IP Law Daily, COPYRIGHT—N.D. Cal.: NVIDIA must face direct infringement claim over ‘shadow library’ downloads, (Jan 21, 2026)

By Thomas Long, J.D.

Authors granted leave to amend complaint to add direct infringement claims alleging illegal downloads and indirect claims over customers’ uses of NVIDIA’s AI training dataset.

Authors suing chipmaker NVIDIA Corp. over its artificial intelligence models have been allowed by the federal district court in Oakland to file an amended complaint adding allegations that NVIDIA illegally downloaded the authors’ copyrighted works from online “shadow libraries.” The court also approved amendments making allegations regarding newly discovered AI models allegedly trained on the illegally downloaded works, as well as new claims of secondary infringement based on NVIDIA’s alleged acts encouraging its customers to download and use its training dataset. In the court’s view, the authors did not unduly delay seeking amendment because they did not have knowledge of the relevant facts until they were turned up during discovery. It also concluded that NVIDIA would not be prejudiced by the amendments, given the early stage of the litigation. The direct infringement claim over illegal downloads—which was added to an amended complaint filed the day after the court gave the go-ahead to amend—is similar to claims filed in Bartz v. Anthropic PBC, No. 3:24-cv-05417-WHA (N.D. Cal.), which ultimately resulted in a class action settlement that included a $1.5 billion payout to class members (Nazemian v. NVIDIA Corp., No. 4:24-cv-01454-JST (N.D. Cal. Jan. 15, 2026)).

Infringement suit. Authors Abdi Nazemian, Brian Keene, and Stewart O’Nan filed a putative class action lawsuit against NVIDIA for using their copyrighted works in training its NeMo Megatron–GPT series of large language models. An LLM is AI software designed to generate naturalistic text outputs in response to user prompts.

Motion to amend. The authors moved to amend their complaint to include claims related to “shadow libraries” known as Anna’s Archive and SlimPajama, among others. Shadow libraries are illegal online databases that contain copyright materials that are made available for free. In support of their request to amend the complaint after the deadline for amending pleadings (but before the close of discovery), the authors asserted that NVIDIA’s discovery productions on September 26, 2025, revealed for the first time that NVIDIA had downloaded books from these libraries.

“Plaintiffs repeatedly and unsuccessfully sought discovery regarding NVIDIA’s use of shadow libraries, which contributed to their delay in seeking amendment,” the court said. “Once they had the relevant information, Plaintiffs moved to amend their complaint within a few weeks (and gave notice of their intent to amend even earlier).” The court concluded that the authors demonstrated good cause for not seeking amendment earlier, and it rejected NVIDIA’s contention that it would be unduly prejudiced by the amendment. Prejudice was lacking because the allegations were largely related to the authors’ original claims and leave to amend was sought months before the close of discovery and prior to the filing of dispositive motions.

The court also permitted the authors to amend their complaint to include LLM model families other than the “NeMo” and “Megatron” models identified in their original complaint. There was good cause to amend because newly produced evidence showed that the five new models proposed to be added to the complaint were also trained on NVIDIA’s allegedly infringing dataset.

Finally, the court determined that newly discovered information supported the authors’ request to add claims of vicarious and contributory infringement based on NVIDIA’s directing and instructing its customers to download and use NVIDIA’s training dataset compiled from the shadow libraries. In the court’s view, “discovery produced in October demonstrated that NVIDIA possessed scripts that facilitated copyright infringement by customers and that it provided these scripts to specific customers who then used them in infringing activities.” The court noted that this knowledge was different from knowledge resulting from earlier discovery productions—specifically “the knowledge that NVIDIA hosted a script that could be utilized to download and process the data … or that the script was publicly disseminated as early as March 2024.”

The authors filed their amended complaint on January 16. It alleges, “NVIDIA copied [the authors’] copyrighted works multiple times to train its language models, including from known pirated libraries (also known as ‘shadow libraries’). Those notorious shadow libraries include The Pile, Bibliotik, and Anna’s Archive.” The amended complaint also asserts that NVIDIA caused third parties to download and store the copyrighted works by encouraging, facilitating, and promoting its customers to download copies of the dataset.

Similar claims added to another lawsuit. Plaintiffs in another AI copyright lawsuit in the Northern District of California—this one concerning the Mosaic AI LLM developed by Databricks Inc.—also have been given the go-ahead to replead with additional claims for direct infringement based on allegations that Databricks unlawfully obtained copyrighted works in its training dataset from online shadow libraries. In the court’s view, the plaintiffs appeared to have been diligent in seeking amendment. It was preferable for the plaintiffs to wait to file their motion until they had uncovered a strong evidentiary basis for their direct infringement claim via discovery, rather than filing prematurely.

Moreover, the court had admonished the plaintiffs to have a stronger basis for the claim when it denied a previous attempt to amend the pleadings. Additionally, the court said the new allegations “do not come as a complete surprise to Defendants,” since the plaintiffs had already attempted to add them to their complaint, and their complaint against codefendant Mosaic ML, Inc., already contained a similar claim (In re Mosaic LLM Litigation, No. 3:24-cv-01451-CRB (N.D. Cal. Jan. 20, 2026)).

Implications for AI copyright litigation. These cases appear to signal that more litigants are trying to reproduce the plaintiffs’ success in Bartz v. Anthropic PBC. This could significantly raise the financial stakes in the widespread litigation over the use of copyrighted materials to train LLMs. The district court preliminarily approved the settlement in Bartz on October 17, 2025. The settlement has not yet received final approval because of procedural complications regarding the enumeration of claims and the procedures for class members to opt out.

The case is No. 4:24-cv-01454-JST.

Judge: Tigar, J.

Attorneys: Elisha Barron (Susman Godfrey LLP) for Abdi Nazemian. Andrew H. Schapiro (Quinn Emanuel Urquhart and Sullivan, LLP) for Nvidia Corp. Cadio R. Zirpoli (Joseph Saveri Law Firm, LLP) for Stewart Onan. Jedediah Wakefield (Fenwick & West LLP) for Databricks, Inc. and Mosaic ML, Inc.

Companies: Nvidia Corp.; Databricks, Inc.; Mosaic ML, Inc.

MainStory: TopStory AINews Copyright TechnologyInternet CaliforniaNews GCNNews