Summary

Generative AI is an unavoidable “macro wind” that Söderström believes will fundamentally reshape today’s mostly one-way consumer interfaces into something conversational. Spotify’s old machine learning optimized a thin uplink of clicks, swipes, and skips; natural language can carry intent nearly as richly as the information coming down. His five-to-10-year wager is that “almost all big consumer products are gonna be a conversation to some extent,” though Spotify has not yet found the final paradigm.
Spotify’s most strategic AI opportunity is a new, high-fidelity stream of user intent at nearly 700 million-user scale. AI Playlist, live in 40 countries, lets someone request EDM with big drops at 160 BPM, retain good tracks, reject artists, and refine the result conversationally. Söderström calls the limit case “deep, ongoing qualitative user research with almost seven hundred million users all the time.”
Enterprise AI is overhyped in measured impact today but potentially transformative once incumbents rebuild around it. Söderström cites roughly 7% developer productivity gains: coding occupies perhaps one of eight working hours, and net-new code is only a sliver of that. The larger unlock is refactoring huge codebases, automatic peer review, and exposing 15 years of internal data through real-time APIs and MCPs—“my biggest job to enable AI is not AI engineering, it’s old school engineering.”
Spotify’s operating system converts strategy into an explicit, company-wide capital-allocation ranking. Roughly 14 VPs pitch 30-50 six-month bets—44 in the cited cycle—before Gustav Söderström and co-president Alex Norström stack-rank them and the organization resources downward, often reaching about 30. A three-hour Tuesday execution meeting prohibits taking issues “offline,” keeping any dependency blocked for at most two-and-a-half working days; the price is heavy planning overhead.
AI’s high marginal cost may force more consumer subscriptions and usage-based tiers, but Spotify has unusual muscle memory for that economy. Unlike traditional software, inference does not quickly approach zero marginal cost; unlike most internet companies, Spotify has always incurred another royalty cost when another stream plays and “could go bankrupt overnight” if free usage outran monetization. Söderström expects some consumers to pay for “tons and tons of inference,” even as falling unit costs trigger far more demand.
Söderström’s music-economics thesis is that scale, not a larger royalty percentage, is the industry’s meaningful upside. Spotify shares about 70% of revenue, says it paid more than $10 billion in the cited year versus roughly $1 billion almost a decade earlier, and has reached its first profitable year after 15 years of reinvestment. With almost 300 million paid users out of roughly 500 million globally, he calls rev share “a red herring”: even paying out 100% would produce only about 1.5× today’s royalties, while billions of paying listeners could expand the pie dramatically.
Podcasts and audiobooks strengthen Spotify’s bundle because their usage has proved incremental rather than cannibalistic. More media raises retention and willingness to pay, supporting price increases, while one simple interface conceals music royalties, podcast advertising and Premium payments, and audiobook allowances and top-ups. The five-year ambition is to cross one billion users, make audiobooks mainstream, and add more verticals to a differentiated “nutritious service.”
Spotify’s appetite for risk rests on admitting strategic mistakes instead of defending sunk decisions. Söderström calls podcast exclusivity a bad bet and says syndication improved the catalog while cutting costs: “The real cost is when you try to defend your past decisions.” Earlier, his premature Moments interface consumed about a year before a faulty A/B test was discovered; Daniel Ek kept backing him, with Söderström invoking a Jeff Bezos-style principle of judging “the inputs you had, not the outputs,” which encouraged more ambition rather than safer bets.

Deep dive

1. AI is a macro wind Spotify cannot opt out of

Söderström’s categorical long-term view is “AI or die,” just as earlier companies faced computer, internet, and smartphone transitions. “It’s not your choice whether you adopt it or not. It’s gonna happen to you.”
Spotify’s smartphone crisis supplies the cautionary precedent: desktop users entered through a free tier, while mobility was paid. Once consumers appeared with only phones, they had no free experience, threatening the acquisition engine and forcing a business-model redesign.
Old statistical machine learning was principally an output mechanism; its emblematic interface is the full-screen TikTok feed, optimized around explore/exploit using a narrow stream of clicks and swipes. Generative AI adds natural-language input and makes the relationship two-way.
Söderström compares existing services to “old-school broadband”: enormous downlink bandwidth but perhaps one megabit down versus 150 kilobits up. He cannot predict the winning interface, but doubts the one-way feed will also define the generative-AI era.

