Amazon is buying up physical books and destroying them for AI

Amazon started out selling paperbacks out of a garage. Thirty years later, it is buying up rare physical books just to slice off their spines and feed them into AI scanners.

An investigation by 404 Media blew the lid off this secretive operation using an Apple AirTag. The outlet slipped the tracker inside a rare book they suspected would end up in tech hands, then followed its journey across the country to an Amazon facility in Las Vegas known as VGT3.

Workers at the warehouse confirmed that their routine involves receiving massive shipments of printed books, cutting off the bindings so the pages can be fed through high-speed scanners, and destroying the originals in the process. The unofficial mascot for this team is a dinosaur holding a book in its claws.

Amazon confirmed the purchases, stating it buys books through commercial marketplaces like Biblio to improve products, including its Nova family of foundation models.

The hunt for pre-AI text

Why go through the hassle of shipping heavy paper across state lines when the entire web is a scrape away?

The short answer is data pollution. LLMs have basically scraped the open web dry, and what is left on today's internet is increasingly diluted by AI-generated junk. When models train on AI slop, they risk what researchers call model collapse, where output quality rapidly degrades.

Physical books printed before 2022 are completely pristine. They contain rare, out-of-print human knowledge that never made it onto the internet, offering an uncorrupted well of natural language.

Unlike Anthropic, which ran into lawsuits over alleged pirated book datasets, Amazon decided to take the legal path of buying the physical copies first, even if it means destroying them right after.

Why it matters

There is deep irony in an online bookstore turning into a book guillotine, but this exposes just how desperate the frontier data race has become.

For builders, it signals that training high-end models is no longer just a compute and algorithmic problem. It is now a physical logistics problem. Clean, pre-2022 human text is a non-renewable resource, and deep-pocketed tech giants are willing to buy up every physical copy they can find to stay ahead.

We think Amazon's maneuver is legally clever but ethically fraught. By purchasing through commercial marketplaces like Biblio, Amazon sidesteps the direct copyright infringement that comes with web scraping or digital piracy.

However, physically destroying rare and out-of-print books is bound to draw intense backlash from archivists and authors. It also highlights a growing barrier to entry: if staying competitive requires buying and shredding mountains of physical archives, smaller AI startups simply cannot afford to play.