Does Music Need a New Royalty for the AI Era?
A concrete (and imperfect) proposal for a “Generative Usage Royalty”
If you’re not already a subscriber to Leveling Up and want to join other curious music industry professionals, creators, investors, and entrepreneurs, enter your email below and you won’t miss out on future newsletters:
GM readers 👋,
Happy May!
It was great seeing several of you at Amplify 2026. I’ll be heading back to NYC for this year’s Music Investor Conference Friday, June 5th through Wednesday, June 10th. If you would like to discuss ways to work with Alderbrook – consulting, investing, newsletter sponsorships, and more – or simply want to grab a coffee and say hello, please let me know by sending a note here. I’ll get back to you soon!
In this month’s newsletter, Drew Thurlow is back with a timely guest post. Every generation or so, a new technology forces the music business to do something that some observers say it struggles with: change. As generative AI music usage increases, the stakes grow for the industry to get the rules of the road right. Drew lays out a concrete – but admittedly an imperfect – proposal to address AI licensing’s unique challenges. It will hopefully spark a conversation, because the architecture the industry builds right now – while deals are still being struck and courts are still deliberating – will likely shape who gets paid (and who doesn’t) for decades.
And you can learn even more about the topic of AI’s impact on music by purchasing Drew’s new book at this link!
And with that, on to the disclaimers…
Note: We write this newsletter to learn in public. This piece is for informational purposes only. None of this is tax, financial, investment, or legal advice. Do your own research!
Now, let’s get after it!
Jimmy
Does Music Need a New Royalty for the AI Era?
“I don’t know what they want from me, it’s like the more money we come across, the more problems we see.” - Notorious B.I.G.’s “Mo Money Mo Problems”
The music industry has a long history of treating every new way to monetize music as an invitation to create a new royalty.
That is the real pattern: not simply that technology changes things, but that technology creates meaningful new commercial uses for music that the existing system cannot measure, price, or administer. When that happens, pressure builds and eventually - usually late, usually with a fight - the industry creates a new royalty.
For example, this is what produced the mechanical royalty. In 1909, player pianos and piano rolls turned songs into a new kind of reproducible product. Music was drawing unprecedented crowds to bars and speakeasies, and songwriters wanted a share of those bar tabs.
It is also what produced the public performance right. In the 1930s, as radio fidelity increased, broadcasting went mainstream and opened up new commercial pathways for music. Songwriters, once again, wanted to be compensated. Congress agreed.
Thomas Edison’s gramophone eventually spawned the recorded music business and the master royalty. And in the late 1990s, we got the digital public performance royalty for sound recordings – after rights holders decided they were not going to let digital radio repeat terrestrial radio’s century-long omission of a master royalty in the United States. An omission, it is worth noting, that still has not been corrected.
Each of those royalties was created out of a specific set of technological and commercial circumstances. Each took years of fighting before coming into existence. None landed cleanly, and in fact, industry stakeholders still fight over them.
And now, we have Generative AI – a technology that could radically change how we make, consume, and monetize music.
How Big Is the AI Opportunity?
Before getting into how a new royalty should work, it is worth establishing what is on the table. The traditional music royalty pools are not small. And I believe (Drew here 👋) that the AI one may not be either.
For example, in 2024, global songwriter and publisher collections cleared €12.6 billion (~$13.6 billion), up 7.2% year over year, according to CISAC’s Global Collections Report. That number splits roughly into thirds: digital streaming and downloads at €5.1B, live and background at €3.5B, and TV and radio at €3.4B. The MLC alone distributed over $800 million in mechanical royalties in 2024 under the US blanket license. Global recorded-music sync was a $650 million business.
Compare that to where AI licensing is going. Goldman Sachs’ 2025 Music in the Air report pegs direct AI music revenues at roughly $400 million today, growing to $2.1 billion by 2030. That is a 30% CAGR, contingent on training and output rules getting clearer! AI music is currently around 0.1% of the royalty pool. That number is going up.
Goldman’s $2.1B estimate is direct AI revenue (i.e., what the AI music platforms generate). The royalty pool that flows back to rights holders is a percentage of that figure. So exactly how much of that potential revenue gets paid out is exactly what a hypothetical AI royalty framework – that I’ll propose below – would attempt to set.
