AI Music Licensing Deals: Who Signed and How They Pay

All three major record companies now license their music to AI companies, and not one of their deals has a disclosed price. Universal Music Group and Warner Music Group turned their 2024 lawsuits against the AI song generators into licences: Universal settled with Udio in October 2025 in a " compensatory legal settlement" plus licences for its recordings and songs, and Warner settled with Udio and then Suno a few days apart in November 2025. Sony Music has licensed Klay Vision and, with the other two majors, Stability AI, but is still suing Suno, and was reported in July 2026 to have sued Udio a second time ( the record). What the public record does show is how the money moves: settlement payments, equity, royalties on what fans make, and opt-in by artists and songwriters.

This guide is for anyone deciding what to do with a music catalogue: a record label, a music publisher, a production-music library or a distributor weighing whether to license to AI companies, and how. It is also for reporters and for the AI companies' own teams who need the deals in one place. It lists every AI music licence in our deal ledger with what was disclosed; traces how the lawsuits against Suno and Udio became licences; sets out the cases still open, from the lyrics suits against Anthropic to GEMA's win against Suno in Munich; explains how the deals pay, as far as any record states it; who can license which right in a song; and what all of it means for a smaller catalogue.

AIData.Broker keeps the deal ledger and the case ledger this guide is built from, so every deal and ruling below links to its record and its source. This guide is research from public sources, not legal, tax or investment advice.

Contents

1. Every AI music licence on record
2. How the lawsuits turned into licences
3. The cases still open
4. What a deal pays, and how
5. Who can license what
6. What it means for a smaller catalogue
7. Questions people ask
8. What this means for you

Every AI music licence on record

The music deals are recent. Before October 2025 the only licences in our ledger that covered music at all were Shutterstock's library deals, which license images, video and music together, ProRata's search engine, which lists Universal among its partners, and two training licences to ElevenLabs from Merlin and Kobalt. Then the settlements came, and within a year every major label group had signed with at least one AI company, and so had Kobalt (which calls itself the world's largest independent music publisher), Merlin (which licenses digital services for independent labels) and the distributor Believe.

The table lists every licence in the deal ledger whose licensor is a music company, plus the two Shutterstock deals that include music, in date order. The column on money shows exactly what the record states; for every licence from a music company that is a phrase such as "not disclosed", a revenue share, or a figure that is not a licence fee. Most rows rest on a party's own announcement or on a named publication quoting one; the rows marked reported rest on press reporting, and are weighed that way.

One more announcement belongs beside the table although it is not a licence in our ledger. On 16 October 2025 Spotify said it would develop AI music products with Sony Music Group, Universal, Warner, Merlin and Believe through "upfront agreements, not by asking for forgiveness later", and that "artists and rightsholders will choose if and how to participate" ( Spotify's announcement). The Spotify rows above are the licences that followed it.

What the table shows

No music licence has a price on record. Across every kind of rights holder, our ledger holds 130 disclosed AI licensing deals and only 23 of them carry a dollar amount; none of those is a licence from a music company. The nearest figures are Shutterstock's, and they price a library of images, video and music together, not music. Stability's round is a sum invested in Stability, not a fee paid to the labels, and the record says so in its own words ( the record). Anyone quoting a price for a label's AI licence is quoting something no party has published.

The licensees are music companies, not the general AI labs. Suno, Udio, Klay, Stability, ElevenLabs and GRAI build music generators; Spotify builds a tool for fans. In the news deals the licensees are OpenAI, Meta, Google and Microsoft; in music no general-purpose AI lab has licensed from a music company, and they appear only as licensees of Shutterstock's mixed libraries and as defendants. Anthropic has argued in court that the publishers' licences with Udio and Klay sit in a different market from training a general model ( the record), which is a reason the line matters.

The deals license two kinds of use. Some license training (Klay, ElevenLabs's first deals, Stability in 2026, Suno v6, GRAI). Others license a product that lets users transform existing songs (the Spotify tool, ElevenLabs's fan platform with Universal), where the licence follows the output and pays on use. Several, starting with Universal and Udio, do both. Only one, ProRata's, is the kind of licence to show content in answers that most news publishers sign.