2. Natural language turns blunt behavior into legible intent

Patrick’s pushback: chat users arrive expecting to write, but will Spotify listeners invest similar effort instead of remaining lazy? Söderström’s answer is that even occasional high-fidelity descriptions are radically better than inference from playback behavior alone.
Playlisting already gave Spotify valuable labels because users deliberately placed songs together. Skips are plentiful but ambiguous: the listener might hate the song, love it but be tired of it, or simply find jazz wrong for the gym.
AI Playlist, available in 40 countries, accepts requests such as a running playlist with EDM, big drops, and 160 BPM. Listeners can preserve successful tracks, reject others or specific artists, and ask for more of what worked.
Spotify has always tried to “reproduce a small part of your neural cortex on our servers.” Natural language makes that approximation easier and turns personalization toward continuous qualitative research across almost 700 million users, alongside conventional interviews and A/B tests.

3. The enterprise bottleneck is legacy infrastructure, not model access

At a large company, Söderström estimates developers may code for only one of eight working hours, with net-new code occupying a small fraction of that. Models must still improve at understanding and refactoring Spotify-scale codebases and performing trusted automatic peer review.
The larger prize may be the other seven hours: communication, planning, design collaboration, meetings, and prototyping. AI’s addressable workflow therefore extends well beyond making an engineer type code faster.
Model Context Protocol, or MCP, can wrap internal services so employees “speak English to your infrastructure.” A designer or PM could screenshot Spotify, ask Cursor for a clickable HTML prototype, then connect it to a real songs feed without becoming a developer.
One nontechnical PM even wrapped Sweden’s tax authority in an MCP and did her taxes through Cursor. Spotify’s harder version is exposing 15 years of cold-stored listening history through real-time APIs: data that previously would have required an engineer to run an SQL job that might take a week. “Old school engineering” comes before AI engineering.

4. Explanations make product judgment transferable

Spotify uses Seven Powers, bundling theory, and Felix Oberholzer-Gee’s value stick to give product and technology teams a strategic vocabulary. The goal is to keep willingness to pay well above price while improving employees’ willingness to sell their services through mission and culture.
Söderström’s deliberately provocative maxim is “Talk is cheap, so we should do a lot of it.” Structured Socratic debate is inexpensive compared with execution and can expose weak reasoning before a company spends resources building it.
Drawing on David Deutsch, he wants explanations that are falsifiable, possess reach, explain why, and are “hard to vary.” A theory that can swap Thor for another angry god without losing predictive power probably explains little.
His “100% science and 0% magic” claim does not dismiss senior intuition; it treats intuition as valuable pattern recognition that remains trapped in one person. Even after a successful A/B test, he asks for a theory of why it worked so the learning can spread and predict new behavior.

5. The bets board forces real capital-allocation choices

Every six months, Spotify’s roughly 14 VPs pitch proposed bets as though presenting startups to a VC. Personal support from Daniel Ek, Söderström, or Norström is not enough; each VP must articulate why the company should fund the work.
The cited cycle contained 44 bets, within a usual range of about 30-50. Söderström and Norström create one global ranking, then teams resource from the top until capacity runs out—perhaps around bet 30—and explicitly commit to delivery.
Stack-ranking is simple in theory and painful in practice: “You have your two darlings, but if you have to kill one of them, which do you kill first?” Declaring initiatives equally important merely pushes conflict down to competing VPs.
During the preceding cycle, teams prototype the prospective future Spotify in Figma and increasingly with generative-AI tools. Misalignment and “fighting” happen before commitment, when leaders can still hold and evaluate an integrated product rather than discover collisions late in execution.