In the below chart, I consider a few potential scenarios for the size of the AI pie. If the Goldman estimates play out as forecasted, we’re likely looking at a mid-hundreds-of-millions royalty pool for rights holders within five years. Run a more aggressive scenario where AI music reaches streaming scale by 2035, and the pool is multiples bigger.
Put another way: I see the AI music opportunity as a sync-sized opportunity in the worst case and a streaming-sized one in the best case. The architecture we build for it now determines who participates in it for decades to come.
Why Existing Royalty Models Break In the AI Era
Generative AI is not one use case. It is several stacked on top of each other. There is a potential transfer of value at:
Ingestion: when music is used to initially train a model;
Inference: when a model is prompted; and
Output: when a model generates, references, imitates, or recommends music in ways that displace or compete with existing catalog
That is a much more complex problem than “here is a new transmission medium, decide what it pays.”
It is also why some of the early licensing deals – which rely on pro-rata streaming data to determine payouts – cannot be the long-term answer. Streaming data tells you what people listened to on Spotify last month. It does not measure training contribution. It does not capture output influence. It does not describe how music functions inside a model at all. Using pro rata streaming payouts for AI generated music is convenient, unprincipled, and maybe even lazy. Clearly, it is mismatched to the technology. As AI becomes more important to the industry, this disconnect will look even more obvious.
These pro-rata licensing deals may be necessary first steps. Along these lines, I would agree that the UMG/Udio and Warner/Suno settlements are doing the important work of establishing that this use is licensable. However, these deals and settlements are not doing the harder work of establishing how different types of use gets paid out.
Answering if vs. how are two different problems. And the second one needs a structural answer. Proxies have a way of becoming permanent, and this is one the industry should not let take hold.
Why Current Legal Disputes Will Enable a Royalty to Follow
Before getting into it, there is a fork in the road that no AI royalty conversation can skip: is training on copyrighted music fair use?
That question is being litigated right now in cases involving the major labels, publishers, and AI music companies. The answer matters enormously, because it determines whether any future royalty regime will be:
Negotiated: Rights holders have leverage, and AI companies have to license catalogs they want. A market emerges around this behavior.
Mandated: Courts find training is infringing, and we need a solution. This results in private agreements and/or Congress steps in with a compulsory license and a rate-setting process.
Foreclosed: Courts find training is fair use. As a result, there is no royalty obligation outside of private agreements from AI companies that want to do the right thing or prefer to have a clean data feed.
Most of this essay assumes that the first outcome or second outcome prevails. The third one would leave rights holders with no real leverage and turn any “framework” into a polite request that Big Tech does not have to acknowledge.
However, my assumption is not guaranteed. It is one of the reasons the licensing deals being struck right now are not just about money; they are also about establishing a market practice that this use is licensable.
But Why Not Just Do Direct Deals?
It is an obvious question, especially right now. UMG settled with Udio in October 2025, agreeing to license its catalog into a new AI platform launching in 2026 - walled garden, no downloads, opt-in artist consent. Warner settled with Suno in November 2025 - “multi-million dollar” settlement, a licensing partnership, and Suno’s acquisition of Songkick thrown in for good measure. Sony is still in court, and its fair-use cases against Udio and Suno are expected to produce a pivotal ruling later this year. The Big Three have all signed deals with KLAY. ElevenLabs has locked in Merlin and Kobalt, with a 50/50 publishing/master parity split and a Most Favored Nations clause for the publisher.
Given this progress, why not just let the market handle it?
I’d argue that there are three problems with allowing direct deals between rightsholders and licensees to reign supreme.
First, direct deals concentrate. Collectives distribute. The Independent Music Publishers Forum has urged its members to reject AI deals that don’t treat recording and composition rights at parity. Meanwhile, the European Composer & Songwriter Alliance is calling for transparency in licensing agreements. This is the streaming-era complaint replayed, and structurally it is the same problem: when only the rights holders with leverage make the deals, everyone else gets a take-it-or-leave-it rate after the precedent is set.
Second, direct deals don’t solve the songwriter problem. A label can license a master. A label cannot license the underlying composition. The ElevenLabs/Kobalt 50/50 is the first deal that gets this right, and it works because the publisher had its own seat at its own table. Without a structural mechanism that forces parity, most AI deals will continue to inherit streaming’s 80/20 split and most songwriters will continue to lose.
