Gustav Söderström - How Spotify Thinks - [Invest Like the Best, EP.424]
What this covers
This is a conversation with Gustav, Spotify's Co-President, Chief Product Officer and Chief Technology Officer. It covers how the company decides what to build and how he thinks AI changes consumer products. Much of it concerns process. Spotify runs as a synchronized company because it chose a super app strategy, so every product has to ship in one app. Gustav also describes refusing to launch a feature that wins an A/B test until a team can explain why it works. That standard draws on David Deutsch and the idea that a good explanation is hard to vary, and on Daniel Kahneman's finding that losses weigh about 1.5 times more than gains, which makes removing a feature costly.
The AI material is concrete. Gustav describes AI Playlisting as the first real signal of what is in a user's mind. He puts measured developer speedups at big companies at around 7 percent, and says the larger gains come from reshaping workflows. He argues there is a product overhang even if model capability froze today. He notes that AI carries a high marginal cost per use, unlike classic software. He also makes the contested forecast that cheaper intelligence will raise total demand, as spreadsheets did.
The last stretch is about Spotify's business. He explains how a free background-shuffle tier came from finding that 91 percent of listening was in the background. He argues that paying artists 100 percent of revenue would lift per-stream pay only about 1.5 times, so growing paying subscribers is the real lever. He also describes the back-end complexity of bundling music, podcasts and audiobooks, which each pay out differently, and recalls that Daniel Ek judged him on the quality of his process rather than on outcomes.
Söderström argues that generative AI is a general-purpose technology shift that forces consumer products toward two-way conversational interfaces and requires big companies to expose their data and reshape their organizations, while Spotify's synchronized planning, explanation-driven product culture, and bundling strategy are how it adapts.
- Generative AI widens the user-to-product input channel from clicks and skips to natural language, so consumer products will become conversational
- Big companies gain less from AI than startups until they expose their data via real-time APIs and rebuild their workflows around reasoning engines
- Spotify's synchronized six-month bets process and bundling model exist because its single-app distribution strategy requires them
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Oberholzer-Gee's value stick adds willingness to sell, the price at which suppliers including employees will provide inputs, to willingness to pay; companies can grow value by lowering what they must pay for talent by offering mission and culture.
“the best companies in the world are not necessarily the ones who actually pay the most. It's the ones with the most interesting mission, the best culture”
Spotify is complex on the back end because music pays through a pool-based royalty model, podcasts are largely ad-based, and audiobooks have included time allowances with top-ups, so one front-end app masks very different business models that the company has to model together.
“music is a pool-based royalty model. Podcast as you know, is advertising-based largely.”
When the cost of a computational resource falls, total demand rises more than proportionally, as the spreadsheet showed by creating vastly more accounting and modeling rather than eliminating accountants, so cheap intelligence will generate far more use.
“from a financial point of view, when the cost of something drops, the demand usually increases more than the drop.”
Pattern recognition only iterates along an existing path, so transformative leaps like geocentric to heliocentric or quantum physics require explanation, which can produce entirely unintuitive conclusions that patterns cannot reach.
“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.”
Deutsch holds that the human mind is infinitely scalable and that there is no fundamental limit to what humans can understand, since explanations allow a break beyond pattern-recognition species.
“He does not think there's a limit to what we can understand because of explanations.”
Unlike classic software, which amortizes a large upfront investment to near-zero marginal cost, AI carries high marginal cost per use, so AI businesses must cover inference costs through subscriptions or ads.
“The marginal cost is high and you need to cover it.”
Drawing on Kahneman and prospect theory, Gustav notes that people value losing something about 1.5 times more than gaining it, so removing a feature is far more costly than launching it, which argues for testing removal with explanation.
“I think people value losing something 1.5 times the value of getting it.”
Measured developer speedups from AI at big companies are around 7 percent, so current impact is overhyped, but the real impact comes when companies reshape their workflows around reasoning engines, which will be tremendous over the longer term.
“the speed up is 7% or something... So I think right now it's a bit overhyped in terms of actual impact”
A good explanation must be hard to vary, meaning that swapping any parameter should break its predictive power; conspiracy theories are easy to vary because the actor can be swapped for another with the same result, which marks them as likely untrue.
“a good explanation. If you move one of the parameters, the entire thing is not predictive anymore, then you're probably close to the truth.”
Even if model capability froze today, there is a large product overhang, and products that look amazing could ship for several years before current model capability is exhausted.
“there's a huge product overhang and if we froze, I think we would see products shipped that look amazing for several years”
Even if Spotify paid out 100 percent of revenue, rather than roughly two-thirds to 70 percent, the per-stream increase would be only about 1.5 times, so the real lever for artists is growing the number of paying subscribers.
