Something is buying Japan’s used books by the ton, and the books are not going to collectors. Since about August 2026, online used bookstores across Japan have been filling bulk orders from a rotating set of accounts, and the titles they are clearing out look nothing like normal demand. NTV, Japan’s national broadcaster, broke the story in an exclusive. The trail it traced runs from a shop counter to a single logistics center in Okayama prefecture, across the Pacific, and toward the growing business of turning books into AI training data. The books are the input. The model is the output. The book does not come back.
It is tempting to think of AI as pure software, but it has a physical supply chain. It has data centers, water, and power, and the fights over those are already public. It also has a paper one, and that supply chain is quieter, because most of it happens on the used book market where almost nobody is watching. This post follows the boxes.
The short version:
- Online used bookstores across Japan have been getting bulk orders from many accounts since around August 2026. A single dealer in the Kanto region sold more than 1,000 books to one destination.
- Some shops report selling 100 books a day and days that run at five times normal sales.
- The genre mix is the tell. Not novels or manga, the normally fast moving stock, but philosophy, history, political history, medicine, law, and Edo period life and culture.
- Every domestic order ships to the same logistics center in Okayama prefecture. The operator declined to comment.
- NTV, analyzing trade data with the firm Sayari, found more than 50 tons of goods labeled JAPANESE BOOKS exported from a Japanese publishing distribution group company to the United States since 2025. At roughly 500 grams a book, that is about 100,000 books.
- The reporting points at a destructive scanning practice internally called Project Panama. Nippan denies any such deal. Treat the link as reported and denied, not proven.
- The outlet 404 Media has separately documented Amazon collecting, slicing, scanning, and discarding used books in a Las Vegas warehouse, Japanese books included.
It Started With Orders That Did Not Look Human
The first sign was volume. Around August 2026, dealers selling used books online in Japan began seeing orders that made no sense for a normal customer. The orders arrived in bulk, they came from many different accounts, and they pointed at a small number of destinations.
One dealer in the Kanto region sold more than 1,000 books to a single destination. That is not a reading habit. That is a pickup. The timing carried the same message. Orders arrived at even intervals, the kind of metronome rhythm that looks automated rather than human, and stock that would normally take weeks to move sold out in a day.
Across the shops NTV spoke to, the numbers repeated. Some reported moving 100 books a day. Some reported days at five times normal sales. None of them described a customer who cared what they were buying.
The Genre Mix Is the Tell
What is being bought is more revealing than how much of it is moving. A collector or a reader chases the books that everyone else wants. These buyers skip them.
The orders concentrate on philosophy, history, political history, medicine, law, and material about daily life and culture in the Edo period. Those are dense, factual, long form titles, and many of them are exactly the books that sit on a used bookstore’s shelves for years because the local market for them is thin.
The books that normally sell, novels and manga, are not the target. That is the tell. A dataset does not want what is popular. It wants what is broad, varied, and durable, the kind of prose that teaches a model how people reason about the world rather than how they entertain themselves. Buying the fast movers would be expensive and redundant. Buying the quiet shelf is cheap, and it is exactly the material a general model is missing.
Put simply: a collector buys the book they want. A dataset buys the shelf.
Follow the Boxes to Okayama
Where the boxes go tightens the picture. Every domestic order the reporting traced shipped to the same logistics center in Okayama prefecture. A single consolidation point is what you build when you are collecting from many sellers and shipping onward in bulk. When NTV asked the operator about it, they declined to comment.
There is a second kind of approach running alongside the orders. A store in western Japan received an email from an overseas company asking to buy tens of thousands of books. That is not a shop order at all. It is an industrial purchase order for raw material.
So the trail has two shapes. Domestic orders flow toward one consolidation point inside Japan. Overseas buyers also approach dealers directly, in volumes no single bookshop could absorb on its own.
Fifty Tons of Japanese Books
The clearest evidence is in trade records. NTV analyzed export data with Sayari and found more than 50 tons of goods labeled JAPANESE BOOKS exported from a Japanese publishing distribution group company to the United States since 2025.
Weight converts to count in a way that makes the scale real. A typical used paperback or hardcover runs around 500 grams. Fifty tons at that weight is roughly 100,000 books. That is not a shipment. That is a library, boxed and sent.
Be precise about what this proves. The export records show that the books left Japan in industrial volume for the US. They do not, on their own, name who unpacked them or what happened next. That part of the story comes from a different kind of document.