Third, direct deals don’t measure output influence. The agreements between AI companies and the Majors cover what was ingested. They don’t touch what comes out on the other side. A song from a small independent label that demonstrably shows up in an AI generation – through melodic similarity, timbral imitation, or attribution – has no path to compensation under a model where only the Big Three’s catalogs are licensed.
Direct deals are how this gets started. They are not how it scales, and they are not how the people who actually wrote the songs get paid. The industry knows this. We built SoundExchange and the MLC for exactly this reason. We are going to need to do it again.
A Concrete (And Imperfect) Proposal
So how do we solve this problem?
What the industry should consider is a two-tier Generative AI-specific royalty. A royalty that I am calling the “Generative Usage Royalty.” Here’s how I envision this royalty working:
Tier one: a training-rights fee. Treat the use of copyrighted recordings and compositions as training data in a similar way as the industry treats sync licensing. In other words, rights holders will be paid up front for a discrete licensable use. Pricing would be set by negotiation in the near term, and ideally by a statutory rate later under a blanket license. The license would be tied to a specific model and a specific training run, with audit rights for rights holders, and disclosure obligations for licensees about what was ingested.
These training deals would have a term (like streaming deals), and any rights holder would have the ability to negotiate or opt out at the end of the term. And while streaming deal terms can typically run three years, that time horizon is too long for the exponential developments that happen in AI. Ideally, these AI deals would be one year.
Tier two: an output royalty. A percentage of revenue derived from the AI outputs paid into a collective pool. This would be a portion of the licensee’s revenue generated from a variety of sources, including subscription fees, API revenue, generation credits for rights holders (similar to how streaming services offer free ad inventory), and B2B licensing of generated music.
This royalty set would weigh three factors: (i) detected output influence (using fingerprinting and attribution modeling), (ii) use of name-and-likeness in the prompt, and (iii) a baseline catalog share. None of those signals are perfect on their own. Together they are closer to “what value did your music actually contribute?”
Both tiers would flow through a new or designated collective rights body, similar to the MLC or SoundExchange, with separate splits for the master side and the publishing side.
This is not the only plausible design. But the two-tier structure has the right shape, because it pays for the ingestion event and it pays for the ongoing commercial use, which is roughly how the rest of the music business already works.
The Master vs. Composition Splits Problem
Here is where most royalty reforms go to die, and the AI conversation will not be different.
A Generative Usage Royalty has to attach to both the sound recording (master) and the underlying composition (publishing). The starting question is the split between them. Streaming settled, after years of fighting, on a roughly 80/20 split of the rights-holder dollar in favor of the master side. It’s worth noting that a number of publishers and songwriters have been pushing back against this split ever since.
For an AI training fee, the cleanest starting point is 50/50, because the model is ingesting both the recording and the underlying work simultaneously, and there is no reason to weigh one over the other. For the output royalty, the same logic makes sense. If you want to get even more detailed, the potential adjustments could be baked in where outputs are demonstrably more influenced by composition (melodic similarity) versus master (timbral or production similarity).
This is how sync works, and the first major AI license deal we had – ElevenLabs and Merlin/ Kobalt – does split revenue 50/50 between the master and publishing.
Songwriters and recording artists also need a defined share of each side. In my opinion, this should be distinct from what their labels and publishers receive and that share should not default to whatever their existing contracts say. (Sorry label friends, but those contracts were not written with AI in mind.)
This is the part of the argument that the labels and publishers will fight hardest, and it is the part that determines whether this framework is fair to the people who actually make the music. However, we have a precedent for this: it is not unlike how SoundExchange was set up to collect digital performance royalties, where 45% of what is collected goes directly to the featured artist. At least, it’s supposed to.
There is no avoiding this fight. There is only choosing whether to have it now, while the framework is being built, or later, after the inertia of doing nothing has hardened into precedent.
Enforcement and Measurement
A royalty system is only as good as its enforcement mechanism. This proposal has to come with an enforcement architecture, and this is where it is worth being honest about what is possible with current technology:
Training disclosure is solvable but politically hard. AI companies generally do not want to publish their training data sets; doing so creates legal exposure and potentially exposes their “secret sauce.” A workable regime would require disclosure to a trusted intermediary – the collective body, an auditor, or a regulator. Under confidentiality, rights holders would be able to verify their works’ presence or absence. We have some analogs to this in AI image generation, and book publishing. Some of these bodies have even taken action when AI companies have violated agreements.
