The Invisible Architecture of Music Streaming

How business and tech have perhaps killed songwriting, originality, and art

There's a moment that often comes back to me. A few years ago, I was at a concert in a small club in Pescara, one of those places where the stage is as high as a single step and you can smell the bass player's sweat. The band was unknown (3 completely crazy kids with an unmistakable smell), the audience maybe thirty people, the sound raw and imperfect. But there was something alive in that room, a direct connection between those playing and those listening that didn't need algorithms to exist. Going home, I opened Spotify to search for that band. They weren't there. Or rather, they were, but buried under thousands of other results, invisible unless you knew exactly what to look for. And I wondered: how many bands like that exist in the digital shadow, technically present but practically nonexistent?

I spent some time studying how music streaming platforms really work. Not the surface—we all know that. I was interested in understanding what happens underneath, in the technical guts of these systems that mediate our relationship with music. What I found made me rethink everything I thought I knew about how we discover music today.

Let's start with a truth that's rarely told: when you click play on Spotify, you're activating one of the most sophisticated logistical machines ever built. That's not an exaggeration. Spotify manages over one hundred petabytes of data and more than two thousand distributed microservices. Every song you listen to doesn't come from nowhere—it comes from a chain of servers, caches, geographically distributed nodes working together to make you believe the music is simply there, ready for you. It's an extraordinary illusion, and like all successful illusions, it hides something important.

The first thing to understand is that not all songs are equal to the system. I'm not talking about artistic quality—I'm talking about something much more prosaic: distribution cost. When you listen to a song, that song has to travel from somewhere to your device. If it's already stored on a server near you, the cost is minimal. If it has to be retrieved from a datacenter on the other side of the world, the cost rises. Multiply this by billions of plays per day and you understand why optimization becomes an obsession.

This is where the concept of predictability comes in, and this is where things get interesting. The system works better when it knows in advance what you'll listen to. If it can predict that tomorrow millions of people will listen to a certain song, it can distribute that song in advance to hundreds of servers around the world. When the request comes, the song is already there, ready. No delay, no extra transfer cost. It's elegant, efficient, economical.

The problem is this: how do you predict what millions of people will listen to? The answer is as simple as it is unsettling. You don't predict it. You decide it yourself.

Editorial playlists are not a service offered to users. They are a consumption control tool. When Spotify creates a playlist like "Today's Top Hits" and presents it to fifty million followers, it's not simply suggesting music. It's creating a predictable flow of plays that can be optimized at the infrastructure level. The system knows that playlist will be listened to by hundreds of thousands of people, knows they'll follow the proposed order, knows the first fifteen songs will receive most of the plays. With this certainty, it can pre-distribute those songs to all necessary servers, ensuring that every play is served from local cache instead of the central datacenter.

I made a rough calculation.

Let's assume one and a half billion streams per day, an average bitrate of 128 kbps, an average song of three minutes. Without optimization, the daily bandwidth cost would be in the order of tens of millions of dollars. With the predictive prefetching made possible by playlists, that cost can be reduced by sixty, seventy percent. On an annual scale, we're talking billions of dollars in savings. This isn't speculation: Spotify has invested hundreds of millions with Google Cloud precisely to optimize this type of flow.

Albums, on the other hand, remain a problem. Not because they're technically different, but because they're unpredictable. When a user saves a twelve-track album, the system doesn't know which tracks they'll listen to, in what order, when. They might only listen to the singles. They might start from track seven. They might never listen to it. This uncertainty means the system must allocate resources conservatively, preparing for every eventuality without knowing which will occur. It's inefficient, expensive, and creates what technicians call cache waste—pre-stored songs that nobody touches.

The consequence is subtle but profound: the system has a structural incentive to promote playlists over albums. Not because someone decided albums are bad, but because the very architecture of the system rewards predictability. And playlists, by definition, are predictable. Albums are not.