6. Tuesday’s execution meeting resolves dependencies in real time

The entire VP group forms an execution team for three hours every Tuesday. With a five-day week, a blocked dependency can therefore reach Söderström, Norström, and the relevant executive within at most two-and-a-half working days.
Participants are “not allowed to say the word offline” or defer resolution until later: the counterpart is already in the room. A dependency owner can explain the constraint, learn that it was unknown, and commit to fixing it immediately.
VPs cannot bring direct reports to explain details. That forces leaders to understand their own work and preserves a stable group whose accumulated rapport permits unusually direct discussion.
The room spans business, product, and technology. Commercial leaders learn AI and monorepos; engineers and product leaders learn the P&L, gross margin, and company goals—knowledge Söderström considers necessary for developing a CEO-level perspective.

7. Spotify’s super-app strategy dictates synchronization

Spotify divides resources among platform, consumer experience, personalization, and the music, podcast, and books verticals. Each block initially plans with predictable resources; only near the end does leadership move people globally to close critical gaps.
The organizational design follows the consumer strategy. As the average number of app-store installs trended below one, Spotify chose the Chinese super-app logic: put music, podcasts, books, and video inside one application and reuse scarce distribution.
One shipped app makes every vertical interdependent, so Spotify cannot fully divide and conquer. Six months balances reaction speed against overhead: a quarter means excessive planning, while a year is too slow. Söderström stresses this is “not the right one. It’s the right one for us.”

8. AI looks overhyped today and enormous after reconstruction

Söderström cites studies finding only about a 7% speedup across a big-company developer’s actual time. “Right now it’s a bit overhyped” relative to realized impact, even though summarization, coding assistance, and personal productivity already occur constantly.
The step-change comes when companies expose all data in real time and rebuild workflows around reasoning engines rather than tack models onto existing systems. Startups feel further ahead because they have no 15-year infrastructure estate to reconstruct.
AI also breaks software’s traditional economics: large upfront investment no longer guarantees negligible marginal cost. More inference must be funded through effective advertising or subscriptions, and Söderström expects tiers based on usage; Spotify is accustomed to balancing paid conversion against per-stream royalty costs.
Patrick argued that cheaper inference could create vast consumer surplus. Söderström agreed there is years of “product overhang” even if models froze, but rejected any practical ceiling on demand: spreadsheets made calculation cheap and produced more accounting, while abundant intelligence may be spent on whether tomorrow’s coffee should be one degree warmer.

9. The free mobile shuffle tier came from first principles

Copying YouTube’s foreground, on-demand model looked obvious during the smartphone transition, but Spotify found only about 9% of listening occurred in the foreground. That pattern would have neglected the 91% background use case.
The desired product was more ambitious: favorite songs playing forever for free with the phone in a pocket. Offering full on-demand access, however, would have replicated Premium closely enough to threaten cannibalization.
Premium data supplied the solution: subscribers voluntarily shuffled playlists about 50% of the time and used specific on-demand selection for the other 50%. Giving away shuffle preserved substantial value while leaving every Premium listener with a reason to remain paid.
Long trials did not solve the underlying need because users knew their playlist investment might disappear when the trial ended; Söderström cites Nokia Comes With Music’s year-long offer. The unintuitive permanent shuffle tier made growth “explode” and remains a differentiation against competing services.

10. Spotify positions its profit engine as music’s R&D department

Spotify emerged from piracy-era Sweden, where one UK label executive reportedly saw a Swedish company’s P&L and said, “That’s not a business. That’s a hobby.” The industry’s desperation made it willing to take risks with Spotify; Söderström says Spotify itself took enormous capital risk, including minimum guarantees.
Since roughly 2012, Söderström has called Spotify “the R&D department of the music industry.” It shares about 70% of revenue and invested the remainder plus more, remaining unprofitable for 15 years before the first profitable year and then recycling profit into people, product, and AI.
Spotify paid out more than $10 billion in the cited year, up from about $1 billion almost a decade earlier. Söderström argues music is now larger than during the CD era; more creators divide the expanded pie, but excluding newcomers is not a defensible remedy.
Royalties are paid from subscriber economics, not a fixed per-stream rate. Spotify’s roughly twice-higher engagement and half the churn of rivals mechanically depress its displayed per-stream figure; even distributing 100% would yield only about 1.5× today, so his answer is billions of payers, not rearranging the split.

