The claims
Austin Beaulier creates 3D models and photogrammetry scans and publishes them on Sketchfab, Thingiverse and Polycam. In the companion Meta case he alleges more than 400 of them carry a Creative Commons attribution license. His theory is that the creator name, title and license terms published with each model are copyright management information under 17 U.S.C. § 1202(c), and that AI developers removed that information when they turned the models into training data. He brings only DMCA claims, which carry statutory damages of $2,500 to $25,000 per violation and do not require proof of infringement of the models themselves.
According to the complaint as described in NVIDIA’s motion, his models were included in Microsoft’s TRELLIS-500K dataset, derived from the academic Objaverse-XL index. NVIDIA is said to have used Objaverse-XL and TRELLIS-500K as a map to download the underlying files, then run them through preprocessing (format conversion, mesh normalization, rendering, voxelization) that separated the models from their CMI. The only specific NVIDIA product identified is its AI Blueprints, which incorporate Microsoft’s TRELLIS model.
NVIDIA’s motion
NVIDIA moved to dismiss on 22 June 2026 (ECF 20) on six grounds: no removal was pleaded, only failure to preserve; no intent, because any loss of CMI was a byproduct of uniform preprocessing; no reason to know the removal would enable or conceal infringement, because no infringing output is alleged; no identical copies, because the pleaded preprocessing transforms the works; no distribution, only internal use; and no actual knowledge that CMI had been removed. It also pointed out that the complaint attributes the omission of license information in TRELLIS-500K to Microsoft, not NVIDIA.
The order
In a docket order entered 23 October 2026, Judge Lee granted the motion with leave to amend. Rather than write a separate opinion, she found Judge Breyer’s August opinion in Beaulier v. Meta Platforms persuasive and adopted its reasoning. Beaulier may file an amended complaint within 21 days, with a redline against the original. NVIDIA must meet and confer at least seven days before moving to dismiss again. The court terminated NVIDIA’s motion to stay as moot and stayed discovery, explaining that a plaintiff whose claims have been dismissed for implausibility is not entitled to discovery on them. The 28 October hearing was vacated and the case management conference moved to 27 January 2027.
The reasoning adopted from Beaulier v. Meta
Judge Breyer’s order (No. 3:26-cv-02632-CRB, ECF 32, 26 Aug 2026) is the opinion the later Beaulier rulings build on. It took each element of § 1202(b)(1) in turn and found only one missing.
| Element | Result | Reasoning |
|---|---|---|
| CMI existed | ✓ Pleaded | Creator, title and Creative Commons terms qualify. A Sketchfab “NoAI” tag counts as a term or condition for use under § 1202(c)(6). |
| CMI was removed | ✓ Pleaded | A plaintiff need not plead the mechanism. CMI present at the source and absent from training inputs is enough at this stage. |
| Removal was intentional | ✕ Not pleaded | Under Logan v. Meta, a uniform process that sheds CMI as a side effect is not intentional removal. The complaint itself called the separation an inherent consequence of training. Contrast NYT v. Microsoft and Concord v. Anthropic, where plaintiffs alleged deliberate extraction choices that dropped CMI. |
| Knew or had reason to know it would enable infringement | ✓ Pleaded | The removed license terms marked the limits of authorized use, so removing them plausibly enabled use beyond those limits. |
| Distribution (§ 1202(b)(3)) | ✕ Not pleaded | Distribution requires a sale or other transfer of ownership to the public. Moving copies within one company for training is internal use. |
Judge Labson Freeman’s 2 October order in Beaulier v. Roblox reached a similar result, though unlike Judge Breyer she also found removal inadequately pleaded. She noted that the parties had not yet briefed the Ninth Circuit’s 16 September decision in Doe v. GitHub. That decision affirmed dismissal of § 1202(b) claims over Copilot output. It rejected a freestanding identicality requirement but held that generating new material without attribution is not removing or altering CMI, and it declined to reach, as forfeited, the theory that training on stripped copies violates § 1202(b).
The four Beaulier suits
| Defendant | Court and number | Judge | Status |
|---|---|---|---|
| Meta Platforms | N.D. Cal. 3:26-cv-02632-CRB | Breyer | Dismissed with leave to amend, 26 Aug 2026 |
| Roblox | N.D. Cal. 5:26-cv-02642-BLF | Labson Freeman | Dismissed with leave to amend, 2 Oct 2026 |
| NVIDIA | N.D. Cal. 5:26-cv-02647-EKL | Lee | Dismissed with leave to amend, 23 Oct 2026 |
| Microsoft | W.D. Wash. 2:26-cv-01031-JNW | Whitehead | Motion to dismiss filed 8 Jun 2026; no ruling located as of 30 Oct 2026 |
Evidence notes
IPLYR’s reading of what an amended complaint, and any later proof, would need. These are observations about method, not predictions about the outcome.
- Intent is now a question about the pipeline.
- The line courts are drawing runs between a process that loses everything that is not training signal and a choice that singles out CMI, such as picking an extraction tool because it drops attribution when an alternative would keep it. Facts of that kind can sometimes be found in public material: dataset cards, released preprocessing code, and the fields kept or dropped in each dataset release.
- Find where in the chain the CMI disappeared.
- As Judge Breyer described the companion complaint against Meta, Objaverse-XL keeps references back to each source file, repository and creator account. The complaint against NVIDIA alleges that Microsoft’s TRELLIS-500K release omitted per-file license information. A field-by-field comparison of the source pages, the Objaverse-XL index and the TRELLIS-500K records would show which release dropped what. That attribution is exactly what NVIDIA says the complaint lacks.
- Model tests answer a different question.
- Extraction and memorization tests show whether a model learned or can reproduce a work. They say little about why metadata was discarded before training. A § 1202(b) claim needs dataset and pipeline evidence; an infringement claim needs model and output evidence. They should be built and reported separately.
Documents
| Date | ECF | Document | Filed by |
|---|---|---|---|
| 22 Jun 2026 | 20 | NVIDIA’s Motion to Dismiss | NVIDIA |
| 6 Jul 2026 | 26 | Plaintiff’s Opposition to Motion to Dismiss | Beaulier |
| 26 Aug 2026 | 32 | Beaulier v. Meta, No. 3:26-cv-02632-CRB: Order Granting Motion to Dismiss (adopted by Judge Lee) | Court |
| 16 Sep 2026 | No. 24-7700 | Doe v. GitHub, Inc., No. 24-7700 (9th Cir.): Opinion | Ninth Circuit |
| 2 Oct 2026 | 47 | Beaulier v. Roblox, No. 5:26-cv-02642-BLF: Order granting motion to dismiss with leave to amend | Court |
| 23 Oct 2026 | Text order | Docket order granting motion to dismiss with leave to amend and staying discovery | Court |
NVIDIA’s motion and the opposition link to RECAP copies on CourtListener. The Meta and Roblox orders link to copies hosted by ChatGPT Is Eating the World. Judge Lee’s order appears as docket text; its terms are summarized above.