Jamendo's Copyright Suit Against NVIDIA and Suno

On June 22, 2026, Jamendo, a music licensing platform owned by Belgium-based Winamp Group, sued NVIDIA in the U.S. District Court for the Northern District of California, accusing the chipmaker of training its Fugatto and Audio Flamingo audio models on Jamendo's music without authorization. A week later, on June 29, 2026, Jamendo filed a parallel suit against Suno in the U.S. District Court for the District of Massachusetts.

Two Suits, One Dataset

The MTG-Jamendo Dataset is a research collection of roughly 55,000 full audio tracks, around 3,500 artists, and 195 genre, instrument, mood, and theme tags. Jamendo built it in 2019 with the Music Technology Group at Universitat Pompeu Fabra in Barcelona, and released it for non-commercial research use only.

Jamendo's theory is that NVIDIA's own published research papers identify MTG-Jamendo as training material for Fugatto and Audio Flamingo, and that feeding a commercial product with a non-commercial dataset crosses a line. The complaint brings four sets of claims:

• direct copyright infringement
• breach of contract (the Creative Commons BY-NC license and Jamendo's terms of use)
• unjust enrichment
• unfair competition

In the Suno complaint, Jamendo says it sent Suno a €16 million ($18 million) invoice in September 2025, broken down at €289 per track, and that Suno confirmed receipt but never paid. Jamendo is now seeking at least €17.8 million ($20 million) in actual damages and the defendants' profits, plus injunctive relief. Most AI music litigation has been about the underlying musical works themselves, rather than a compilation of metadata.

Instead, Jamendo is claiming infringement based on two distinct copyright assets. First, the musical works and sound recordings for which Jamendo holds exclusive licenses. Second, and more importantly, the MTG-Jamendo Dataset itself, which Jamendo describes as containing its "original selection, coordination, and arrangement of musical works and sound recordings and associated metadata, including distinctive and granular tagging of each track".

That second claim is the novel part. It is not really about the songs. It is about the curated metadata wrapped around them.

The Dual-Track Infringement Theory

Why run two infringement theories in parallel? Because they shore each other up. The contract claims are arguably the strongest plank. The dataset was released under a non-commercial Creative Commons license and an express research-use notice, so using it to ship a commercial product looks like a clean breach of contract regardless of how the copyright questions shake out.

Copyright in a Compilation of Metadata

A relevant and frequently asked question is: Can you copyright a dataset of songs and tags?

The short answer is: sometimes, but the copyright only grants you “thin” protection. U.S. copyright law protects the original "selection, coordination, and arrangement" of a compilation, not the underlying facts or data themselves. So Jamendo is not claiming ownership of the genre labels or the audio files as raw data. It is claiming ownership of the creative way it chose, organized, and tagged them.

That matters because most AI music cases have trained their fire on the works, not the wrapper. Jamendo is essentially arguing that the wrapper, the granular tagging framework that makes the dataset commercially useful for AI training, is itself a protected work.

Thin Protection for Factual Compilations

Suno and NVIDIA will likely respond that the Dataset is not protectable by copyright at all, or at most enjoys only "thin" protection. They have a heavyweight Supreme Court precedent to point to.

The "sweat of the brow" Trap

In Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340 (1991), the Court held that "the copyright in a factual compilation is thin" because copyright can never extend to the facts themselves, only to the creative choices made in assembling them. An alphabetical white-pages phone directory, no matter how much sweat went into gathering the numbers, lacked the "minimal degree of creativity" the Constitution requires. Feist demolished the old "sweat of the brow" theory: effort alone earns nothing.

The lesson for Jamendo is blunt. The months of curation, the 120 tracks its employees reportedly review each day, the granular tagging, all of that labor buys it nothing unless it reflects genuinely original choices rather than mechanical collection.

Can Jamendo Survive a Motion to Dismiss?

To get past a dispositive motion to dismiss, Jamendo will need to establish that its compilation is sufficiently original to warrant copyright protection in its own right. That requires proving the selection of tracks, the structure of the metadata, the tagging system, and the overall organization reflects protectable creativity, not just a functional or factual arrangement.

This is plausible, but not automatic. Jamendo did register the MTG-Jamendo Dataset as a compilation with the U.S. Copyright Office on June 17, 2026, just days before filing, which gives it a registration to stand on. But registration is only a ticket in the door. It does not settle whether the AI defendants copied enough of the expressive elements of the MTG-Jamendo Dataset to constitute copyright infringement, and the late timing could limit the remedies available, since statutory damages and attorneys' fees generally require timely registration.

What It Means for AI Developers and Rights-Holders

If you build generative audio, the takeaway is uncomfortable. Using a "research dataset" in a commercial product can create exposure even while the broader fair-use question remains unsettled. A non-commercial license is not a loophole. It is a contract, and breaching it is a separate cause of action that does not care whether your training is transformative.

For rights-holders, the playbook is widening. Instead of suing only over copied songs, plaintiffs can target the curated data layer, the metadata, the tagging framework, the organization that made a dataset valuable to train on in the first place. That turns curation into an asset worth defending.

Jamendo's cases against NVIDIA and Suno are less about who owns the music and more about who owns the map of the music. The contract claims give Jamendo a sturdy path that sidesteps the fair-use brawl. The copyright claims, though, run straight into Feist and its thin-protection doctrine, and there the outcome is far from certain.

The real message is simple. In the AI era, the dataset you trained on may matter as much as the songs inside it, and the license you ignored may hurt you more than the copyright you allegedly violated.

If you have an Artificial intelligence intellectual property concern, contact our team today.