I've often wondered if those who designed these systems were aware of the cultural implications. Probably not, at least initially. The engineers were solving a distribution problem, not thinking about redefining how music is discovered and consumed. But the effect is the same, intentional or not.

There's a study from the National Bureau of Economic Research that particularly struck me. It measured the impact of Spotify's editorial playlists on streams. The results are staggering. Appearing on "Today's Top Hits" increases streams by twenty to forty times in the following week. Appearing on "New Music Friday" by five to fifteen times. This isn't simply promotion. It's success creation. By deciding which songs to put in those playlists, Spotify literally decides who will succeed and who won't.

The feedback loop that results is even more insidious. A song enters an editorial playlist. It receives millions of plays. The algorithm perceives it as popular and starts suggesting it organically to other users. The song receives even more plays. The infrastructure pre-caches it even more aggressively, further reducing distribution costs. And meanwhile, somewhere, an unknown band plays in a small club in Pescara, technically present on the platform but practically invisible.

I don't want to paint all this as an evil conspiracy. It's something more complex and, in some ways, more disturbing. It's the natural emergence of a network of incentives: technical, economic, market-driven. No one decided that experimental music should be penalized. It's simply that experimental music is unpredictable, and unpredictability costs. No one decided that organic discovery should be stifled. It's simply that controlled discovery is more optimizable. The result is the same, but the cause isn't a villain in a room pulling strings. It's a system doing what it was designed to do.

Discovery Mode is perhaps the clearest example of this dynamic. Introduced in 2023, it allows artists to accept a reduction in royalties—up to half—in exchange for priority in recommendation algorithms. It's brilliant, in a way that makes me uncomfortable. Spotify isn't directly manipulating the algorithm—it's incentivizing artists to voluntarily give up part of their earnings to gain visibility. The result is that those who can afford to earn less get more exposure, while those who need every cent remain in the shadows. And from an infrastructure perspective, the system wins twice: it pays fewer royalties and gets predictable traffic, perfectly optimizable.

The underlying economic model amplifies all of this. Spotify doesn't pay artists per stream linearly. It uses a pool system: all subscriptions go into a pot, and that pot is divided proportionally among artists based on their streams. This means that when an artist gains streams, they're not adding to the pie—they're taking a bigger slice of a pie that stays the same size. And who are the artists gaining more streams? Those in editorial playlists. The system rewards those already rewarded, penalizes those already penalized.

An independent artist would need two hundred fifty thousand streams to earn one thousand dollars gross. If half goes to distributors and rights, that leaves five hundred dollars net for the work of composing, recording, promoting. For an entire album. Meanwhile, an artist who gets an editorial playlist can receive millions of streams in a week, generating tens of thousands of dollars. It's a lottery, and the house controls who wins.

What strikes me is how all this is invisible to the average user. You open Spotify, you see an apparently infinite selection of music, you feel like you have the world at your fingertips. But that selection is filtered, sorted, prioritized by systems that have their own incentives, their own technical necessities, their own business models. You're not exploring a neutral archive. You're navigating a territory shaped by forces you don't see.

Researchers talk about filter bubbles and echo chambers. Recommendation algorithms tend to suggest music similar to what you've already listened to, creating self-reinforcing cycles. You listen to indie pop, the algorithm suggests more indie pop, you listen to even more indie pop, and before you realize it you're trapped in a genre you might never have consciously chosen as your musical identity. The diversity of the offering doesn't translate to diversity of experience.

There's something deeply ironic in all this. Music streaming was born with the promise of democratizing access to music. Any artist can upload their music, any listener can access millions of songs. In theory, the barriers to entry have disappeared. In practice, they've been replaced by different barriers—more subtle, harder to see and therefore harder to contest. It's not that you can't get in. It's that once inside, you're invisible unless the system decides to show you.