“Even if we were a charity and we paid out 100%, that would only be 1.5x what you get today.”
The music business today is bigger than in the CD era, with payouts rising from about 1 billion to over 10 billion, and Gustav believes it is undervalued and will grow much larger as billions of people pay for music.
“This year we paid out over 10 billion and that's up from 1 billion, I think almost 10 years ago.”
Startups move faster because large companies' data sits in cold storage and is not exposed as real-time APIs, so the main bottleneck for an incumbent adopting AI is old-school engineering to expose its data, not AI engineering.
“my biggest job to enable AI is not AI engineering, it's old-school engineering exposing all this data”
The key shift of generative AI is that the user-to-product channel (uplink) can now carry natural language nearly as rich as the product's output, turning interfaces from one-way feeds into two-way interactions.
“the big shift is that it's two-way.”
Gustav argues that product intuition called 'magic' is actually pattern recognition from seniority that the person simply cannot explain; a good theory converts that intuition into knowledge the whole company can use.
“the reason you call it magic is because that person can't explain it. It isn't actually magic. It's just science.”
A subscription business model focuses on retention because users vote with their wallet monthly, whereas an advertising model steers a company toward any additional engagement, which is why Spotify does not chase engagement at any cost.
“If you're an advertising business model, mostly you're going to be steered towards any additional engagement.”
Gustav refuses to launch features that win A/B tests until teams have a theory explaining why they work, because a theory spreads across the company and builds an accumulating consumer theory that predicts unlikely outcomes.
“even if we see something working in a B test, I try to tell them that I don't want to launch it until you at least have a theory of why it works.”
Daniel Ek judged Gustav by the quality of his inputs, ideas and structured process, not by outcomes, because judging by outputs rewards luck and penalizes good process that failed, so more chances are given to structured experimentation.
“he was more like Jeff Bezos measures the inputs not the outputs.”
Spotify found that about 50 percent of premium listening was shuffle mode, so offering shuffle-only free listening on mobile captured background use without cannibalizing premium users who still wanted on-demand.
“we looked at our data and saw that 50% of premium users, we're listening in shuffle mode.”
Reasoning from first principles, Spotify found that 91 percent of listening was in the background, which led to a free background shuffle tier that counterintuitively drove growth, while free trials and time caps failed because users expected playlist investments to disappear.
“Turns out back then it was 9% or something. So you have 91% of the use case being in the background.”
Spotify adopted a Chinese-style super app strategy when app installs became scarce, reasoning that distribution was the hardest problem and that bundling podcasts and books into the same app leverages existing distribution.
“distribution became the most important thing. And then we chose ... we're going to build it in the same application because then we can leverage our own distribution.”
Pre-generative machine learning shaped consumer UIs as one-way output mechanisms that maximize engagement metrics via explore-exploit, with the full-screen TikTok feed as the epitome of that age.
“the epitome of that age is the full screen TikTok feed. UIs shape themselves after technology that powers them to maximize metrics.”
AI Playlisting, an LLM trained on listening data and world knowledge that takes English prompts, gives Spotify the first real signal of what is in a user's mind, which makes approximating the individual far easier than click streams allowed.
“So for us, it's the first time that we get that kind of fidelity of what is actually in the user's mind.”
Spotify needs a synchronized company because it chose a single-app super-app strategy to exploit its distribution, so every product must ship in one app and no team can run ahead independently.
“So because of our consumer strategy, the company needed to be synchronized.”
Spotify's E Team meets three hours weekly with all VPs in one room; the rule that nothing can be taken 'offline' or 'later' ensures blockers are resolved in real time, and blocking is capped at about two and a half days.
“you're not allowed to say the word 'offline' or 'later' because that person is in the room.”
Banning VPs from bringing direct reports to the E Team forces them to know the details themselves, and keeping the same group together builds trust that allows honest debate.
“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 to yourself.”
Non-developers increasingly use Cursor because the Model Context Protocol lets internal services be wrapped so people can speak English to infrastructure and prototype features without coding.
“if you take your internal services and you wrap them in an MCP, you can speak English to your infrastructure.”
Spotify positions itself as the R&D department of the music industry, having invested heavily in product and machine learning for about 15 years of losses while the labels profited, a model that cannot be sustained indefinitely.
“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.”
Spotify's exclusive podcast deals with celebrities were a bad strategic decision because podcasts are cheap to produce and exclusivity demands content-picking skill, so Spotify pivoted to maximum catalog and syndication, saving cost and improving the catalog.
“It's a bad strategic decision that we did.”