Output similarity is solvable but technically uncertain. Audio fingerprinting works well for exact and near-exact matches but degrades quickly for stylistic similarity and determining plagiarism.
Inference-level attribution – knowing which training works contributed to a specific generation – is still mired in research. There are interesting attribution methods being discussed, but nothing ready to be deployed, at least at scale. Any honest version of this proposal punts on per-inference attribution. Instead, for now, attribution should be determined by the name-and-likeness used in a prompt plus a pool distribution.
The web3 era taught the music business that clever tech does not solve all problems. Metadata hygiene, governance, adoption, and enforcement were the hard parts then, and they are the hard parts here. The proposal above is designed to work with imperfect tech and improve as the tech improves, rather than waiting for perfect measurement that will not arrive. Again, we have never let perfect be the enemy of good in the music business.
How Does a New AI Royalty Impact Different Stakeholders?
Most of the conversation about AI royalties has focused on artists and creators. That undersells what is actually at stake to music business stakeholders. A Generative Usage Royalty rearranges incentives for almost every industry player. Some of those parties are obvious winners. Some are conflicted.
Major labels. Structural winners on the master side. They own the catalog AI models most want, users want to prompt, and a defined royalty stream means a new revenue stream to bulk up profits as streaming growth slows. But they may lose some control if a statutory rate is introduced because it sets a floor. The Majors would be unable to undercut one favored AI partner, and an output royalty paid into a collective pool means they can’t take all the upside. As a result, they’ll fight the splits-to-talent piece hardest, because that is the part that comes out of their share.
Publishers and PROs. They are the biggest structural winners. The 50/50 parity baked into a statutory framework is the thing they’ve been begging for in streaming for 15 years. Getting it written into the AI royalties from day one resets the bargaining position for everything else in music and it positions PROs as natural administrators of the new collective body.
Spotify and DSPs. Interestingly conflicted. On the one hand, AI-generated music threatens to flood the platform with content that costs DSPs nothing in royalties. On the other hand, if AI outputs trigger the output royalty when they’re streamed, DSPs end up paying for content they thought was (essentially) free. The framework protects the DSP/artist relationships at the cost of its margin on AI slop. Expect them to lobby hard for narrow definitions of “covered output.”
YouTube and TikTok. Both are already AI producers (e.g., YouTube’s Dream Track, Google’s new FlowMusic, TikTok’s various AI music experiments). For them, this is a tax on a capability they were planning to get for free. Expect resistance. But it also gives them legal cover because a defined royalty is cheaper than an enterprise-killing lawsuit, which is the calculation that likely drove the UMG/Udio and Warner/Suno settlements in the first place.
AI music companies (Suno, Udio, ElevenLabs, KLAY, etc.). Winners. A sanctioned royalty framework legitimizes these companies in the eyes of artists, the industry, and investors. Suno and Udio both settled with the Majors when given the choice between paying and fighting. Many feel these generative AI companies are illegal and unethical. A formal royalty framework would go a long way to combat those narratives.
Indie labels, indie publishers, and the long tail. The biggest winners under my proposed framework. Without it, they will likely be left to negotiate alone against AI companies that have already spent their budgets on the Big Three. This is the constituency that a Generative Usage Royalty most exists to serve. It is also the constituency, paradoxically, with the least power to push it forward.
In summary, a new royalty is not a tax on the AI industry. It is the thing that lets the AI industry exist inside a music business that does not try to sue it out of existence every six months. From my perspective, each party above has a version of this deal where they win.
What This Buys and What It Costs
A Generative Usage Royalty designed along these lines will not come easy. It would create new overhead, new disputes, and new rate-setting fights. It would require labels, publishers, PROs, AI companies, lawmakers, and creator advocates to actually sit at one table, which is not a thing the industry is famous for doing well.
What it would give back is an architecture that pays for the right thing. It would price training as the licensable event it is. It would tie ongoing payments to the actual product AI companies are selling, not to a proxy borrowed from a different business (and one increasingly dominated by those at the top). It would acknowledge that creators deserve a defined share of both sides of these new revenue streams. And it would do so in a structure that can be exported internationally, which matters because AI does not respect national borders.