I often think about what all this means for music as an art form. Experimental, niche, innovative music—the kind that by definition doesn't fit predictable patterns, the kind you can't pre-cache because you don't know who'll listen to it—is structurally penalized. Not because someone hates it, but because the system's architecture doesn't favor it. The economic incentive goes toward homogenization, toward repeating formulas that work, toward predictability that reduces costs.

I obviously don't have solutions. I don't believe Spotify is evil, nor that we should go back to vinyl or CDs. Technological progress has brought real benefits: universal access, reduced costs for listeners, the possibility for independent artists to reach a global audience without going through major labels. But it seems important to understand what we're losing while gaining all this. To understand that when we open an editorial playlist we're not simply listening to curated music—we're participating in an economic and technological system that has its interests, its logics, its consequences.

Perhaps the most important thing is to keep alive a form of conscious resistance. Actively seek music outside algorithmic recommendations. Go to concerts, discover local bands, follow independent blogs and magazines, talk to friends who have different tastes than ours. Use streaming for what it is—an extraordinarily convenient tool—without forgetting that every tool shapes the use we make of it.

That evening in the club in Pescara, going home after the concert, I realized the experience I had lived wasn't replicable on Spotify. Not because the music was different, but because the context was different. The discovery had happened through chance, physical presence, word of mouth from a friend (thanks Kris!) who had dragged me there. No algorithm had mediated it, no system had optimized it. It was inefficient, unpredictable, impossible to scale. And perhaps precisely because of that, it was alive in a way no editorial playlist can ever be.

What leaves the most bitter taste in my mouth is the awareness that all this isn't an accident. It's the predictable result of a system that put business in command and used technology as a control tool. Songwriting, art, musical exploration—everything that by its nature is unpredictable, personal, irreducible to formula—has been systematically marginalized. No conspiracy was needed. It was enough to build an infrastructure that rewards the homogeneous and penalizes the different, and then let the market do its work.

We hear the result every day. Songs that seem to come from the same factory, identical structures, predictable drops, interchangeable lyrics. Not because musicians are less talented than before, but because talent that doesn't fit the format gets filtered out before it even reaches our ears. The algorithm selects, the playlist amplifies, the loop closes. And whoever doesn't enter the loop simply doesn't exist.

What I find almost comical, in a tragic way, is that the same music industry is now trembling before Suno and other AI-based music generators. Suddenly everyone realizes that maybe, just maybe, they should have valued what makes an artist irreplaceable instead of treating them as a provider of fungible content. They spent twenty years compressing royalties, pushing toward homogenization, building systems where music is a commodity to be optimized. And now they're surprised that someone built a machine capable of producing musical commodities at zero cost?

The irony is perfect. They trained the public to accept generic, predictable, interchangeable music. They built playlists where one song is as good as another, where the artist's identity is irrelevant, where all that matters is that the sound fills the background without disturbing. And now that same music can be generated by an AI in thirty seconds, without paying anyone. What did they think would happen?

I have no compassion for an industry that chose to kill what made it necessary. If you spent decades convincing the world that music is just background entertainment, you can't complain when someone finds a cheaper way to produce background entertainment. The value of art lies in its uniqueness, in personal vision, in human imperfection. Everything you systematically eliminated in the name of optimization.

That evening in the club in Pescara, and then going home after the concert, I understood something I couldn't articulate at the time. The music I had heard wasn't optimizable. It didn't fit any playlist, didn't adapt to any algorithm, didn't generate predictable patterns. It was inefficient, uncomfortable, impossible to scale. And it was alive in a way that no system based on predictability will ever be able to replicate, neither human nor artificial.

Songwriting might survive. Art might survive. But not thanks to the streaming industry or the music industry at large. It will survive despite it, in small clubs, in independent productions, in niches that algorithms can't see. And when the current business model collapses under the weight of its own contradictions, crushed by music generators that do exactly what the system has always asked for, perhaps someone will remember there was an alternative. That it was possible to choose to value art instead of optimizing it.

And perhaps it will be too late.