How the lawsuits turned into licences

The music licences did not come from negotiation alone; most of the largest came out of litigation. In June 2024 the three major label groups, coordinated by the RIAA (the US recording industry's trade body), sued Suno in Massachusetts and Udio in New York, alleging both trained their generators on the labels' recordings without permission and seeking up to $150,000 per infringed work, the statutory maximum for willful infringement under 17 U.S.C. 504(c). Statutory damages run per work, so the claimed exposure multiplies with every work in a catalogue, and the suits became the frame for the negotiations that followed.

The defence was fair use, the US doctrine that allows some unlicensed copying, judged case by case on four factors. The fourth factor asks about the effect on "the potential market for or value of the copyrighted work". That made the licences themselves part of the dispute: every licence a label signs can be offered as evidence that a market to license music for training exists.

Universal and Warner settle, late 2025

Universal settled first. On 29 October 2025 it announced that it and Udio had settled, with " a compensatory legal settlement" and new licences for recorded music and publishing, and a new platform for 2026 "trained on authorized and licensed music" ( the case record). Universal's own release adds that Udio's existing product would stay available in the meantime "with creations controlled within a walled garden", with fingerprinting and filtering added before the new service launched ( Universal's announcement).

Warner followed with two settlements a few days apart. It settled with Udio on 19 November 2025 and agreed to co-develop a licensed platform with an opt-in model for artists ( the record). On 25 November it settled with Suno in a broader partnership ( the record). Warner's announcement sets out what Suno agreed to change in 2026 ( the record):

1. Licensed models. Suno would launch new licensed models and deprecate its current ones when they arrive.
2. Downloads. Downloading audio would require a paid account, songs made on the free tier would not be downloadable, and paid users would get monthly download caps.
3. Opt-in for artists and songwriters over "the use of their name, image, likeness, voice and compositions in new AI songs."
4. An asset. Suno bought Songkick, Warner's concert-discovery service.

None of the three settlements states an amount. The case ledger records each as "not disclosed", and the ledger's count of 9 settlements over training data includes them.

What the settled companies did next

Udio went on to sign the independent sector. Kobalt licensed it in April 2026 ( the record), and by July 2026 it had also signed training licences with Merlin, which licenses digital services for independent labels, with the distributor Believe, and with the National Music Publishers' Association, as listed in Sony's second suit ( reported).

Suno rebuilt its models. In August 2026 it licensed data from BMG for Suno v6, which Suno says is not trained on the data used for earlier versions ( the record), and in September it launched v6, "developed using licensed data from music labels and distributors such as Warner Music Group, BMG, and Believe" ( the record).

Who is still suing

The settlements left most of the claims standing. Universal and Sony are not reported to have settled with Suno, nor Sony with Udio ( the record). Their case against Suno has grown since:

• In August 2026 the court let the labels add a claim that Suno got its training audio by stream-ripping YouTube, getting round a measure that blocks downloading, a claim under the anti-circumvention law in 17 U.S.C. 1201 that is separate from copying ( the record).
• In September 2026 Sony and Universal filed a second suit aimed at the v6 models, reported to name 60,202 recordings ( reported by Variety).
• In October 2026 the labels opposed a motion by Lightspeed, the investor that led a Suno funding round, to quash a subpoena about what it knew of Suno's training data, arguing that this goes to willfulness and statutory damages ( the record).

Publishers have joined in. Round Hill Music sued Suno and the web data company Bright Data in August 2026 ( reported), and Reservoir Media sued Suno in October 2026 over 1,050 compositions, totalling statutory damages at $157,500,000. Reservoir relies on Suno's own admission in the labels' case that it trained on "essentially all music files of reasonable quality that are accessible on the open Internet" ( the record).

