AI Music Generation Revenue: Where the "It's a Rounding Error" Advice Breaks Down

The brief landed on a Tuesday: nine cues, 118–128 BPM, "driving but not aggressive," no vocals, delivered 48kHz WAV plus stems, full buyout, ten days. I have written that cue a hundred times — a filtered saw pad, a sidechained sub, a clap that lands slightly behind the grid so the whole thing breathes. What had changed was the rate, which had drifted from a number I would have taken without thinking to a number I had to think about, and a line near the bottom noting how many composers had received the same brief.

One brief, one composer, one inbox. That is an anecdote, not a market. But it is the anecdote that sent me looking hard at what AI music generation revenue actually adds up to, because the advice circulating among people whose job is to have opinions about this did not match what my inbox was doing.

Here is the advice, stated as plainly as anyone states it:

AI music barely earns anything. The listening share is tiny, the per-track payouts are noise, the handful of viral AI "acts" are freak outcomes. Make better music and stop reading think-pieces.

I want to be fair to that advice, because it is the most defensible position anyone has offered, and because the alternative narrative — that a wave of synthetic tracks is about to drain the royalty pool by Christmas — has been wrong about its timeline every single year it has been asserted. So this piece does the boring thing. It works out where the advice is right, where it stops being right, and what a more honest version of the rule sounds like.

Where the advice is roughly right

Start with the mechanism, because most arguments about this skip it.

The major subscription services pay out of a pot, not per listener. Subscription and ad money for a given market and month goes into a pool, and rights holders draw from it in proportion to their share of qualifying streams. There is no per-stream rate written into a contract anywhere. The figures you see quoted — a third of a cent, four tenths of a cent, the range depends entirely on who is doing the arithmetic — are outputs, reverse-engineered after the fact by dividing payouts by streams. They move with subscription price, market mix, ad rates, and the size of the denominator. Anyone quoting you a fixed per-stream rate as if it were a posted price is telling you they got the number from someone else.

On top of that, the floor is real. As of writing, Spotify does not pay anything at all on a track until it clears a minimum annual play count — a thousand streams in the prior twelve months. That threshold was controversial when it landed and it is worth understanding on its own terms: it exists specifically to stop the long tail of near-zero-play uploads from converting into millions of sub-penny payments that cost more to process than they are worth. A generated track that nobody finds does not earn a fraction of a cent. It earns zero, and it will keep earning zero.

Which brings up the part the panic narrative consistently underweights: generation got cheap, discovery did not. Making a competent-sounding four-minute track is now a commodity operation. Getting anyone to hear it is the same brutal problem it has always been, and arguably a worse one, because every uploader now faces a bigger pile. The AI acts you have read about did not succeed because the audio was extraordinary. They succeeded because someone ran a distribution and press play — a plausible band biography, a shareable origin story, playlist placements that were pitched like anyone else pitches them.

The outliers deserve their scare quotes and their asterisks. The Velvet Sundown, the synthetic "band" that accumulated a serious monthly listener count in 2025 before the project acknowledged what it was, is the reference case everyone cites. It got there on novelty and on the news cycle about itself — a self-consuming loop that is not repeatable at scale, because the second and fiftieth instances are not news. Xania Monet, the AI-voiced R&B project reported to have signed a multi-million-dollar deal, is a genuine data point about where label capital is willing to go, and a terrible data point about typical earnings, in the same way that a lottery winner is a terrible data point about the returns of buying tickets.

So yes. If your question is "will synthetic tracks meaningfully reduce my streaming royalty statement next quarter," the honest answer for almost every working artist is no, and the people telling you otherwise are usually selling a newsletter subscription or a policy position.

How to read any AI-revenue number you get handed

Before the part where the advice breaks, a filter. Every figure in this debate arrives pre-spun, mine included. Six criteria I run on anything before repeating it:

Who commissioned it. The most-cited forecast in this space — the CISAC-commissioned study projecting a large share of music creator revenue at risk within a few years — was produced for a global body of authors' societies whose institutional purpose is to argue that authors' revenue needs protecting. That does not make it wrong. Forecasts commissioned by interested parties tend to be directionally useful and numerically soft, and this one is a model of substitution effects out several years, not a measurement of anything that has happened. Read it as a well-argued hypothesis with a number attached.

