Sony and Universal have pinned a theoretical $9 billion price on Suno and more than $4 billion on Udio. Here's where those numbers come from, why nobody expects them to be paid, and what AI builders should take away.

If you run an AI company, this number should get your attention: $9,153,900,000.

You get it by multiplying 61,026 sound recordings by **$150,000 **each, which is the highest statutory damage US law allows for willful copyright infringement. That's the theoretical maximum Sony Music and Universal Music Group have been pointing at Suno

No serious person expects Suno to pay nine billion dollars. The number was never supposed to be realistic. Its job is to be scary, and it's doing that job well.

I'm not a musician. My work is data pipelines and scrapers. Still, the AI music lawsuit fight is the best preview I've found of how training-data disputes will go for the rest of us. Let's walk through it.

How we got here, fast

Here's a quick summary of a long and messy timeline:

June 2024: The RIAA sues Suno and Udio for the major labels.
October 2025: Independent artists begin filing class actions against Suno and Udio.Later suits go after Mureka and Google's Lyria 3.
Late 2025: Warner Music settles with Suno, and Udio signs licensing deals with Universal and Warner.
May 21, 2026: Sony and UMG ask to add 61,026 recordings to their Suno case. They identified them with the Audible Magic fingerprinting service.
June 29, 2026: A judge won't let Sony add tens of thousands of new works to its Udio case.
July 20, 2026: Sony files a brand new lawsuit against Udio that covers more than 30,000 recordings.
September 2026: Sony and UMG sue Suno a second time, now aiming at its new v6 model.
So this is a whole wave of lawsuits, and each new one is bigger than the one before.

Where the $150,000 number comes from

Under US copyright law, a rights holder can pick statutory damages rather than proving what they actually lost. The normal range is $750 to $30,000 per work. When the infringement is willful, a court can award up to $150,000 per work.

Everything hinges on that per-work rule. How much a label lost on any one song doesn't matter. The math simply counts songs.

Here's how that counting adds up:

Original Suno complaint: 560 works, about $84 million at the max
Expanded Suno list: 61,026 works, about $9.15 billion at the max
Sony's second Udio suit: 30,000+ works, about $4.5 billion at the max

With statutory damages, your training dataset becomes a bill, and every file is a line on it.

The new twist: "model laundering"

To me, the September Suno suit is the most interesting one, because it targets a new model.

Suno says v6 was trained on licensed material, community interactions, and what its team learned. The labels say there's more to it. Their complaint claims Suno used knowledge distillation, meaning v6 learned from the outputs of older models that had been trained on unlicensed tracks. Variety quotes the complaint calling v6 "the fruit of the same poisoned tree."
If the court accepts that argument, it changes a lot. Retraining alone wouldn't wipe a model's history clean. The taint follows the lineage.

A huge number of AI products today come from distilling a large model into a smaller one. If courts decide distilled models inherit their parent's legal problems, plenty of companies will need to take a close look at their family tree.

Why the labels won't actually collect $9 billion

Headlines love these figures, so I want to be clear: the maximums are upper limits, and nobody should read them as forecasts.

Here's why they almost never get paid in full:

Willfulness has to be proven, and courts usually award much less than $150,000 per work
Judges resist cases that keep ballooning, as the Udio ruling showed
Fair use is still unsettled, and results vary from case to case
Most of these fights wrap up with licensing deals rather than damage awards
That last point is the pattern to watch. Warner settled and signed a license. UMG and Warner licensed Udio. Suno now has deals with Warner, BMG, and Believe. These lawsuits act less like a path to a verdict and more like a negotiation backed by a very big stick.

The artists who aren't in the room

The label cases grab the headlines, but the independent artist cases could end up mattering more.

The Class actions filed since October 2025 speak for songwriters, session musicians, producers, vocalists, and small labels. Their argument is straightforward: the systems learned from full tracks, vocal stems, and lyric sheets, and now produce music that competes with the very people who made the originals.

This matters because a major label's licensing deal only covers that label's catalog. A session player who appeared on hundreds of recordings, or an indie artist who released music alone, sits outside that deal. Warner or Universal getting a seat at the table doesn't give them one.

Even if every major label signs on, a long tail of rights holders can still sue. For AI companies, "we signed the big three" makes the risk smaller, but some of it is still there.

It also leads to a tougher question that nobody has solved yet: how do you fairly pay thousands of small creators for a model trained on millions of files? Per-stream royalties took years to work out, and per-training-example royalties haven't even begun.

What this means for AI companies outside music

This is the part that matters most to me. Music is only the first test. The same reasoning covers images, code, news articles, product data, and anything else collected at scale.

Your exposure is your file count

Because damages are counted per work, a company trained on 10 million copyrighted items has an exposure on paper that sounds ridiculous. That figure can still surface in investor due diligence, acquisition talks, and insurance quotes, even if it never reaches a courtroom.

Provenance is becoming a product feature

Saying "here's exactly where our training data came from" is becoming something you can sell. Udio and Suno both advertise their licensing deals now, and people buying AI tools are starting to ask about it.

Fingerprinting makes hiding hard

The labels used Audible Magic to find 61,026 of their tracks. Similar tools exist for images, text, and code. Betting that nobody will check your dataset is a bad bet.

People used to ask "is this legal?" Now they ask "can you prove where it came from?"

A practical checklist if you train or fine-tune anything

Scraping data for clients is my everyday work, so I've thought about this a lot. This is what I'd tell any small AI team.

Keep a data manifest. Record the source, the collection date, the license or terms, and how you used it. It's a dull spreadsheet that pays off later.
Separate training data from reference data. Collecting public prices for analysis carries a very different risk from feeding creative works into a generator.
Check the terms of every source. Just because something is public doesn't mean you can train on it.
Track your model lineage. If you distill from another model, note which one it was and what it was trained on, as best you know.
Prefer licensed or open datasets for anything that produces output competing with the original creators.
Talk to a lawyer before launch, instead of after a demand letter arrives. I'm not a lawyer, and this isn't legal advice.

Dates worth putting in your calendar

If you'd like to keep up without reading every filing, these are the key dates in the main Suno case, according to Music Business Worldwide:
September 30, 2026: fact discovery and depositions end. Any evidence about training data is now on the record.
April 9, 2027: dispositive motions are due. Each side will ask the judge to decide big questions, such as fair use, without a full trial.
Keep an eye on whether the court allows the 61,026-work expansion. The Udio judge rejected a similar request. If the Suno judge does the same, expect more brand-new lawsuits instead of bigger existing ones, just like Sony did with Udio.

The honest read

My feelings about this fight are mixed.

I understand why musicians are angry. If your work were used without permission to build a machine that competes with you, you'd be furious too. The independent artist class actions seem especially fair to me.

At the same time, "$9 billion" doesn't help anyone understand what's really happening. It's a number built to apply pressure. What actually happens, over and over, is a licensing deal where the biggest labels get paid and smaller creators wait to see if anything reaches them.

For AI builders, the lesson is to keep building and know your data the way you know your finances. In 2026, your dataset can turn into an itemized bill, and the people holding the receipts have gotten very good at counting.

Enjoyed this read? Be sure to follow me for more stories that make people feel things. You can also connect with me on LinkedIn or Telegram @kawsarlog for more updates on my work.

I do freelance scraping projects, so if you need some data collected feel free to contact me on LinkedIn or email at hi@kawsarlog.com