The deals as evidence

Both sides now use the licences in court. Reservoir argues that Suno's settlement with Warner, its deal with BMG and its partnership with Believe show that a functioning market for licensing music as training data exists, so that Suno's use of unlicensed works is not excused ( the record). Sony makes the same argument against Udio: because Udio signed with Universal, Warner, Kobalt, Merlin, Believe and the publishers' association, Sony argues, in Unite.AI's words, that "a licensing market for AI training data plainly exists" ( reported).

Anthropic argues the other way in the lyrics case. Its economic expert says "no market exists or is likely to form to license training data for general-purpose LLMs", and it calls the publishers' own licences with Udio and Klay a different market, that of single-purpose music generators ( the record). For a holder the consequence is concrete: a licence is no longer only revenue, it is an exhibit, and a rights holder in litigation has a reason to sign deals that show a market exists, whether or not it publishes their terms. Our guide to fair use and AI training sets out how the courts have weighed the market factor so far.

Licensing one, suing another

As with the news publishers, the same companies license some AI firms and sue others. The table puts each one's licences beside its suits.

The cases still open

Our case ledger tracks 123 lawsuits, rulings, settlements and regulatory actions over training data. Two clusters of music cases matter most for a catalogue holder: the music publishers' lyrics suits against Anthropic, which test whether training a general chatbot on songs is fair use, and the suits against the song generators, where GEMA has the only judgment so far.

The lyrics suits against Anthropic

The lyrics cases are mostly music publishers' suits, not labels', because a text model copies words, and in US law the words of a song belong to the musical work, which copyright treats as " musical works, including any accompanying words". The oldest, Concord I, is also the furthest along. Anthropic has asked the court to hold that training Claude on the lyrics of the publishers' songs is fair use as a matter of law, and argues that a lost licence fee is not harm the law recognises ( the record). At the hearing on 21 October 2026 the judge questioned the publishers' claim that Claude's ability to write lyrics threatens their market, and did not rule ( reported).

The later suits lean on piracy rather than on training as such. Lyrics II ( the record), Sony Music Publishing's and Warner Chappell's suit ( the record) and Reservoir's ( the record) all allege that works came into Anthropic's hands through torrenting from pirate libraries, building on what came out in the authors' case, Bartz v. Anthropic. That case ended in a settlement that puts a price on unlicensed training, about $3,000 per work out of a $1.5 billion fund for books. Reservoir says at least five suits over musical compositions are pending against Anthropic in the same court ( the record).

GEMA in Munich

The only judgments on the merits of AI music claims in our case ledger come from Germany, and both went to GEMA, which brings claims for the composers, lyricists and publishers it represents. In November 2025 the Munich Regional Court I held that memorising lyrics in OpenAI's models and reproducing them in ChatGPT's outputs infringed, and that this was not covered by the text and data mining exception, the EU rule that lets lawfully accessible works be mined unless the right holder has expressly reserved them ( the record).

In July 2026 the same chamber largely ruled for GEMA against Suno. It found that Suno had stream-ripped the works from YouTube, getting round a protection measure; that the works were memorised in its v3.5 and v4 models, which sat on servers in Germany; and that, applying US law to the training done in the US, fair use did not apply because Suno's outputs were substantially similar to the originals ( the record). Neither judgment is final, and neither states an amount of damages. What they establish for now is that, in Germany, a music generator that can reproduce a song from a short prompt is exposed wherever it trained.

What a deal pays, and how

No record in our ledger states what any AI company pays any music company, so the question of how much has no public answer. What the records do state is the shape of the payment, and the shapes differ more than in news, where the headline deals are fixed fees for a term (see our news guide). In music the money comes as a settlement, as equity, as royalties on what users make, as a share of a product's revenue, or as fees for a dataset, and several deals combine them.

The table sets out each form with the deals on record that use it. Where a record states a figure, it is linked; where it states none, the cell says so.