Streams versus payouts. Third-party analytics services estimate earnings by multiplying observed stream counts by an assumed rate. That assumption is doing all the work. A piece that says an act "earned" a specific six-figure sum is usually reporting streams × a guessed constant, which is a defensible method and not a bank statement.

Gross versus landed. Even a correct gross figure is not what anyone receives. Distributor cut, label share, publishing splits, and the writer/performer division all happen downstream. A headline payout number and a person's actual income differ by a factor that varies by deal.

Forecast versus measurement. "Fifty percent of uploads" is a measurement. "Billions at risk by 2028" is a model. These get quoted in the same breath constantly and they are not the same species of claim.

What counts as "AI." This is the messiest one. Fully generated instrumental, generated with a human topline, human performance with generated backing, human song with a synthetic voice model, human everything with AI mastering — these all end up inside the same percentage in most reporting. The DDEX disclosure fields that would let anyone separate them exist, and adoption is early and voluntary, which means the underlying data is thin on the exact distinction the argument depends on.

Uploads versus listening versus payouts. Three different numbers, routinely conflated:

Deezer, the one major service publishing an upload figure, has revised it upward repeatedly — from a low-double-digit percentage of daily deliveries to a share it has described as approaching and then exceeding a third within roughly a year. Every revision has moved the same direction. That is a genuinely striking number about supply. It is not a claim about listening, and Deezer has been careful to say so, alongside a stated policy of excluding fully generated tracks from recommendation and from royalty payouts. Spotify, separately, said it removed on the order of 75 million spam tracks over a twelve-month stretch — the company's own figure, unaudited by anyone outside it, and one that describes a cleanup operation rather than a market share.

Where the advice breaks down

Five places. Some are more urgent than others.

1. The pool is fixed, so the mechanism is dilution, not competition

The framing "can an AI track beat my track" is the wrong shape. In a pro-rata pool, nothing has to beat you. Every qualifying stream anywhere in the system shrinks the value of every other qualifying stream, because the denominator grows and the numerator does not. A synthetic track that captures fifteen minutes of somebody's sleep playlist has taken money from the pot without competing with you for a listener in any sense you would recognise as competition.

That effect is currently small — single-digit-percent small, on the listening-share numbers anyone will stand behind. But it compounds silently and it does not require a hit. It requires volume, and volume is precisely the thing that got cheap.

2. The floor of the paid market moves first, and streaming data cannot see it

This is the part the streaming-share argument structurally cannot address, and it is the part I feel in my own invoicing.

The money most working music people actually live on is not streaming royalties. It is the corporate video cue, the podcast bed, the game menu loop, the retail in-store bed, the library track licensed at a rate that would embarrass a session player. That market has three properties that make it exposed in a way a catalogue of released songs is not: the brief is functional, the buyer is not a fan, and the license is often a flat buyout with no downstream reporting to anyone.

When an ad agency's junior producer generates a 30-second bed instead of licensing one, no stream is registered, no royalty is not-paid, and no study counts it. The transaction simply does not occur. AI music generation revenue, measured as money flowing to AI outputs, undercounts this entirely — because the relevant number is not what the generated track earned, it is what the licensed track stopped earning.

I do not have a clean figure for that, and I am suspicious of anyone who says they do. What I have is a rate card that has moved, a brief count that has gone up, and conversations with other library composers who describe the same drift. Anecdote, plural. Treat it accordingly.

3. Functional listening is the soft tissue

Some listening is about who made it. A lot is not. Sleep, focus, study, ambient, background lo-fi, workout beds — categories where the listener's requirement is "appropriate and unobtrusive for forty minutes" and where playlist real estate is the entire distribution mechanism. In that space, provenance carries close to zero weight with the end listener, and the incumbent human-made catalogue was already competing against library-grade production commissioned specifically to fill those slots.

