using them. I saw them being hacked into the product at Hack Week every year. People wanted them there, and we just said, “Our developers are a small sample of the world. What if they're a good sample of the world?”

So that was one push. We saw people using it internally and hacking it. But what made us decide on it was that it was this format. Everything in the world was getting more and more short-form. People were consuming bite-sized content, and attention spans were going down.

There was this counterforce, which was long-form, deep discussions. People spoke in full sentences about quantum physics or whatever, and that just felt like something very important and good for the world. So we looked at it, and we saw that it seemed to be growing from a small base. We saw the biggest competitors being asleep at the wheel.

We did basically the Peter Thiel idea that it's better to go after small markets early and bet on organic growth than to try to take a small share of a mature market. It looks less risky in the mature market to get 1%, but the thing people miss is that the customer acquisition cost in a mature market is just massive.

Mm-hmm.

Whereas the customer acquisition cost in a new market is usually small. So we decided to go for it because we thought it was something that was in line with music. If you lose an hour in a deep podcast, you come out feeling like you learned something.

Mm-hmm.

And this is the reason we also went into books, because it's in line with that. We want to be this nutritious service. And there are 2 litmus tests for this. One is: if you lose an hour on Spotify, how do you come out feeling, versus if you lose an hour doomscrolling in the bathroom, how do you feel about that?

In one case, you feel like you ate a lot of candy, like you had a lot of energy in you, but it's bad calories. In the case of Spotify, you feel like you learned something.

The other litmus test that we have is that you see a lot of parents restricting screen time for their kids and saying—

“That's Spotify, yeah.”

“Go to Spotify instead,” which expresses how they feel about it and how we feel about it.

So that was one of the reasons to go into podcasts. It was partly philosophical, but we also saw the market opportunity of a small market that was poised to grow. We saw the need from early adopters trying to hack it in. Then we made this bet on leveraging our own distribution, combining it with music, saying that the market is this big right now, but what if we could expose podcasts to people who listen to music? Could we grow the market?

So that's the bet we made. And the truth is, audiobooks are something similar. In the US, audiobooks were a very niche behavior—10 or 11 million people, or something, who paid for them à la carte. The idea was that that's a limitation because of the business model. When you pay à la carte, you're not going to explore new books at $15 per book cost.

Just as in music, when you pay 99 cents per song, you're not going to soundtrack your sleep. It's too expensive at 99 cents per 3 minutes. But what if we have a no-marginal-cost model, where you can just explore? Is the audiobook market much bigger than it looks? Is the business model what's wrong?

So again, it was seeing a market that looked pretty small, but you can see in the Nordics, where you have the access model, that it's getting very mainstream. So it was a bet on the market, but it was the same philosophical discussion: Are these good calories? Is this nutritious? Is this in line with Spotify's mission of being the place where you go when you want to feel good about yourself instead of when you want to feel bad about yourself?

I remember the very first time I ever talked to Daniel, walking along the West Side Highway here many years ago. He talked about this notion of Spotify needing to be better than free.

Yeah.

If you think about podcasting, it's very different from music. When someone listens to this show on Spotify, you don't owe me anything. How do you think about the way that podcasting—and then, obviously, books are maybe a little bit more like music—and I'd like to hear how you think about it?

How do you think about the supply of the stuff that people are listening to on Spotify or watching on Spotify, the ways in which that affects your business model and the bundle?

We didn't know before we started whether podcasting, and later audiobooks, would be cannibalistic to the other media types or not. But it turns out they are not.

The easiest model to think about is that Spotify is a bundle now. You pay some price, or you have the advertising-based tier, and you get a bunch of value. Our job is to try to increase the value you get, so you value it more. Then, over time, maybe we can capture some of that value through price increases.

We've raised prices a few times, which is part of why we're profitable now. But that's because we had user surplus in value. That's because we kept just stacking value.

Value is 2 things. Value is features like personalization and just a really good product, but the other value is different types of media. What we see is that if you have a user who uses music, they have a certain amount of consumption. When you add podcasts, it's just more. It's not more—it's an infinite game, it looks like, at least for now. We haven't run out of time in the background yet.