Settlements, and the price of not settling

The settlements are the least known of these payments. Universal's word for its Udio payment, "compensatory", is the only description on record ( the record). What frames them is the alternative: statutory damages of up to $150,000 per work for willful infringement, which Reservoir multiplies out to $157,500,000 for its works against Suno alone. No music settlement figure is public, but the settled books case puts one benchmark on what unlicensed use has cost, about $3,000 per work, against a statutory ceiling of $150,000 per work. That benchmark is for books in a class action, not for music.

Equity

A structure that sets music apart from news is that the labels take a stake. In Stability AI's funding round in August 2026, Sony, Warner and Universal invested and "also licensed their catalogs for training as part of the deal" ( the record). The record notes that Stability's own announcement names Electronic Arts as a further backer and that the round size rests on TechCrunch alone. Equity aligns the rights holder with the product's success, but it pays mainly if the company succeeds and is sold or listed, and a holder valuing such an offer should value the licence and the shares separately.

Royalties on what users make

The Spotify deals pay on output rather than on input. Believe's announcement says "streams of these remixes will generate royalties for the participating artists and rights holders behind the original works", and that the tool is a paid add-on for Premium subscribers, "creating an additional revenue stream" on top of what artists already earn on Spotify ( the record). Kobalt's deal ties the same tool to "participating songwriters" ( the record), and Merlin's lets artists on its member labels choose whether to take part ( the record). No announcement states the rate, the split between recording and publishing, or whether there is a minimum.

Opt-in by artists and songwriters

Opt-in is the term that recurs most across the announcements that describe their terms, and it is a payment term as much as a consent term: only the works and artists who opt in earn. Warner's deal with Suno gives artists and songwriters "full control over whether and how their names, images, likenesses, voices, and compositions are used in new AI-generated music" ( the record). Believe applies a two-step consent: artists first opt in to being considered for AI programmes, then decide on each partnership ( the record). MassiveMusic trains GRAI only on a dataset rightsholders approved for training through its opt-in system, and the training licence "does not itself authorize downstream creative uses" ( the record). For a catalogue holder this means the size of what it can license is the size of what its artists and writers agree to, not the size of its catalogue.

What an agent or a library keeps

Holders who license through someone else give up a share, and the few published shares are not music-specific. A licensing agent for catalogues of text, audio, video and data, Credtent, says it sends 85% of gross revenue from every deal to the rights holders ( the lever). Shutterstock pays its contributors a 20 percent average royalty on data-licence revenue, split by the volume of each contributor's content in the datasets sold. Our page on negotiating sets these splits beside the ones in books and news.

Who can license what

A song is not one asset. Before any catalogue can be licensed for AI, its holder has to know which rights it holds in each recording and each song, and which belong to someone else. Most of the deals above were signed twice for that reason, once for the recordings and once for the songs.

Masters and compositions

US copyright law lists " musical works, including any accompanying words" and "sound recordings" as separate categories of work. The musical work is the composition, melody and lyrics, and is usually held by a music publisher on behalf of the songwriters. The sound recording, the master, is usually held by a record label. In the majors the two sit in different companies of the same group (Universal Music and Universal Music Publishing; Sony Music and Sony Music Publishing; Warner Music and Warner Chappell), and the AI deals reflect it: Klay signed each major's label and publishing arm separately ( the record), and Udio's deal with Universal covers "recorded music and publishing" ( the record).

The two rights reach different things. A label's right in a master is limited to copies that " recapture the actual sounds fixed in the recording", and it does not reach a new recording that only imitates them. Training on audio copies the recordings themselves, which is the labels' claim against the generators. An output that sounds like a song without copying its audio is outside the master right, and what protects the song then is the composition, which is why the publishers' suits and GEMA's case also target outputs that reproduce the work, and why the deals add opt-in for an artist's voice and likeness.

Artists' and songwriters' consent

A label that holds a master does not automatically hold everything a buyer wants. Whether a recording contract signed before generative AI lets the label license the artist's voice for a model is a question of that contract. The deals answer it by asking the artist: Warner's settlements with Suno and Udio both add an opt-in for artists ( the record). In California a contract clause that lets someone create a digital replica of a performer's voice in place of their work is unenforceable unless it describes the intended uses with reasonable specificity or the performer had counsel or a union ( Labor Code 927).