Closing Thoughts
The history at the top of this essay is not as clean as the industry likes to tell it. The mechanical royalty took a Supreme Court loss, and an act of Congress. The digital public performance right exists for sound recordings only because rights holders refused, repeatedly, to be steamrolled the way they were on terrestrial radio. The terrestrial sound recording performance right still does not exist in the United States, because the broadcasters won that fight and have kept winning it.
New royalties get built when reality forces the issue and when rights holders have enough leverage to make it work. They do not get built because they are obvious or because they are necessarily fair.
The AI moment may or may not produce that leverage. The licensing deals being signed now will determine a lot of it. The court rulings on training will determine more. And the willingness of labels and publishers to share the upside with the artists and writers they represent will determine whether the framework is durable or if it is litigated apart by its own beneficiaries. We know which path we usually take in music.
A Generative Usage Royalty, structured roughly as described above, is the right solution. It will be imperfect. It will be fought over. It will be revised. That has been true of every royalty in this business. We should build it now – while it is still early – and build it with intention. Otherwise, the industry will wake up to a system congealed around bad proxies and worse precedent.
Thanks to Matt and Adam for the feedback, input, and editing!
Leveling Up’s work is provided for informational purposes only and should not be construed as legal, business, investment, or tax advice. You should always do your own research and consult advisors on these subjects. In addition, our work may feature entities in which Alderbrook Companies, LLC or the author has invested, has worked, and/or has provided consulting services.
📚 Music Business, Tech, and Investing Content Worth Consuming
Here is some of the best content that our team has consumed over the past month or so –
We need to rethink what ‘indie means’ (Mark Mulligan: link): “In 2025 independent labels represented, on a distribution basis, 30% of the global recorded music market. But the term independent is becoming progressively less useful…Why? Because the influx of capital into the music market has driven the rise of a new breed of commercially-focused, global scale labels…In today’s music business, data and marketing tech & resources are the superpowers. It takes scale to build powerful enough capabilities. Most smaller indies cannot afford them.”
Michael Jackson’s Catalog Is Still Setting Records In ‘Michael’ Biopic’s Third Week (Billboard: link): “For the third consecutive week, Michael Jackson obliterates his personal-best weekly streaming count as impact from Michael biopic continues to spell massive streaming returns and more chart placements for many of his best-known songs. For the May 1-7 tracking week, Jackson’s solo catalog registered 181.6 million official on-demand song streams in the United States, according to Luminate.”
As ABS Deal Reshape The Music Rights Landscape, KBRA Says It’s Rated $12.9B In Music Royalty Backed Bonds Since 2020 - But Expects Issuance To Fall 25% In 2026 (MBW: link): “KBRA expects music ABS issuance to fall by approximately 25% in 2026, dropping from over $3.3 billion in each of the past two years to slightly more than $2.5 billion – “primarily because of continued issuer consolidation.”
If You Thought The US Music Industry Was Concerned About AI, You Should Hear What They’re Saying In China (MBW: link): “The Chinese music streaming market is facing “industry chaos” due to platforms allowing masses of copyright-infringing, AI-made tracks to fill up their catalogs…[Tencent Music Entertainment’s Executive Chairman Cussion] Pang told analysts that the “proliferation of unauthorized AI-generated content… creates headwinds for our music subscription growth” and “undermines creators’ rights and dilutes the long-term value of the music ecosystem as a whole.”
🤝 Want to Work Together? Get in touch!
Alderbrook has helped dozens of companies create value for their stakeholders. We’d love to have the opportunity to work with you too!
Here’s what we offer music publishers, labels, tech companies, and investors:
Growth and Strategy Consulting: Projects covering corporate & growth strategy, commercial & market analysis, stakeholder education materials, and more.
M&A and Investment Advisory: Commercial due diligence for buyers, fundraising support for sellers, investment sourcing, valuation assessments, fractional VP of Finance / Strategy services, and more.
Early-Stage Investing: Actively investing in start-ups, with a current focus on Pre-Seed and Seed rounds.
If you’re interested in learning more and exploring how we could work together, you can either reply directly to this email or click the link below to fill out our contact form. We’ll be back in touch soon!
📭 Share Leveling Up with a friend!
If you enjoy Leveling Up, please consider sharing it with friends and colleagues (link to share)!
Enjoyed the article!
I wouldn’t dismiss per-inference attribution. AaaS (attribution as a service) tech exists in the market, foundational to the econ/business model know as ‘attribution share’: https://formsandshapes.substack.com/p/the-ai-era-building-sustainable-ai