The Anthropic Thread, and Why It Is a Suspicion
The practice the reporting points at is destructive scanning. You buy the book, you cut or scan it to capture the text, and then you destroy the physical copy so it cannot be resold or compared. One internal name for this kind of collection has surfaced in court: Project Panama.
Project Panama appeared in the authors’ lawsuit against Anthropic, the case that produced the company’s 1.5 billion dollar copyright settlement. Trade press has separately reported a book buying contract dated November 21, 2024 involving Nippan, formally Nippon Shuppan Hanbai, a major Japanese publishing distribution company.
Here is where the story has to be handled carefully. Nippan denies the deal. The group says it never sold anything knowing it was for AI training. The connection between the Japanese books leaving the country and a specific AI lab is reported and denied. It is not proven, and it should not be repeated as if it were.
The human side of that same pipeline is its own story. The industry hires people to label and correct training data, then fired contractors for using AI to do the very job it had hired them for. That episode, reported by 404 Media, is a reminder that the layer under AI is made of people, not just paper.
Two further facts sharpen the stakes without settling them. Anthropic’s settlement does not cover Japanese authors, so the writers whose work may sit inside these scanned books are outside the only large payout in the case so far. And the scale in the records, tens of tons and six figure book counts, is the scale of a company building training data, not a company filling a warehouse.
Amazon Is Running a Version of This Too
This is not one lab’s private method. 404 Media has reported that Amazon collects used books, slices and scans them, and discards the physical copies in a Las Vegas warehouse, with Japanese books among the material. The details differ from the Japan case, but the shape is the same. Buy the secondhand copy cheaply, convert it to text, throw the object away.
When two or more large buyers do the same thing, it stops looking like a quirk and starts looking like a market. The used book trade is being treated as a cheap source of raw text, and the supply is finite.
In Japan, This Is Mostly Legal
None of this requires anyone to break a Japanese law. Under Japan’s copyright framework, scanning books for AI training is permitted unless it can be shown to unfairly harm the interests of rights holders. That is the key sentence in the whole story. The default is that you may scan. The burden falls on the person who says the scanning hurt them.
Compare that with the United States, where authors sued over exactly this kind of collection and won a large settlement. The same behavior that produced a billion dollar payout in one legal system can be close to routine in another. That gap is a large part of why the books are being bought here.
The Japanese rule is not a loophole nobody noticed. It was written to allow text and data analysis. The live argument is whether building a commercial model from millions of scanned books counts as the analysis the law meant, or as the unfair harm the law excludes. That argument has not been settled.
What the Shop Owners Are Left With
The dealers are of two minds, and both minds are reasonable. Sales are up. Books that would have sat for years are moving. For a small online shop, an order for a thousand books is a good month arriving all at once.
Then there is the other worry, and it is not really about money. A used book market keeps books circulating. A copy sold once and resold ten times is still a copy that people can read. Destructive scanning ends that. Once a book is shredded, it is not just sold. It is removed, and the pool of physical copies in the world is permanently one smaller. For rare or out of print titles, especially the local and regional histories these buyers seem to favor, there may be no second copy coming.
That is the part that reads less like a sales boom and more like a slow emptying. The cultural record is being pulled out of circulation by the ton, and the shops doing the selling can see it happening and cannot quite say no.
The Takeaway
The important thing here is not one company’s name. It is the arrival of a buyer whose demand ignores taste. Every previous player in the used book market, collectors, readers, libraries, students, wanted particular books for particular reasons. This buyer wants the shelf. Dense, varied, unfashionable, and cheap is the ideal purchase, which is nearly the opposite of what makes a bestseller.
That is a strange and powerful kind of demand, and it lands on a market that was never designed to notice it. It also lands on ordinary institutions. The same pattern of automated systems quietly reshaping who gets what already runs through hiring screens that decide which resume a human ever sees, and through the way schools are working out what AI means in classwork. The books are simply the newest, and the most physical, version of the same shift.
Two things are worth watching. First, whether the link between the exported books and a specific AI company is ever confirmed on the record, rather than reported and denied. Second, whether Japan’s permissive default for training data holds, or whether the argument about unfair harm finally forces a change. Until one of those moves, the quiet part of AI’s supply chain will keep running, one pallet of used paperbacks at a time.
The human layer under AI is also a labor story
The jobs that build training data are done by people, and those people are the easiest part of the pipeline to cut. Our report on OpenAI’s fired contractors follows what happened when reviewers reached for AI to do the work they were being paid to do.
About The Author