Audiobooks were a niche behavior in the US at about 10 to 11 million à la carte buyers; the breakthrough was bundling them into Spotify Premium with no marginal cost, which led to exploration and growth just as music grew after its access model.
“The real innovation there was the business model to be able to bundle audiobooks into Spotify Premium.”
Making leaders stack-rank initiatives, rather than declaring several equally important, forces explicit tradeoffs and prevents downstream VPs from fighting over conflicting priorities.
“if you as a leader don't bring clarity, you're going to set your org up for fighting.”
Technology alone produces disruption and piracy-style havoc, but lasting value creation only occurs when someone also invents the business model that monetizes the technology.
“without a business model, there's seldom like large-scale change. You can destroy a lot of things, but you never really create value.”
Spotify's per-stream payout looks lower than competitors' because the industry pays per subscriber, and Spotify's users listen about twice as much per subscription dollar, so better engagement mechanically reduces the per-stream figure.
“So if you take the same $10, you listen twice as much on Spotify. The per stream is half.”
The real cost in strategy is defending past decisions; there are two ways to always be right, guess correctly or change your mind quickly when wrong, and Spotify chose the latter.
“The real cost is when you try to defend your past decisions.”
Daniel Ek's willingness to let Gustav make and survive large mistakes, and his judgment based on inputs, increased Gustav's appetite for ambitious risk.
“being allowed to take a lot of risk by Daniel.”
Spotify has always paid per-stream royalties to labels, so it could go bankrupt from success on its free tier, unlike Twitter or Facebook, which is why it always had to manage the free-to-paid balance rather than grow at any cost.
“We grew up in a world where if we were too successful on the free tier, we could go bankrupt overnight.”
Entering a small emerging market early is less risky than taking 1 percent of a mature market, because customer acquisition costs in mature markets are massive while those in new markets are small, following Peter Thiel's logic.
“the customer acquisition cost in a mature market is just massive, whereas the customer acquisition cost in the new market is usually small.”
Generative AI lets a company run continuous deep qualitative user research across nearly 700 million users, substituting for the interview-then-A/B-test cycle that previously only small samples received.
“The promise of generative AI is really a deep, ongoing qualitative user research with almost 700 million users all the time.”
Every six months Spotify's roughly 14 VPs pitch proposed bets as if they were startups to VC-style decision makers; the co-presidents then produce a global stack rank (44 bets this cycle), and teams resource down the list as far as possible.
“they pitch, literally pitch, as if we were a VC and they were a startup.”
Resourcing every developer globally against the bets list would make planning intractable and disempower VPs, so Spotify divides resources into organizational blocks, lets each keep its people, and only moves staff at the margin.
“That planning process is hell if everyone is up for grabs.”
A single experience organization owns the consumer experience across mobile, desktop, car, and speakers so that competing internal P&L incentives across business units never ship a fragmented product to the user.
“It's like the gatekeeper against the org, holding them back from the user behind them, protecting the user.”
Adding podcasts and audiobooks to Spotify did not cannibalize music listening; each new media type added consumption, retention, and willingness to pay, because the bundle creates more total value.
“it turns out it's not. So the easiest model to think about it is Spotify is a bundle.”
The Moments interface, which auto-played music with swipe navigation, failed because it was ahead of the machine learning needed to deliver relevant tracks in a few swipes, and an A/B test bug led to rollout that underperformed and was rolled back.
“the mistake was the machine learning was not there. We were not good enough to get you there in enough swipes.”
Spotify can differentiate by offering a unique combination of music, podcasts, and audiobooks in one service, which is harder for competitors to copy than exclusive content, so bundling is the more durable differentiator.
“You cannot really find the combination of music, podcasts and audiobooks. That's a unique thing.”
Exclusive content demands being right about which creators to pick, a hard skill that Netflix executes well, whereas Spotify's maximum-catalog approach with machine learning avoids that risk by serving all content to each user.
“As soon as you go into exclusivity game, you got to be right. You got to be a content picker.”
Long free trials fail to convert users because users anticipate losing the playlists they invest in when the trial ends, so they never start investing; a permanent free tier avoids this.
“the user knew that if I start investing in playlists now, a year from now, my playlist investment is going to disappear.”
Spotify's main behavioral signal was skips, which are blunt: a skip can mean hatred, over-listening, or wrong context, so skip data alone can never achieve perfect personalization.
“All of those just look like a skip to us.”
Spotify gives business people technology understanding and gives product people and engineers full P&L and gross margin understanding, because a CEO perspective requires understanding all of the business.
“If you're the CEO, you have to understand all of it.”