Time limits the grant as well. An individual author who grants rights, however "perpetual" the contract says they are, can end the grant in a window that opens 35 years after signing, unless the work was made for hire. A label or publisher licensing older works that came to it by assignment should check which grants are open to termination before promising a buyer a long term. Our guide to data licence term length covers the clause.

Collecting societies

In Europe many composers and publishers give their rights to a collecting society (GEMA in Germany is the one in court), which licenses and sues on their behalf. That is why GEMA, not a publisher, won the Munich cases. What a society may license depends on what its members gave it. In the UK the rules require that a right holder consent "specifically for each of the matters" a society manages, in documentary form, and let it withdraw the rights of its choice on reasonable notice ( the regulations). A publisher whose writers belong to a society should read the membership terms before licensing training rights itself, and a buyer will ask.

EU law adds a step that a right holder can take without any licence. The text and data mining exception applies only if the use "has not been expressly reserved by their rightholders", for content online by machine-readable means, and a provider of a general-purpose model in the EU must have a policy to identify and comply with those reservations. There, reserving rights is what turns a catalogue from material a model may mine into material it has to license. Our rights hub covers the rules country by country.

Distributors and aggregators

Independent labels and self-releasing artists mostly license through someone who already represents them. Merlin licenses digital services for independent labels and distributors and signed ElevenLabs, Udio and Spotify, with artists on its member labels free to choose whether to take part. Believe, a distributor, and its platform for self-releasing artists, TuneCore, signed Udio, Suno and Spotify, and for Spotify's tool it asks its artists through a two-step consent ( the record). Kobalt, a publisher, licenses for participating songwriters ( the record). And MassiveMusic, part of Songtradr, runs an opt-in system that turned rightsholders' approvals into a training dataset for GRAI ( the record). For a small label these are the routes already open: if it is a Merlin member or distributes through Believe, it may be able to opt in to a licence already signed.

Production libraries that hold their catalogue outright

A production-music library, which writes and records music to license for film, television and advertising, can be in the strongest position of all, because it may hold both rights in every track. Whether it does depends on how each track was made:

• Music written by the library's employees in the course of their work is a work made for hire, and the employer " owns all of the rights comprised in the copyright" unless a signed writing says otherwise.
• Music commissioned from freelance composers is a work made for hire only in the categories the law lists, such as part of a motion picture or other audiovisual work, and only with a signed agreement that says so ( the definition). A track commissioned as stand-alone library music may not fit.
• Otherwise the library needs a written assignment: a transfer of copyright " is not valid unless" it is in a signed writing.

A library that can show that chain for every track, and holds the masters as well, can license training on its own signature. That is scarce. The large stock libraries, whose music sits inside collections that also hold images and video, were among the first to license: Shutterstock's data licensing page lists 6 million or more music tracks and sound effects beside its images and footage.

What it means for a smaller catalogue

The deals on record were signed by the largest holders in the business, but the buyers behind them have not stopped at the majors. Udio's and Suno's licences with Merlin, Believe, Kobalt and BMG show the generators going past the majors to the independents within months of the first settlements, and GRAI trains on an opt-in dataset that MassiveMusic assembled from rightsholders' approvals. A smaller holder has three ways in: opt in through an aggregator that already has a deal, license direct to a music AI company, or license through a library or agent that licenses on to AI companies.

Who takes catalogues

The general-purpose AI labs are not on the list. None has licensed from a music company in the ledger (their only music comes inside Shutterstock's library deals), and one of them argues in court that no such market exists for its kind of model ( the record). Our buyer board and vendor register list who buys other kinds of data.

