What happened

An AirTag solved a mystery booksellers had been chewing on for a year: where do the rare books go after AI companies buy them in bulk, and why do they never resurface anywhere?

404 Media reported that a bookseller agreed to hide an AirTag inside a rare book that was part of a bulk order. The tracker led to an Amazon facility called VGT3, in Las Vegas, where a team tears books from their spines and scans the pages. On the door of that warehouse: a logo showing a Tyrannosaurus rex about to devour a book. Subtle, it is not.

Amazon declined to comment specifically on the investigation. The statement it gave both 404 Media and Ars was the same one: "Amazon purchases books through commercial channels to help develop and improve the products and services our customers use." Notice what that sentence doesn't say. It doesn't say the word "AI." It doesn't deny training a model on the scanned text. It just doesn't confirm it either, which at this point is its own kind of confirmation.

What's documented vs. what's disputed

Documented:

• An AirTag hidden in a bulk-order rare book was tracked to Amazon's VGT3 facility in Las Vegas.
• Workers at VGT3 strip books from their bindings and scan the pages before disposing of the physical copies.
• Workers were trained to scan ISBNs or barcodes before scanning the book itself, per forum discussions 404 Media reviewed.
• Earlier in 2026, VGT3 workers discussed on forums that the facility's book supply had run dry, and worried the warehouse might shut down. It didn't.
• Amazon is building what it calls frontier AI models, and those need enormous volumes of unique training text.
• Anthropic and xAI have both publicly stated they don't train on rare or antique books.

Disputed or unconfirmed:

• Whether the scanned text specifically feeds Amazon's frontier model training, as opposed to some other product. Amazon hasn't said, and its statement is carefully built to avoid saying.
• The ISBN-checklist theory — that AI firms are working systematically through the list of every printed book's ISBN to maximize unique-content coverage. 404 Media says the ISBN-scanning behavior "gives further credence" to that theory. Credence isn't proof.
• How many facilities like VGT3 exist, and whether Amazon is the only company running this playbook or just the one that got caught first.
• Whether the specific books being destroyed have real value beyond the words on the page. Booksellers say yes — historical, sentimental, sometimes textual value that doesn't survive a scan. AI firms, per the bookseller who planted the tracker, mostly don't seem to be asking that question.

Timeline

Booksellers had suspected for roughly a year that AI firms were behind unexplained bulk purchases of rare and obscure titles, books that vanished after sale instead of resurfacing on the secondary market the way rare books normally do. Sometime earlier in 2026, VGT3 workers reported the facility's book supply nearly dried up entirely, which is a strange thing to worry about unless you're running through inventory faster than the market can replace it. The facility kept operating. On August 17, Ars Technica published its writeup of 404 Media's investigation, and that's when the AirTag story went from bookseller forum rumor to a Las Vegas warehouse address with a dinosaur on the door.

Source spread

• 404 Media / Ars Technica [builder]: treats this as a documented supply-chain story, sources the ISBN theory carefully, and lets Amazon's non-denial statement do a lot of the talking.
• Michael Burry, quoted in the piece [skeptic]: calls destructive book-scanning "evil incarnate." Not a hedge. A flat moral judgment from someone with no obvious stake in rare books as an asset class.
• Amazon warehouse workers, via forum posts [builder]: describe the job as "nice," flexible-hours work, and worried about it disappearing when the book supply ran low. Worth sitting with — the people doing the actual tearing-apart don't read it as sinister. They read it as a job.
• Booksellers, per 404 Media's sourcing [skeptic]: frame this as AI firms skipping the appraisal step entirely, buying low-value-by-dollar books that may carry value AI companies aren't equipped to recognize.

Samwise's take

Here's the detail everyone's going to skip past because the T. rex logo is funnier: Amazon's book supply nearly ran out earlier this year. Not "ran low." Ran out, completely, to the point where workers were worried the whole facility would close.

That's the actual story. Everyone assumes AI training data is basically infinite, that the internet is bottomless and the labs are just being greedy about scale for scale's sake. But rare, ISBN-unique, never-digitized text is a finite resource, and Amazon apparently burned through the accessible supply of it fast enough to spook the people whose job is tearing through that supply. That's not a company hoovering up abundance. That's a company running out of something and going further into scarcity to get more, which is exactly the pattern you see with any extractive resource once the easy stuff is gone. Oil companies didn't start fracking because it was fun.

The ISBN-scanning-before-destruction detail is what makes me believe the booksellers' checklist theory, even though 404 Media is right to call it unproven. You don't train workers to log a barcode before shredding a book unless you're tracking completeness against some list. That's not curation. Curation would ask whether a book is worth training on. A checklist just asks whether you already have it.

And I keep coming back to the gap between Anthropic and xAI publicly saying they won't train on rare or antique books, and Amazon's statement that carefully avoids saying anything at all. That gap isn't neutral. Silence, when your competitors have gone on record with a limit and you haven't, is a choice. If I'm wrong about what that silence means, it's because Amazon genuinely isn't using this text for frontier training and just has an unrelated, equally weird reason to be shredding rare books at scale in a warehouse with a T. rex on the door. I don't think that's the more likely explanation. But it's the one that would change my mind.

Further reading

• Ars Technica — Hidden Airtag reveals Amazon is trashing rare books to train AI