Ninety Thousand Tracks a Day: Living With AI Music Oversupply

Somewhere in the last two years, "I generated a track for this" stopped being a sentence worth saying out loud, and the reason isn't that the tracks got worse. It's that the number of people who could say the same sentence, about a track that took them thirty seconds and no musical training, grew faster than anyone's ability to care. The data on exactly how fast is now public, and it's a cleaner argument for rethinking what generated music is actually good for than any opinion piece could make on its own.

The number that matters more than the headline

Deezer reported that AI-generated tracks made up over 50% of daily uploads to the platform at peak in June 2026, at a volume of roughly 90,000 fully AI-generated tracks landing every single day. That's the headline number, and on its own it reads like a supply story — a flood of new music.

The number that actually reframes it is sitting right next to it in the same report: despite that volume, AI-generated music accounts for only 1 to 3% of total listening on the platform. Ninety thousand tracks a day, half of everything uploaded, and it barely registers as a fraction of what people actually play. That gap — enormous supply, close to flat demand — is the whole story. It's not that AI music is being rejected; it's that volume at this scale stopped functioning as a signal of anything a listener would seek out. A flood doesn't get more interesting per gallon.

What the growth curve says

This isn't a one-time spike, either. Two months earlier, Deezer put the same figures at 44% of daily uploads and roughly 75,000 tracks per day. Do the arithmetic across that window and upload volume grew about 20% while the majority-share crossed the 50% line for the first time — a curve that's still climbing, not one that found a ceiling and stopped. Whatever "abundant" means today, it's the floor for what's coming, not the peak.

What doesn't move with it is the listening share. Both reports land in the same 1-3% band for actual consumption, months apart and tens of thousands of tracks-per-day apart. Supply roughly doubled in the time listening share stayed flat. That's not a market finding its equilibrium — it's a market where generation cost dropped so far that production volume decoupled entirely from demand, and nothing in the trend line suggests that reconnects on its own.

Why "I made a song" stopped being information

Here's the mechanism underneath the numbers. A credential is only worth something if it's scarce enough to signal effort or skill. "I recorded a song" used to imply access to instruments, studio time, or at minimum enough musical training to produce something listenable. None of that is true anymore. A usable, competently arranged, genre-appropriate track is now a text prompt and thirty seconds away for anyone, which means the fact of having generated one carries close to zero information about whether it's any good, whether it fits your project, or whether it was made with any more care than the 89,999 other tracks uploaded the same day.

This is the same pattern that played out with stock photography and, before that, with desktop publishing: the skill floor for producing a version of the thing collapsed, and the market response wasn't "everyone's output got equally valuable" — it was "the baseline stopped being worth anything on its own, and value moved to whatever the baseline still can't do for you automatically." Music generation just crossed that line faster and more visibly than most categories, because the volume numbers are public and the gap to actual listening is this stark.

What the platforms already did about it

Deezer's own response to this volume is the clearest evidence of where value went: the platform automatically excludes AI-detected tracks from its algorithmic recommendations and editorial playlists. That's a major streaming service treating raw AI-generated volume not as content to promote, but as noise to filter out of the discovery layer by default. The signal there isn't "platforms hate AI music" — Deezer's own reporting frames it plainly as a normal, growing part of the catalog. The signal is narrower and more useful: being AI-generated is not, on its own, a reason for a platform (or a listener) to pay attention. It's a production method, not a merit.

The two things that didn't get commoditized

None of this means generated music lost its usefulness — it means the usefulness moved to two specific properties that volume doesn't erode, because they're about fit to a specific use, not about the track being impressive in isolation.

Fit. A track generated against your actual brief — this footage's length, this scene's mood, this edit's tempo — solves a real production problem a library search doesn't: matching duration and feel without a compromise. A stock track that's "close enough" in mood but eleven seconds too long, or perfect in length but wrong in energy, is the everyday failure mode of library music. A track built to the brief sidesteps both failures at once, and that's a production win regardless of how many other tracks got generated that same hour by someone else, for someone else's brief.

Clearance. The other property that doesn't commoditize is licensing certainty. A track generated specifically for your content doesn't carry a rights holder's claim the way pulling a real artist's track does — a genuinely durable advantage over needle-drop licensing risk, independent of how "impressive" the composition is, and independent of oversupply entirely. Ninety thousand other AI tracks getting uploaded the same day changes nothing about whether the one you generated is safe to put behind your own video.

Both of these are about a track's relationship to a specific piece of content, not about the track as a standalone artifact competing for attention in a feed. That's the reframe: stop treating a generated track as a thing that has to be good enough to notice, and start treating it as infrastructure that has to be the right length, the right mood, and clear to use — the same standard you'd apply to a stock photo license, not the standard you'd apply to someone's demo reel.

A Versely walkthrough: generating for fit, not for attention

The practical version of "value is fit, not novelty" is a two-step generation, not a one-shot prompt:

1. Generate longer than the final cut needs, briefing genre, instrumentation, tempo, and mood rather than trying to prompt an exact arrangement — text-to-music models are far more reliable at texture and groove than at landing a structural beat (a drop, a build) at a specific timestamp, so give yourself material to cut into rather than betting on one generation landing the structure perfectly.
2. If the first pass fades out too early for your runtime, extend it from where it left off rather than starting over — this continues a previously generated track from a chosen point, which is the fix for a track that's the right mood but the wrong length, without losing the take you already liked.
3. Mix it under your existing audio, not over it — Versely's add-music workflow attaches the finished track to your video in mix mode so it sits under any dialogue or voiceover at a lower volume rather than replacing it, which is how a soundtrack should actually sit in a real edit.

That's the whole workflow, and none of it depends on the track being novel. It depends on the track being the right length, the right mood, and cleared for the use — which is exactly what ninety thousand tracks a day makes more valuable, not less, because everyone can generate a track now, and almost nobody bothers to generate the right one. For the fuller picture of where AI-generated audio, video and imagery fit into a real content operation rather than a novelty demo, Versely's audience hub breaks the workflow down by who's actually building it.