If your income leans on that kind of placement, the tiny aggregate listening share is misleading, because the exposure is not distributed evenly across genres. It is concentrated exactly where you are.

4. The capital is not chasing per-track royalties

Here is the investment answer buried under the royalty argument. Serious money entering this space is not buying streams of generated tracks. It is buying the generation tools, the rights positions that make training defensible, the catalogues that get licensed into models, and the distribution layer that sits between the two. Label deals with model companies, catalogue acquisitions, litigation settling into licensing — that is where the value is being assigned.

Which means the return on a per-track basis being lousy is completely consistent with a great deal of capital moving. Those are separate markets. If you are a stakeholder trying to work out whether this is real, per-stream payout is close to the least informative number you could pick.

5. Disclosure is where value gets assigned, and it is barely built

Whoever ends up with a working, enforced disclosure standard gets to decide what "AI-generated" means for payout purposes, and therefore who gets paid. The DDEX credits exist. Spotify has moved to support them; Deezer tags albums it detects. Adoption is voluntary, detection is imperfect, and the incentive to under-disclose is obvious — the Madonna-cover episode, where a generative credit appeared only after the track had done its numbers, is the small ugly version of a problem that gets bigger with the stakes.

Until that plumbing works, every revenue figure in this argument — including the ones I have used above — has an error bar wider than most people quoting it will admit.

If you make electronic music, you are standing closer

I will be direct with the EDM and electronic producers, because the general advice is worse for you than for a singer-songwriter.

Your genre's briefs are the most describable in music. "124 BPM, minor key, rolling sub bass, filtered build, no vocal, sixteen-bar loop, clean at the top for VO" is a complete specification. It is also, word for word, a prompt. Genres whose value is legible as parameters are the ones a parameter-driven system serves best. That has nothing to do with whether the output is good — plenty of it is mushy, the transients smear, the low end is often a mono blur that falls apart on a real system, and anything asking for genuine arrangement tension across six minutes still tends to come back as a loop wearing a costume.

But the filler tier of your market was never asking for arrangement tension. It was asking for a competent 124 BPM bed by Thursday.

What does not compress into a prompt: a live set that reads a room, sound design with a signature somebody can identify in two bars, an artist relationship a supervisor calls directly, adaptive audio that has to respond to game state, and mixes that hold up on a large system. That is not a consolation list. It is a description of where the defensible margin sits.

Who should act on this, and who should stop reading about it

Act now: library and production-music composers, anyone whose income skews to buyouts and functional briefs, catalogue owners with heavy exposure to mood and background playlists, and distributors — the fraud and spam-filtering burden lands on you first.

Watch closely, act slowly: publishers and rights administrators. The disclosure-metadata question will determine your economics more than the listening share will, and it is being decided now in standards bodies rather than in the press.

Stop reading about it: artists with a real audience relationship, live acts, anyone whose listeners can name them. The pool dilution is genuine and it is currently a small tax, not a threat, and the hours you spend tracking this are hours not spent on the thing that actually protects you. That is the part of the standard advice that survives intact.

The honest version of the rule

The original rule — AI music earns almost nothing, ignore it — is right about the wrong quantity. It measures money arriving at generated tracks, when the number that determines your year is money that stopped arriving at commissioned ones. It is right about outliers, right about discovery, right that generation being free does not make attention free, and right that most synthetic uploads earn precisely zero.

It is wrong that the aggregate being small means your exposure is small, because exposure is concentrated by genre and by contract type, and it is wrong that streaming statistics can see the market where this actually bites. The rewritten version: AI music generation revenue is currently a rounding error, and the substitution it causes in the buyout and library market is not, and no published number tracks the second one yet.

So here is the rule for tonight: if a client could describe what they want in one sentence you could paste into a prompt box, that line of your income is already priced against a machine — spend your next hour on the work that only exists because you were in the room.