Think about the last time a song genuinely changed how you listen to music. Not a track which the algorithm surfaced between two songs you already knew. Something a person put in front of you. A friend who wouldn't shut up about a band. The owner of a record store that said, "Try this." A DJ on late-night radio playing something that had no business being on the playlist. That moment where a human being with taste and conviction put something unfamiliar in your path, and it rewired you. When did that last happen?
That moment has a business model underneath it. And we broke it.
The CD era is often described as a golden age of music, which is true, from a business standpoint, but the important thing about CDs wasn't the sound quality or the format. It was the margin. A CD costs as little as 75¢ to manufacture and sold for $12 to $18. During the boom years of 1984 to 1995, label profit margins averaged above 30%. In 1999, the year CD sales peaked, the US recorded music industry hit $23.7 billion in revenue. Adjusted for inflation, that number is higher than anything before or since.
80% of albums released by major labels during this period lost money. 8 out of 10 records failed commercially. That sounds catastrophic. It was actually an industry with enough breathing room to invest in the long term. A label could sign a band that wasn't going to sell a million copies in year one, fund their development, give them time to find their sound, and absorb the loss if it didn't work out, knowing that some of those long bets would eventually become the most valuable artists on their roster. That's how Frank Zappa ended up on a major label. That's how Radiohead were given the space to pivot from The Bends to OK Computer. An A&R executive could sit in a room with a musician who had no numbers, no following, no data trail, and feel something. The way they held a room. The conviction in their voice. The sense that this person had spent years developing something that hadn't surfaced yet but was about to. Thats judgment: the ability to recognise something you've never encountered before.
A recognisable sight for any millenial
Then Napster proved that people would always take free music over paid music if the option existed. Apple responded with iTunes, selling individual tracks for 99¢, which killed the album as an economic unit overnight. You no longer had to buy twelve songs to get the two you wanted. Revenue collapsed. By 2008, US recorded music income had dropped from $14.6 billion to $9 billion. Then Spotify arrived with a pitch that, from the consumer's perspective, was irresistible: every song ever recorded, available instantly, for the price of a couple of coffees a month. In the name of democratising music, the entire economic relationship between artist and listener was restructured. We stopped buying. We started renting. Streaming widened access and lowered barriers to entry in ways that were genuinely transformative. It also replaced human curation with a system that rewards familiarity at scale. Spotify pays artists roughly $0.003 to $0.005 per stream. A million streams earns between $3,000 and $5,000, most of which goes to the label. The margin that sustained creative risk-taking has evaporated.
Thus, the money moved. For fifty years, touring promoted records. The live show existed to sell the album. Streaming inverted that completely. The recording became a loss-leader for the live business. Artists now make records to build enough algorithmic visibility to sell tickets. The gig is the product. The song is the marketing material for the gig. You can press unlimited CDs, but there are a finite number of stages, a finite number of nights, and you have to physically turn up. The abundance that streaming promised collided with the scarcity of the only business model left standing.
There now exists a creativity incentive paradox. You cannot tour without an audience. You cannot build an audience unless the algorithm recommends you. The algorithm only recommends what has already worked. Spotify's own research illustrates the problem: active listeners who choose their own music listen more widely; passive listeners who let the algorithm choose go narrower. Every recommendation pulls toward the centre. So artists respond rationally, treating songs as products to be optimised for the platform's specifications. Intros are engineered down from twenty seconds to five. Hooks are front-loaded as the primary deliverable. Songs open with the chorus because the chorus is the only component the system needs to register engagement. A Samsung report projected that by 2030, the average hit will be roughly two minutes long, half the length of a 1990’s track. The song is no longer the product. The hook is the product. The rest is packaging.