Then when you add audiobooks, it's just more retention, more time spent, and more willingness to pay. So that's how we think about it as a business model.

Then, on the back end, they have very different business models. Music is a pool-based royalty model. Podcasts, as you know, are largely advertising-based, but now we also have this Spotify Partner Program, where you don't have Spotify ads in the premium tier if you're paying, so you get more uninterrupted listening. So that's another business model, which is part of the premium bundle.

Then you have audiobooks, where the publishing industry works in a third way, very differently. We also have a certain amount of time included in the premium tier and then a top-up if you run over that.

One of the really complicated things about Spotify that I don't think is appreciated is that on the front end, it's one app, one consumer, and you just go between them. But there are very different implications of where you click in that UI in terms of triggering different business models and so forth.

So to model a company financially is actually quite hard. We have to predict your user behavior. Where you click matters, and we have personalization that has different impacts in terms of cost and so forth.

We've had to build a system—we call it the Spotify machine—and that's why I said I have 1 experience organization. The job of this experience organization is to make sure that all of this complexity, all of these teams that theoretically could be set up to compete with each other to fix their P&L, never ships to the user.

There's 1 person who is the responsible person for the consumer experience, and that person's job is to make sure that as you go between mobile and desktop and car and speakers, the thing makes sense.

It's like the gatekeeper against the org, holding them back from the user, behind them, protecting the user. But it's also the same in personalization. I have a personalization organization because you have the same incentives of programming music versus podcasts versus books. Everyone wants to take market share and so forth. So it's the same problem. We have to optimize for the user and protect the user from the internal incentives of teams and business models.

That makes Spotify a pretty unique company. We're one thing on the front end, and we're many different things on the back end with different products.

If you think about, let's say, 5 years from now and you dream as big as you can possibly dream for where Spotify might go from where it is today to where it will be in 5 years, paint us that picture.

Certainly, I hope we've cracked the 1 billion-user line, but as a subscription, I hope we're becoming one of the biggest media subscriptions in the world and we add more and more value to that. Hopefully, music is bigger than it ever was. I'm hoping that audiobooks are a mainstream phenomenon, as they are in Scandinavia, where there are almost as many people who listen to audiobooks as listen to music. I think that would be a net good for the world.

But I also hope we've added a few more of these verticals. I can't say what they are. With the subscription model—the bundling model that we didn't talk so much about—we can differentiate on product or on content, but largely, we tend to license commodity content. We don't work with exclusivities, at least not anymore. We tried it in podcasts for a while.

So you can differentiate on the product and consumption of the commodity content, but you can also differentiate it on the offering. For example, if you look at Spotify now versus other offerings, some other offerings have the same music. Some other offerings have some of the same podcasts. You can't really find the combination of music, podcasts, and audiobooks. That's a unique thing.

So to use bundling theory to create more and more of a differentiated, unique thing that is Spotify, I think is very exciting. I think you will see more innovation on the bundling business model in addition to the product. I'm the product guy, but I'm very interested in business models. I've been a CEO myself, so I think you'll see a lot of innovation there.

What's the key to a good bundle? I'm also curious: You said you experimented with exclusive content that was only available on the platform, and there's less of that now. What drives a decision like that, and how do you think about other people that might want to create a bundle somewhere else?

When we looked at podcasts, you look at something like Netflix and this beautiful business model, with insanely good execution on top of that, and it looked to us like that could be interesting. I think when you're a product company that works with commodity content, you always have this envy of, what if we could differentiate through content? Then life is going to be super easy.

You always think the other thing that someone else is doing is easy and your thing is hard, and it's usually very hard to do the other thing. So we tried exclusivity in podcasts as a way to differentiate the service, but I think it was ultimately a bad bet because the macro trend for the whole thing with podcasts was that the production cost was so low. Joe Rogan was initially sitting in his trailer. The production cost was low, and then going in and doing exclusivities on top of that is kind of counter to the purpose in a way.

The whole point is more like YouTube, in that this is very cheap content, so you can get a lot of it. You don't have to be right. As soon as you go into the exclusivity game, you have to be right. You have to be a content picker, and that's a very hard skill that Netflix does extremely well.

