Amazon has found a new source of training data for its large language models, and it involves literally gutting rare books. According to reporting from 404 Media, the company is buying up out-of-print and hard-to-find volumes, slicing off their spines, and scanning the pages at a facility in Las Vegas called VGT3 — a site that rather pointedly uses a dinosaur clutching a book as its logo. Amazon confirmed the purchases, telling the outlet it “purchases books through commercial channels to improve the products and services customers use.”
Why go through the trouble of destroying physical books? Because the open internet is largely tapped out. LLMs have already ingested most readily available text, and some companies — Anthropic, notably — have faced legal trouble over pirated books. Rare and out-of-print titles represent a largely untapped reservoir of high-quality prose. And there’s another advantage: anything published before 2022 was almost certainly written by a human, which matters as AI-generated content floods the web.
That distinction isn’t academic. When models train on their own output, researchers warn of “model collapse” — a degradation in quality as errors compound across generations. Pre-2022 human-authored books are essentially a clean dataset. Amazon isn’t just buying content; it’s buying provenance.
The image of a company destroying rare books to train chatbots will strike many as grim. But the economics are straightforward: unique text is now a strategic asset, and physical books are one of the last places to find it at scale. The question is whether other companies will follow Amazon’s lead — and what disappears from the historical record in the process.