11. Podcasts and books make the subscription more nutritious

Podcast demand first appeared in Spotify’s Hack Weeks, where employees repeatedly forced the format into the product. Management saw long-form, full-sentence discussion as a counterforce to shrinking attention spans and chose a small, organically growing market with lower acquisition costs than a mature category.
The philosophical test is how a user feels after losing an hour: Spotify wants to resemble nutritious food, not the “bad calories” of doomscrolling. Parents directing children away from screens and toward Spotify reinforce that positioning.
Audiobooks presented a similar market-design bet. Only roughly 10-11 million people in the US paid à la carte, where a $15 book discourages exploration just as $0.99 songs discourage soundtracking sleep; Scandinavian access models suggested bundled listening could become mainstream.
Spotify found the formats were not cannibalistic: music consumption remained, while podcasts and audiobooks added listening, retention, and willingness to pay. Repeatedly stacking product and media value created the consumer surplus that made price increases—and profitability—possible.

12. One simple interface conceals several businesses

Music uses pooled royalties; podcasts primarily use advertising, alongside a Spotify Partner Program under which Premium listeners get more uninterrupted listening without Spotify ads. Audiobooks include a listening allowance inside Premium and charge top-ups when users exceed it.
Consequently, where a user clicks changes Spotify’s costs and financial outcome. The company built what it calls the “Spotify machine” to forecast behavior across these different regimes while personalization decides among music, podcasts, books, and video.
A single experience organization protects the consumer across mobile, desktop, cars, and speakers; a separate personalization organization prevents vertical teams from programming solely for their own P&Ls. Spotify is “one thing on the front end” and many things on the back end.
Five years out, Söderström hopes Spotify has crossed one billion users, become one of the largest media subscriptions, made audiobooks mainstream, and added undisclosed verticals. With largely nonexclusive commodity content, the distinct combination of media and product becomes the differentiation.

13. Podcast exclusivity was a costly category error

Netflix made exclusive podcasting look attractive, but Spotify misread the format’s economics. Podcasts thrived because production could be as cheap as Joe Rogan recording in a trailer; exclusivity countered that low-cost model and required Spotify to become an unusually accurate content picker.
Celebrity status also failed to guarantee hosting ability, while strong podcasters emerged through an organic system. Spotify already had the alternative advantage: acquire the broad catalog, then use machine learning to match different shows to different listeners.
Management pivoted to “the age of syndication,” accepting that creators want to be everywhere. Abandoning exclusivity saved money, expanded the catalog, and improved podcast viewing: “The real cost is when you try to defend your past decisions.”

14. Input-based accountability preserves ambition after failure

Söderström stays energized by periodically returning to code and new tools, then ranging through physics, mathematics, and philosophy. After years on consciousness, his honest conclusion is that he “didn’t crack it”; the pursuit still enriched his product conversations.
Brazilian jujitsu supplies a physical version of the same lesson: someone half his size can dominate through leverage without sweating. Its open competition tests technique, while control can be calibrated without the harms of punching—humility joined to practical usefulness.
The professional kindness that mattered most was Daniel Ek allowing him to “screw up a bunch of things” without expecting dismissal. Repeated second chances raised Söderström’s ambition rather than teaching him to minimize visible risk.
His Moments interface, conceived before Musical.ly, autoplayed content and used swipes across genres. Editors could not supply the personalization that later machine learning might have enabled; after launch, Spotify discovered a bug in the A/B test that had looked positive, rolled back a drastically underperforming product, and lost roughly a year. Söderström says Ek responded in the spirit of a Jeff Bezos principle: judge “the inputs you had, not the outputs.”

From pharma to AGI hype, and developing AI in finance: Martin Shkreli’s journey The RAISE Act: Minimum Standards for Frontier AI Development, with NY Assembly Member Alex Bores