But we had this opportunity. We didn't have to pick content. We could just get all of it and use machine learning to serve you what you wanted and me what I wanted, and there wasn't this capital-intensive need there that there is in producing costume dramas. It was a bad strategic decision that we made.

We also bet a lot on celebrities, and they are celebrities, but they're not always good podcast hosts. The podcast hosts that were really good grew up through this organic system. So there are 2 ways to always be right. One is to always guess right. The other is to just change your mind whenever you're wrong.

So we decided to change our mind and say, "This looks like the age of syndication." Creators actually want to be everywhere. They create a video, or they create music, and they want to be everywhere. Okay, let's embrace that.

In music, we were always a platform. We never played with exclusivity. We said we wanted the maximum catalog. In books, we're doing maximum catalog. Let's just embrace that in podcasts as well. So we pivoted strategy. It saved us a lot of cost, which is part of what we're doing well, and it also improved the catalog greatly. Now we're on a really good trajectory with our podcast viewing.

So it was an example of bad strategy, and I think the important thing is to admit it and change your mind. The real cost is when you try to defend your past decisions.

Stick with a bad idea. What things do you do outside of Spotify in your life that most prepare you, or make you capable, to do the best job that you can in Spotify?

I think the world is moving very fast, so a lot of my time is just trying to keep up with what is happening. I was on vacation in Lisbon with my family recently, and I spent a lot of time with them, seeing Lisbon, which is a beautiful city. But then I asked them for 1 day off from work and from the family to indulge myself.

This time, I was going back to trying to code a bit, use all these new tools, and stay on top of what's happening. Sometimes it's reading physics or math or something. It's a combination of keeping up with what is happening, which is hard because it's moving so fast, but also stimulating myself mentally.

I have to have something that I'm excited about at any point in time, and it can be new things, like AI and what it would mean, but it can be age-old things that I just didn't know, like learning more about physics or math or something. I've read a lot of philosophy for a while because it's just an interesting area.

You think through all the big questions of intelligence and consciousness and all of those things, and you can spend 10 years there just reading all of that. Now I feel capped out a little bit. When you start reading, you think, "Yeah, I'm gonna crack this," and then it turns out I didn't crack it.

People have been thinking about this for a while.

People have been trying to crack consciousness for a while, but it's so deeply interesting. It kept me excited about life for a very long time. I was never a big math person in school. I was okay, but not great. But I found myself getting very excited about math the older I got.

As you start reading a bit of philosophy, you get into things like Gödel's incompleteness theorem and constructive mathematics, and these things are loosely related to work.

But they keep you energized.

But they keep me energized. Actually, it turns out that a lot of my product people and engineers are deeply interested in these things. So I just have something very interesting to talk to the people around me about.

And then I do sports. I do Brazilian jiu-jitsu with my kids, which is very rewarding.

What has that taught you?

Humbleness. You come in and you think you can do something, and you get absolutely smashed by someone half your size—and they're not even sweating. And you're like, "Okay, technique matters."

It's technique. It's leverage.

Yeah. The beautiful thing about Brazilian jiu-jitsu is that the belt system is real. There's a long story behind it, but the net is that a Japanese person brought jiu-jitsu to a Brazilian family. There were a bunch of brothers there who fought a lot. There was 1 brother who was just underdeveloped compared with the others. He was just not very strong, so he could not beat his brothers.

So he started taking Japanese jiu-jitsu and figuring out how he could use physics—just leverage—and slowly, he started beating all his brothers, and that became Brazilian jiu-jitsu. So literally, he had to solve the problem. He could not use power.

Wow, I didn't know that.

And then this family put on all these competitions. What I like about it from an evolutionary product point of view is that they said, "Okay, anyone can come here. Karate, kickboxing, just try it. It's open." They fought in these basements, just evolving the sport and proving that it was real.

A lot of martial arts are magic and secret, and they never test their skills. It works very well in practice. The other thing I like about it is that I've done a lot of other martial arts—boxing and Thai boxing and stuff. Those things are great as exercise, but for self-protection, they're not very good. You cannot punch someone in the face. You're gonna get sued.