Notice the language. Songs as products. Hooks as deliverables. Listeners as engagement data. We have adopted the vocabulary of product development to describe an art form, which tells us how far the pendulum has swung. In the 1970s, an artist might spend years formulating, experimenting, and growing their own voice before you ever heard a note. All this internal creative development is invisible to an algorithm. An algorithm can only see output: streams, skips, saves, and playlist adds. It cannot sense potential. It cannot recognise that someone is developing something unprecedented. The judgment that defined A&R in the CD era has been replaced by pattern recognition, and those are fundamentally different capabilities. Judgment can identify something it has never encountered before. Pattern recognition, by definition, cannot. Each cycle of algorithmic recommendation flattens the landscape further, ironing the cultural tapestry into something which a system that cannot experience joy nor develop taste has been programmed to surface. We are analogue. We discover what we love by being surprised. The algorithm can only predict our past and call it our future.
Which brings us to the business consequence of all of this.
Spotify paid $11 billion to rights holders in 2025, roughly 70% of its total streaming revenue. This was their single biggest cost. When your largest expense is paying other people for the thing you distribute, and you're simultaneously building AI tools that can generate that thing at zero marginal cost, the direction of travel is obvious. On Deezer, 28% of all tracks uploaded daily are now fully AI-generated. Music Business Worldwide found at least 13 AI-generated artists on Spotify with 4.1 million cumulative monthly listeners, from just a few days of research.
In June 2025, an AI-generated band called The Velvet Sundown released two albums and, within weeks, had 1.4 million monthly listeners, charting number one on Spotify's Viral 50 in multiple countries. No member has ever existed. The music, the band, the backstory, the album art: all generated. The algorithm recommended them exactly like it would a real band, because the music was indistinguishable from the algorithm-optimised human music surrounding it. That is the threshold. When a creative output has been optimised to the point where it can be fully specified by data, it becomes automatable. The Velvet Sundown is proof that recorded music has crossed that line. Spotify's 70% cost, their single biggest expense, has just become eliminable.
Anyone familiar with the hype cycle will recognise where streaming sits. The excitement of the early 2010s has curdled into disillusionment. Artists cannot make a living. The algorithm narrows taste. No genuinely new genre has emerged in over a decade. The biggest-selling artist of 2025, Taylor Swift, released her first album twenty years ago. AI music sits at the hinge between disillusionment and whatever comes next: the moment the system reveals what it was always heading towards. The convenience you demanded is now mandatory. Culture is in its own hall of mirrors.
However, business models emerging outside the algorithmic view suggest the correction is already underway. Vinyl sales have grown every year for nearly two decades, driven by people actively choosing to discover and own music in a way the algorithm can't mediate. Bandcamp built a profitable platform around direct artist-to-listener sales with zero algorithmic curation. Independent labels are growing by curating rosters based on taste, as A&R departments used to. The interesting question is whether this infrastructure can scale quickly enough to offer musicians a genuine alternative, before the algorithmic model automates them out entirely.
The Velvet Sundown accrued 1.4 million listeners by doing exactly what the system asked for. The question for every industry watching is: how much of what you make could a machine already replicate? And if the answer makes you uncomfortable, what are you going to do differently?
Free Bird by Lynyrd Skynyrd is nine minutes long with a four-minute guitar solo. No hook in the first thirty seconds. It peaked at number 19 on its first release. By today's algorithmic standards, that's a song that gets buried, a moderate performer with no viral moment, too long, too slow, too risky. The label kept it on the roster. Audiences found it over the years. It entered the Grammy Hall of Fame, sold millions, and became one of the most iconic recordings in the history of popular music.
Fifty years later, that four-minute guitar solo is one of the most shared pieces of audio on social media. Fifteen seconds of it, clipped and looped over chaos. The algorithm is very good at distributing fragments of culture that it could never have produced. The system that would bury Free Bird at the thirty-second mark is the same system that feeds off its legacy every day. It can circulate the cultural capital that human judgment and patience built. It is simply unable to create any more of it.
Something to think about next time Discover Weekly serves you another two-minute track that sounds exactly like the last one.
See you on the next one.
