Sony's 30,000-Song Move Got Blocked. The Claim That Actually Threatens Udio Is a Different One

What Happened

Sony Music asked a court to fold 30,442 of its recordings into its copyright case against the AI music generator Udio, and lost. On July 2, Judge Alvin K. Hellerstein of the Southern District of New York denied the request, ruling that adding more than 30,000 works near the close of document discovery would require substantial additional production. Sony had filed the motion on May 22. With it denied, the suit stays at 333 works. The fight goes back to June 2024, when the RIAA sued Suno and Udio on behalf of the major labels for what it called mass copyright infringement, training generative models on copyrighted recordings without permission. Since then the ground has shifted. Universal Music settled with Udio in October 2025 and Warner in November, each signing a licensing deal. Of the three majors, Sony is the only one still across the table from Udio in court.

What This Means for Founders

The number grabs attention, but the count is not the story. Blocking the 30,000-song expansion was a procedural call, not a loss on the merits. Sony’s real leverage sits elsewhere. Udio admitted it pulled audio from YouTube to train its models, and in April, Hellerstein declined to dismiss Sony’s claim that doing so circumvented YouTube’s technical protections under the DMCA. How the training data was obtained now matters more than how much of it there was. A fair-use fight is long and uncertain. A circumvention claim is sharper.

The larger shift is in posture. Universal and Warner stopped trying to win and started licensing, the same arc playing out in NYT v OpenAI, where the training-data question is drifting toward deals rather than verdicts. For anyone building a generative model on scraped data, audio or otherwise, that points two ways. First, licensing is now the clean path to content, and the labels have shown they will take the deal. Second, the provenance and consent behind your training set is becoming both your moat and your liability. A startup that ships licensed generative music, or tooling that clears rights for training data, is stepping into a gap the majors just validated by settling.

What You Can Do Now

If you train generative models, document where every slice of your data came from and how you got it. Relying on fair use and scraping around a technical barrier carry very different weight in a courtroom. If you work with rights holders, the opening is to design a licensing path to revenue now, not a lawsuit. Universal and Warner already picked that lane.