For self-protection, the beautiful thing about the martial art, which is called the gentle art, is that you control people and constrain them, and you can adapt the level of violence. This is why police use jiu-jitsu and not Thai boxing.

Hmm.

Because you can regulate the violence to the other person. You can control them without hurting them. So that's why I think everyone should use it—

Hmm.

And practice it. It's good both for self-discipline, because you get humble. It's also actually useful, and you can use it without harming other people.

Spotify—you and Daniel especially—have probably been the most influential people, and certainly Spotify the most influential company, on me and on how I've thought about building our businesses over time.

And a lot of that comes back to the stuff that you don't see. I've purposefully talked about a bunch of it today with you: the bets board, the complexity that's hidden behind a beautiful consumer experience. You were the first person, years ago, to describe the bets board concept to me, and we've used that very effectively. And there have been so many lessons from Daniel on how to think about what matters to users.

I think Spotify is not only an incredible product, but it's also one where the product is a reflection of the company behind it.

Yeah.

I think it's one worth studying by listening to conversations like this one, because for me, what it's done is raise the bar of ambition and the standard of excellence for how a company should be constructed to mirror its needs—its unique needs—but also the character and discipline of the people running it. It's been so fun to do this with you, and thank you so much for all the lessons over the years. You know the closing question that I have for everyone: What is the kindest thing that anyone's ever done for you?

The thing that made me really excel in my role was being allowed to take a lot of risk by Daniel. I've actually screwed up a bunch of things at Spotify that didn't work, and I never felt that I was going to get fired for it. He actually encouraged that, and I got a second chance. That's what gave me higher ambition instead of holding back for fear of failure. So I think it's a series of those things—being allowed to mess things up—that has probably had the biggest impact on my professional career.

Can you give an example of a bad mistake that you made, and how he and the organization made you feel through that process so that you could be re-emboldened to take more risk again?

9. Second Chances Create Bold Leaders

I was interested in new user interfaces. Many years ago, I took the company very hard on a journey for an interface that, at the time, was very provocative. The idea was that Spotify just starts playing things. You swipe up to get to the next genre, and you swipe left or right to get other things within the same genre. Now you would say, “That sounds almost like TikTok.”

Mm-hmm.

This was before Musical.ly. But two things happened. It was very provocative: it started playing things without you asking, so people were upset. But I pushed pretty hard because I was convinced that immediacy—the idea that you just swipe your way to what you want to hear in a very low-friction interface—was the future. It was maybe a decent idea, but it was before machine learning.

Mm.

It just did not work at all. You could not get there. We built this—it was called Moments, the UI. We used editors on the back end, which just did not work at all. So the idea was far, far ahead of where the technology was.

It cost a lot of money. We actually announced it. There's a video of us presenting this user interface and so forth. People luckily forgot it, but it just didn't work. We had A/B-tested it, and it looked okay, which is why we launched. Then we discovered there was a bug in the A/B test—

Oh, my God.

—when it was live, and it actually underperformed drastically compared with what we had. So we had to roll it back, and I had taken the entire organization on this excursion that lost a year or something—

Mm.

—in a very competitive business. That was a good opportunity to get fired, and I didn't. Daniel Ek was like, “I understand. I agreed with the thoughts and the ideas. What was the mistake?”

The mistake was that the machine learning was not there. We were not good enough to get you there in enough swipes. He was more like Jeff Bezos, who measures the inputs, not the outputs. If the inputs are bad—if the ideas are fuzzy and stupid—that's a problem. But you're not always going to be right, even with good ideas.

I heard him say this Jeff Bezos quote: “I judge you by the inputs you had, not the outputs.” The problem with judging the outputs is that you could just get lucky, and you get promoted even though you're not very good, just by luck. Whereas if you look at the inputs, you give more chances. If the ideation and execution are structured, you're eventually going to get it right.

So I just got more chances. That made me actually take more risk instead of scaling down on the risk. But I felt very, very, very burned for a long time. There are jokes internally about Moments.

What a powerful story and mindset for us all to adopt. Such a great closing story, Gustav. Thanks so much for your time.

Thanks for having me.