In March 2024, a group of authors sued nVidia for copying, storing and using their registered works, which were part of a data set that nVidia used to train its NeMo Megatron large language model. The authors want an injunction to bar nVidia from further use and to remove all instances of the works from its models. The case is still a work-in-progress.
The authors claim that NVIDIA obtained copyrighted material from so-called “shadow libraries.” Internal records reportedly suggest NVIDIA felt “competitive pressure” to use these sources to keep pace with AI rivals.
In late April 2026, the presiding Judge ordered nVidia to produce discovery information regarding the ‘shadow library’ data sets used to train the NeMo models “within a month.”
Labyrinthine sourcing
According to the Saveri law firm, which filed the lawsuit on behalf of the authors, the model cards for Megatron, as hosted on the Hugging Face Web site, stated that ‘The Pile‘ was its source dataset. In turn, The Pile includes the Books3 dataset, which further in turn derives from a collection of e-books called Bibliotik; which hosts unlicensed copyright materials.
Other datasets named by the Plaintiffs include a dataset called SlimPajama – which contains derivatives of the Books3 dataset – and another called Anna’s Archive, which the Complaint claims to host “mllions” of pirated books.
An amended version of the authors’ complaint contended that “NVIDIA “got the green light” to use the datasets. “NVIDIA did not hesitate in using pirated books from these illicit sources of copyrighted material, regardless of the “risk” or the harm to authors like the Plaintiffs,” said the Amended Complaint.
Impeding the case
The authors contend that while nVidia has incrementally produced some information about the sources that trained its models, and has provided reasoning as to its timing; the slow pace of production stands to put the timing to take depositions at risk.
“NVIDIA has produced some model checkpoints and training data and made certain data available for inspection on a remote terminal. But NVIDIA has neither fully identified the datasets and data sources containing pirated books that it intends to produce, nor the full scope of its production.”
nVidia claimed that the delay was the fault of the Plaintiffs, which painted a different picture: that the isolated inspection environment that NVIDIA set up—and the file formats provided—made it entirely infeasible to review and search the training data. Several weeks later, NVIDIA again produced data in the same inspection environment, with the same technical limitations and constraints.
Why it matters
The Judge felt that nVidia’s slow pace has been unwarranted, and has imposed a deadline for full production of discovery content. Only then, can it be impartiallyl determined whether or not nVidia is attempting to hide a case of piracy.
nVidia is fighting several copyright infringement lawsuits as of this writing, including a class action filed by a 3D artist which targeted nVidia, Roblox and Meta for using 3D data for AI training.
Earlier in April, YouTube reportedly took down an nVidia trailer about its DLSS5 technology after a takedown request by an Italian TV channel.
Further reading
Joint Status Report. Abdi Nazemian, et al., Plaintiffs, vs. NVIDIA Corporation, Defendant. Document 291. Case 4:24-cv-01454-JST. US District Court for the Northern District of California, Oakland Division.
Amended Complaint. Abdi Nazemian, et al., Plaintiffs, vs. NVIDIA Corporation, Defendant. Document 226. Case 4:24-cv-01454-JST. US District Court for the Northern District of California, Oakland Division.
Complaint. Abdi Nazemian, et al., Plaintiffs, vs. NVIDIA Corporation, Defendant. Document 1. Case 4:24-cv-01454-JST. US District Court for the Northern District of California, San Francisco Division.
Case: nVidia large language model litigation. Article. Accessed April 21, 2026. Saveri Law Firm LLP