The largest coding impact from AI is yet to come, because models will soon understand large codebases well enough to refactor them and perform automatic peer review, and the other seven hours of developer time (communication, planning, design, prototyping) may matter even more.
“the biggest impact is yet to come when it comes to coding.”
Companies that anchored their strategy on a single technology wave (cars, computers) failed to become ubiquitous, while recent companies like Amazon and Google treat technology itself as the strategy and may be the first who never die.
“these could be the first companies that never die because they're ubiquitous technology companies.”
Within five to ten years, almost all big consumer products will be conversations to some extent rather than services used passively, because information flows will become more symmetric between user and product.
“almost all big consumer products are going to be a conversation to some extent rather than this service that you use.”
Spotify converts about 40 percent of users to paid in Sweden, and markets trend toward Sweden's conversion rate over time, so global expansion of paying subscribers is the path to a bigger music pie.
“If you look at markets like Sweden, on average, you can just look at the public numbers, we convert about 40%.”
Computronium is the theoretical maximum computation a universe could perform; present computing is very far from that limit, so Gustav expects computing to keep scaling toward it.
“Theoretically, we're very far from that limit. So I think we're going to go all the way there before we stop.”
Spotify prototypes next-six-month product visions in Figma and generative AI before committing resources, so misalignment is surfaced up front rather than discovered late in the cycle as teams fight over integration.
“So all this so-called fighting happens before you actually commit to doing something and you have something you can hold in your hand.”
The drawback of a fully synchronized company is that planning becomes very expensive, and Spotify had to build its own planning tooling because external tools were not good enough.
“The drawback of that model is that the planning is very expensive. ... we've had to build our own tooling.”
Most large companies currently layer AI onto existing processes rather than reshaping them, which limits the impact until they rebuild around real-time reasoning over company data.
“Right now we're just tacking it on top.”
Companies should avoid building heavily around current AI workflows because models will become more capable and tooling will change quickly, so today's prompt-sharing and hacky workflows are point-in-time.
“So you don't want to overfit too much to the moment.”
When smartphones became dominant, Spotify's desktop-based free tier model died because consumers without desktops had no free experience, forcing a redesign of the business model around a mobile free tier that would not cannibalize paid mobility.
“And our entire model died.”
Spotify entered podcasting partly because everything was becoming short-form, and long-form discussion was a counterforce that Gustav believed was valuable for the world and matched Spotify's mission.
“which was long-form discussions, deep, people spoke in full sentences about quantum physics or whatever.”
Generative AI has so far had most impact on writing net-new code, but at big companies coding is about one-eighth of developer time and net-new code is a small fraction of that, so the current coding gains are limited.
“you basically code one out of every eight hours in a day. So not only is coding only one-eighth of the time, of that 1/8th of the time, net new code is very small.”
Spotify found that only around 9 percent of listening happened in the foreground, implying that roughly 91 percent occurred with the phone in the background, which justified building a background free experience.
“Turns out back then it was 9% or something. So you have 91% of the use case being in the background.”
When users build playlists, they are effectively labeling tracks in relation to each other and investing high effort, producing a rare high-quality dataset describing their taste.
“you're sort of labeling these tracks, at least in relation to each other and you're putting a lot of effort into it.”
Because some consumers will want large amounts of inference and it carries marginal cost, consumer products will likely tier pricing by how much inference a user consumes, as Spotify already does with its tiers.
“I think you're going to see more tiering of consumer products based on how much inference you want.”
Deutsch's criteria for a good explanation require falsifiability and reach, meaning it scales across domains and scales, unlike local explanations such as a geocentric model that fails for other planets.
“a good explanation has to scale, has to have reach.”
Spotify executes in six-month cycles because a year is too slow to react and a quarter has too much planning overhead relative to execution.
“We execute for six months in order for the overhead to not get too big, but we can't go to a year. ... A quarter is too short.”
Brazilian Jiu-Jitsu was created when a smaller, weaker Brazilian brother used Japanese jiu-jitsu leverage and physics to beat stronger siblings, which Gustav uses to illustrate that technique and leverage beat raw power.
“So he started taking Japanese Jiu-Jitsu and figuring out how he could just use physics, just leverage.”
Introducing named strategic frameworks such as Seven Powers, bundling theory, and the value stick gives teams shared vocabulary, which structures strategic thinking in product and technology teams, not just business teams.
“It makes people have a vocabulary. We can talk about network effects, amortization, brand power”
Spotify uses a litmus test: after an hour of Spotify users feel they learned something, whereas an hour of doomscrolling feels like eating candy, and parents restricting screen time direct kids to Spotify, indicating the product is seen as nutritious.
“In one case, you feel like you ate a lot of candy, like you got a lot of energy in you, but it's bad calories.”
