Amazon’s VGT3 Warehouse Slices Up Rare Books to Train Its Nova AI Models

A 404 Media investigation, reported by TechCrunch on August 17, 2026, used a hidden AirTag to trace a rare-book shipment to an Amazon facility in Las Vegas where workers describe…

The operational fact here is simple: a rare book that used to circulate among collectors, libraries, and secondhand shops now has a one-way ticket to a scanner and a shredder. TechCrunch’s Amanda Silberling, relaying 404 Media’s AirTag-tracked investigation published August 17, 2026, describes an Amazon facility in Las Vegas, internally branded VGT3 and marked with a dinosaur clutching a book, where workers say incoming shipments are stripped of their bindings so pages can be scanned faster. Amazon told 404 Media it “purchases books through commercial channels” without addressing what happens to the books afterward — a gap Gadget Review’s reporting flags directly.

What makes destruction the point rather than a byproduct is the legal scaffolding underneath it. Judge William Alsup’s June 2025 order in the Anthropic authors’ litigation, cited by both the-decoder.com and Mashable, found that scanning legally purchased books qualifies as transformative fair use in part because the paper originals were destroyed rather than resold — no surviving copy, no competing market. That ruling didn’t just tolerate book destruction; it built a compliance incentive around it, which is presumably why Amazon’s Nova training pipeline and Anthropic’s now-notorious “Project Panama” both settled on the identical workflow of slicing spines and discarding pages.

Amazon and Anthropic didn’t converge on cutting up books by coincidence — the 2025 fair-use ruling made destroying the originals the legally safer move, not merely the cheaper one.

The secondhand-book market is already feeling the demand shock the source coverage only gestures at. The Guardian’s August 15, 2026 reporting found UK and Irish booksellers filling anonymous bulk orders with no thematic logic — one dealer sold “hundreds” of books for roughly £4,000, another has taken orders for 6,000 books since January, both to buyers paying full price with no discount haggling, an unusual posture for the trade. Layer in Mashable’s account of ISBNdb quietly marketing “the world’s best AI training data” sitting on shelves, under an NDA explicitly designed to avoid an “AI company destroys two million books” headline, and the picture is an industry that understands the optics problem well enough to contract around it rather than stop.

Watch for two things: whether Amazon confirms or denies that Las Vegas-scanned text feeds Nova training data specifically, and whether any regulator or publisher trade group tries to attach conditions — provenance disclosure, digitization-without-destruction mandates — to the fair-use precedent Alsup set. Right now no one outside the AI labs and their scanning contractors knows how many books have gone through this pipeline, which is exactly the visibility gap the NDAs are built to preserve.

Amazon is buying tons of rare books, cutting off their spines, and scanning them for AI training, according to 404 Media, which placed a tracking device in a rare book that ultimately arrived at an Amazon facility in Las Vegas. Amazon told 404 Media in a statement that it "purchases books through commercial channels to improve the products and services customers use."

TechCrunch AI

Read the full story at TechCrunch AI →

The Data Commenter, in your inbox

Data markets, alt data, and the AI training-data economy. No spam, unsubscribe anytime.

Discussion lives in the inline notes attached to article passages.