AI Companies Are About to Relearn Napster's Most Expensive Lesson

What is a song worth to a machine that already learned from it?

A court in Munich started to answer that on July 31, 2026. The 42nd Civil Chamber of the Munich Regional Court ruled that Suno infringed copyrights held by GEMA, the German music rights society, by training its music models on protected works without a license. Suno must disclose the revenue tied to that use. Suno must pay damages. Suno can also appeal, so nothing here is settled yet.

It is GEMA's second win against an AI company in under a year. A Munich court sided with GEMA against OpenAI last November over song lyrics reproduced in ChatGPT. This time the target is training data itself, and the court added something new: training conducted in the US does not shield a company once the output is stored and reproduced in Europe.

I have watched this movie from inside the booth. Early in my career I worked at Liquid Audio, and then at Napster. I remember the hallways. Smart engineers, a product people loved, and a slow certainty settling over the building that the law would arrive before the business model did. It did.

Strip away the legal language and the Munich ruling says something plain. Training data is an input. Inputs have owners. Owners get paid.

For years the industry has argued that a model learning from a work is not the same as copying it. Courts in several countries are testing that argument. In Munich it lost.

The real question under this headline is older than AI. Who captures the value that a new technology creates?

Every song Suno trained on had a writer. Someone sat with a guitar in a small room and finished a chorus. GEMA exists to collect for those people. That is the human weight behind the word "damages." A disclosure order is an accounting exercise to a court. To a working songwriter, it is the first time anyone has been made to show the receipt.

Here is where the history rhymes, and I lived this one.

The music business met its first digital disruption with a sequence I still see in boardrooms. Ignore it. Then shame the people using it. Then regulate it. Napster opened in 1999. A court shut it down in 2001. Apple opened the iTunes Store in 2003 and sold songs for 99 cents. The demand never went away. It got a price, and the industry lost years it did not need to lose.

I am not here to pile on Suno. Different company, different decade, different product. The response is the part that rhymes. What Munich did was pull the licensing reckoning forward. That is a gift, even to the companies paying for it. Every year a technology runs without a price is a year of value nobody can plan around.

What changes for AI companies that train on music

For an enterprise leader, this lands as a supply chain story. Your models learn from something. Somebody owns that something. If your legal team cannot name the source of your training data today, you are carrying a bill you have not seen. That is the same position Napster was in, and the bill came anyway.

This is where judgment earns its keep. The work that lasts sits Above the Algorithm: deciding what your company feeds a model, and who signs their name to that choice. No tool makes that call. A person does, in a meeting, on a Tuesday.

And if you do not own the model or the budget, this still belongs to you. The work you paste into a tool at your desk came from somewhere too. Ask where, and ask out loud. That question is cheap now and expensive later.

Zoom out and the stakes get wider than one company. The value AI creates has to land somewhere. It can pool at the top of a few platforms, or it can flow back to the people whose work made the models useful. Courts settle who pays. Leaders decide who learns.

That second part is the one boards keep skipping. A licensing bill is a one-time cost. The skills your people need to work with licensed, governed AI have a half-life, and that clock runs whether or not you appeal.

If you want the wider view on how rules like this land on enterprise teams, I keep it here: AI governance.

I do not know how the appeal goes. Nobody does. But the question travels well, and it costs nothing to ask in your own building. Where did our training data come from, and who would we owe if a court asked us that on Monday? Take it into your next meeting and watch how fast the room answers. The answer you get is the real state of your AI program. And if you see it differently, I am easy to find.

Sources: Munich Regional Court (Landgericht München I), 42nd Civil Chamber, ruling of July 31, 2026, GEMA v. Suno · GEMA press statements, July 31, 2026 · Music Week, "GEMA wins court ruling on breach of copyright by AI music firm Suno."

Here is what makes Alex a credible voice on this topic: he worked at Napster and Liquid Audio during the music industry's first licensing reckoning, and has spent the two decades since helping companies get the value question right before a court decides it for them.

For more, book a conversation →