What a catalogue is worth

There is no public price for licensing a music catalogue for AI training, per track, per hour or per catalogue: no party to a deal has published one. The nearest public figures are not music. A studio's catalogue of sound effects, 53,000 files, lists at €60,000 for a one-off purchase or €20,000 a year on one data listing site and at $78,900 on another; those are asking prices for sound effects, not sales of music. A musician and YouTuber, Peter Hollens, was reported to have made $33,000 in a year from licensing a thousand hours of old music videos to AI companies, a sum for video with music in it, not for a music licence ( reported by Semafor). The speech prices in our audio guide do not transfer either: a buyer of speech pays for hours of voices, a buyer of music for rights in songs.

What does carry over from the deals is what raises a catalogue's value in a buyer's eyes: both rights under one signature, the consent of the artists and writers already collected, clean provenance (the labels' case against Suno turns in part on how its audio was obtained), and files a model can learn from, such as multitrack stems and rich metadata. GRAI, for one, says it uses MassiveMusic's metadata on mood, energy, instrumentation, key and era for search and recommendations ( GRAI's announcement).

A worked example: a production-music library

Example. Take an independent production-music library in the US: an instrumental catalogue built over two decades, written partly by two staff composers and partly by freelancers, recorded in its own studio, with stems for most tracks and detailed metadata. It has never licensed for AI and has had an approach from a music start-up.

1. Sort the catalogue by chain of title. The staff composers' tracks are works made for hire, and the library owns all of the rights in them. For the freelance tracks it pulls each agreement: a signed assignment works ( a signed writing); a "work made for hire" label on stand-alone library music may not, because such music is not one of the listed categories. Tracks with neither go to a separate pile to clear or leave out.
2. Check what the composers kept. Some freelancers may have reserved rights, or registered with a collecting society that holds rights for them. Where a society is involved, its terms decide whether the library can license training ( in the UK, each right is handed over by specific consent).
3. Take out what is not the library's. Samples, licensed loops and session vocalists whose releases predate AI come out of the lot, or need the other party's consent.
4. Decide what to license. Training only, as in the MassiveMusic licence that " does not itself authorize downstream creative uses", is the narrowest grant and keeps the library's own product business separate. A licence that also lets users generate from the catalogue is a different deal and should pay on use.
5. Know the anchors and their limits. No music price is public, so the library's leverage is its rights, its stems and its metadata, and the fact that more than one buyer licenses music now. It should get competing offers in writing rather than take the first one, and read our guide to exclusive and non-exclusive licences before granting exclusivity.

A second example: an independent label

Example. An independent label holds the masters of its releases; the songs are published partly by its own small publishing company and partly by outside publishers; it distributes through an aggregator that is a Merlin member. It can license its masters itself, but a generator that trains on the recordings also copies the songs in them, and the buyers on record sign publishers separately. Its quickest route is the one already open: opt in to the licences Merlin has signed, after asking its artists through a consent step like Believe's. A direct deal makes sense for the part of the catalogue where it also holds the publishing, so the buyer gets both rights in one place.

Our part

AIData.Broker values a catalogue against the public record and, for a holder based in the United States, the United Kingdom and Canada, describes it without the holder's name to every buyer that fits at the same time; the holder licenses it itself, and pays our fee, set out on the pricing page, only out of a sale.

Questions people ask

What AI music licensing deals have the record labels signed?

All three majors have signed. Universal licensed Udio, Stability AI, Klay, ElevenLabs and Spotify's fan remix tool ( Udio, ElevenLabs); Warner licensed Suno, Udio, Stability AI and Klay ( Suno); Sony licensed Klay and, with the other two, Stability AI ( the record). None of the deals has a disclosed price.

What is the Warner Music AI deal with Suno?

Warner settled its lawsuit against Suno in November 2025 and became its partner: Suno agreed to launch licensed models in 2026 and retire its old ones, to limit downloads to paid accounts with monthly caps, to let artists and songwriters opt in to the use of their names, likenesses, voices and compositions, and it bought Warner's Songkick ( the record). No amount was disclosed ( the case record). Warner settled with Udio the same month ( the record).

What is the Universal Music AI deal with Udio?