Gustav argues that talk is cheap so organizations should talk more, using Socratic debate and structured ideation as a counter to move-fast-and-break-things, because reasoning has historically produced large advances.
“Talk is cheap, so we should do a lot of it.”
Spotify has never had enough people, which is why limited effort and resource allocation are central to how it decides what to build.
“There's never enough people.”
Spotify was well positioned for the internet era because it offered a free tier on desktop, which acquired users who built playlists and retained themselves, while mobility was a paid feature.
“we had a free tier on desktop. And that's where we acquired users and they built a playlist and they retained themselves.”
Stack ranking is simple conceptually but rarely done because leaders avoid saying one thing matters more than another; when pressed, they claim equal importance but then cannot rank.
“Very few people manage to say this is actually more important than that.”
Spotify's goal as a bundled service is to keep users' willingness to pay far above its price by continually adding value, creating consumer surplus that makes the service feel like a great deal.
“And the job of us is to keep the willingness to pay quite far from the actual price. That gap is how much consumer surplus you're giving.”
Spotify grew up in Sweden's piracy era, when a UK label executive reportedly called a Swedish label's P&L 'not a business, that's a hobby', and this brokenness made the music industry willing to take risks on Spotify.
“That's not a business, that's a hobby. That's how broken it was. And that's actually why Spotify could happen”
Generative recommendation systems are beginning to replace older collaborative systems, though Spotify's advantage has been the explicit labeling people did when building playlists.
“even though generative recommendation systems are starting to take over from these more old-school collaborative systems.”
Spotify reached its first profitable year after roughly 15 years of losses, and it reinvests nearly all profit into people, product, and AI rather than returning it to outside investors.
“this year is the first year of profitability for Spotify since its founding.”
Generative AI is a macro shift that companies cannot opt out of; like smartphones and the internet before it, adoption is not a choice, so firms must reposition themselves to ride the wave rather than resist it.
“This is one of those shifts that it's not your choice, whether you adopt it or not, it's going to happen to you.”
Spotify bet that exposing podcasts to its existing music listeners, using its own distribution, could grow the podcast market beyond its small base of deep listeners.
“what if we could expose podcasts to people who listen to music? Could we grow the market? So that's the bet we did.”
Because Spotify is synchronized, it can make global changes like redesigning the entire UI quickly, but it is probably slower than peers at trying small experiments because every change must pass through heavy planning.
“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 trying something”
Spotify's org is split into a platform organization, an experience organization owning the consumer product across devices, a personalization organization for AI and recommendations, and three business verticals (music, podcast, books).
“Then a personalization organization, because that's so important to us, that does all the AI and the recommendations”
Cursor's audience is widening beyond developers to designers, product managers, and business people, with Gustav's New Zealand-born PM using it to prepare Swedish tax filings through an MCP wrapper.
“she did her taxes in Cursor, managed to wrap the Swedish tax authority in an MCP, not a developer.”
In five years Gustav hopes Spotify will have crossed a billion users, be among the largest media subscriptions globally, make audiobooks mainstream as in Scandinavia, and differentiate through a unique bundle of music, podcasts and audiobooks.
“You cannot really find the combination of music, podcasts and audiobooks. That's a unique thing.”
Spotify's long-run aim is to reproduce a small part of each user's neocortex on its servers, which click-stream skip data made hard but natural-language input makes much easier.
“One way to think about Spotify is we always try to reproduce a small part of your neocortex on our servers.”
AI companies offer reasoning engines approaching AGI, some open source, at almost no cost, so organizations get capabilities they assumed would be prohibitively expensive.
“you get it almost for free. It's a gift.”
Spotify's leadership is small enough that all VPs fit in one room, and Gustav treats this smallness as an advantage over much larger competitors rather than as a constraint.
“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.”
Modeling Spotify's economics is hard because user click behavior determines which royalty or cost model triggers, and personalization affects cost, so the company built an integrated financial system it calls the Spotify machine.
“So to model a company financially is actually quite hard. We have to predict your user behavior where you click matters.”
Gustav says Brazilian Jiu-Jitsu teaches that technique and leverage beat size and strength, a lesson he finds humbling and useful for self-protection without serious injury.
“technique matters. ... you get absolutely smashed by someone half your size”
Unlike earlier shifts, AI touches both the consumer product and the company's own productivity and competitiveness, so firms face multiple simultaneous angles of change.
“It touches the consumer product, but it also touches your productivity and competitiveness as a company.”
Spotify does not yet know what the next consumer paradigm is; the product team's job is to experiment and identify it, and Gustav says he would not disclose it even if he knew.
“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”