Universal settled its copyright case against Udio in October 2025 with " a compensatory legal settlement" and new licences for its recorded music and publishing, and the two agreed to launch a subscription platform in 2026 trained on authorised and licensed music. Universal has since licensed ElevenLabs, Stability AI, Klay and Spotify's tool, and is still suing Suno ( the record) and, through its publisher, Anthropic.

Has Sony Music signed an AI deal?

Yes, with Klay Vision in November 2025 ( the record) and with Stability AI in August 2026, when Sony, Warner and Universal invested in Stability and licensed their catalogues for training ( the record). Sony is not reported to have settled with Suno or Udio: with Universal it filed a second suit against Suno over the v6 models ( reported), and it was reported in July 2026 to have sued Udio a second time ( the record).

How do AI music licensing royalties work?

The only royalty structure described on record pays on output: each stream of a fan-made remix on Spotify's tool generates royalties for the participating artists and the rights holders behind the original ( the record). Other deals pay through settlements, licence fees and equity, and in the deals with an opt-in only those who opt in earn. No deal states a royalty rate.

Is Suno licensed now? (Suno licence)

Partly. Suno settled with Warner and says its v6 models, launched in September 2026, were developed with licensed data from Warner, BMG and Believe ( the record), but Universal, Sony, Round Hill and Reservoir are still suing it, and GEMA won against it in Munich ( the record). For users, Suno's terms assign whatever rights Suno holds in a song to Pro and Premier subscribers while promising that no copyright exists in it, and keep free and basic users to personal, non-commercial use ( Suno's terms).

Is Udio licensed? (Udio licence)

Udio has licences from Universal and Warner, both of which settled with it, and from Kobalt, Merlin, Believe and the National Music Publishers' Association ( Kobalt; the others, reported). Sony is not reported to have settled, and its second suit uses Udio's own licences to argue that a market for training licences exists. Udio's terms for users were not read for this guide.

How much do AI companies pay to license music?

No one has said. Not one music licence in our ledger carries an amount; the only published figures near it are Shutterstock's early range for whole libraries of images, video and music, $25 million to $50 million each, and the statutory damages a company risks without a licence, up to $150,000 per work for willful infringement.

What this means for you

If you hold a music catalogue, the public record gives you the structures, the buyers and the legal pressure behind the deals, but no price. The steps that follow from it are the same for a label, a publisher and a library.

1. Map your rights before you talk to a buyer. For each recording and each song: who holds the master, who holds the composition, how it came to you (made for hire, assignment, licence) and what the artists and writers kept. The datasheet and our guide to training data provenance set out what a buyer will ask for.
2. Reserve your rights. In the EU, a machine-readable reservation is what takes your catalogue out of the mining exception; everywhere, a clear position helps if you later need to enforce. Licensing one company and suing another is normal in this market.
3. Collect consent as you go. The deals that describe their terms pay only for the artists and writers who opt in, so the consent you hold is the catalogue you can sell.
4. Separate the uses. A training-only licence, a licence for a product that transforms your songs, and a release of past claims are three things; price each, and do not let a buyer fold them into one number. Our licence terms set out where to stand on each clause, and what is a data licensing agreement explains the contract.
5. Value equity and cash separately, and read any royalty for its base (what it is a share of), its rate and whether it has a floor.
6. Use an existing route if you have one. If you are a Merlin member, distribute through Believe or TuneCore, or publish through Kobalt, ask what you can opt in to before negotiating alone.
7. Get a number from the record, not from a headline. Our valuation puts a range on what you hold from the public prices that exist, and the offer check compares an offer you already have with the record. For a sale we run, we take on a lot only where the expected sale reaches $48,000; below that, an aggregator or a direct licence is usually the better route, as the page on going direct or through us explains.

For the rest of the market, see our guides to AI licensing deals with news publishers, Reddit's data deals, selling videos to AI companies (which covers the music inside footage) and how big the AI training data market is. Every deal and case here is on the records page with its source, and how we source every number explains the method.