The Physical Supply Chain of AI

Artificial intelligence may feel immaterial. We talk about models, prompts, tokens, and cloud infrastructure as if AI lives entirely in software. But behind the sleek interfaces and futuristic branding is a very old-world system: trucks, warehouses, secondhand bookstores, scanners, cutters, and crates of printed books moving through a hidden supply chain.

That physical supply chain matters because the appetite for high-quality training data has pushed AI companies far beyond the open web. As publicly accessible text becomes scarcer, more competitive, and increasingly polluted with AI-generated content, companies are turning to books—especially long-form, human-edited, pre-2022 books—as premium training material. In some cases, the process is not just about digitizing books. It is about buying, cutting apart, scanning, and then discarding them.

This development has sparked a fierce debate about copyright, ethics, preservation, and the future of knowledge itself.

Why books became valuable AI fuel

For years, large language models were trained on a mixture of internet pages, forums, articles, code, and digitized archives. That worked well at first, but the open web has limitations. It is noisy, repetitive, and increasingly saturated with machine-generated text. As a result, books have become especially attractive.

Books are different from random web pages for several reasons:

They are usually written with more structure and editorial care.
They often cover topics in greater depth than web content.
They contain coherent, long-form arguments and narrative flow.
They are more likely to be authored by humans rather than AI systems.
They can provide better examples of style, syntax, and domain knowledge.

From a machine-learning perspective, that makes books highly valuable. For publishers, authors, and librarians, it raises a difficult question: should the most durable form of human knowledge be turned into training material for systems that may compete with the very people who wrote it?

The rise of destructive scanning

One of the most controversial practices linked to AI training is destructive scanning. In this process, books are purchased, their bindings are removed, each page is scanned, and the physical copies are destroyed or discarded.

This is not the same as careful archival digitization. Traditional preservation work usually aims to protect both the content and the physical object. Destructive scanning is far more utilitarian: the goal is not to preserve the book as an artifact, but to extract its text as efficiently as possible.

The method is blunt, but it is also fast and scalable. By using industrial cutting machines and bulk scanning workflows, organizations can process thousands of books at a time. That makes the physical book supply chain suddenly relevant to AI. Rare book dealers, secondhand bookstores, wholesalers, and metadata brokers can become upstream suppliers in a training-data pipeline.

For many observers, the unsettling part is not merely that books are being scanned. It is that books may be treated as disposable raw material.

Project Panama and the hidden procurement machine

Court documents and investigative reporting have brought unusual attention to internal AI data acquisition projects. One such initiative, described in litigation as “Project Panama,” reportedly involved an effort to destructively scan large numbers of books for model training. The language used in these disclosures made the scale sound almost industrial: not just collecting books, but building a pipeline capable of processing them at volume.

According to reporting, a former Google Books executive was brought in to lead acquisition efforts, underscoring how the expertise developed in large-scale digitization projects has become valuable in the AI era. But unlike library preservation programs, the new use case is driven by model performance, not public access.

The key point is that AI companies do not only need data; they need dependable, legally usable, human-written data at scale. That demand creates pressure to secure books through intermediaries, bulk purchase arrangements, and specialized sourcing channels.

A market built for AI appetite

Once demand becomes large enough, markets adapt. In this case, a secondary market has emerged around the needs of AI training. Some metadata and book-sourcing companies have reportedly offered bulk acquisition services aimed at large language model developers. Secondhand dealers, especially those handling academic, technical, and out-of-print books, may suddenly find themselves approached for thousands of titles at a time.

That kind of demand changes the economics of the book trade.

What AI buyers are looking for

AI companies and their procurement partners tend to value books that are:

Recent enough to avoid overused public-domain material.
Human-authored and professionally edited.
Dense with factual, instructional, or narrative language.
Not easily available elsewhere on the web.
Available in bulk through secondary markets.

Academic titles, specialist nonfiction, and niche reference works can be especially attractive because they cover subject matter that is hard to find in clean, organized digital form.

Why this concerns booksellers

For sellers, a bulk purchase request can look like a windfall. But for the wider ecosystem, the implications are murkier.

Rare or low-print-run titles may disappear from circulation.
Out-of-print works may be permanently lost if destroyed after scanning.
Libraries and archives may lose physical copies that still have research value.
The resale market may be distorted by sudden, large-scale acquisition demand.

In other words, AI’s hunger for training data can change the fate of books that were never written with machine learning in mind.

Copyright, fair use, and legal uncertainty

The legal picture is complicated and still evolving. In the United States, courts have begun to distinguish between different kinds of AI training use. In some cases, scanning legally purchased books has been argued to be transformative and therefore protected under fair use. In other cases, lawsuits have focused on allegedly unauthorized use of pirated books or shadow-library material.

That distinction matters.

If a company buys a book lawfully and scans it for internal model training, courts may treat that differently from downloading millions of copyrighted books from unauthorized sources. But even when a practice is technically legal, it may still be controversial from an ethical or cultural standpoint.

The core legal questions include:

Does training on copyrighted books count as transformative use?
Does the act of copying for model training substitute for the original market?
Are authors entitled to compensation when their work helps train commercial AI systems?
Should AI companies disclose exactly what they trained on?

These questions will shape the future of AI development as much as model architecture or chip supply.

The preservation dilemma

There is also a deeper cultural issue here: preservation versus extraction.

Digitization has long been a force for access. Libraries scan books so that more people can read them, study them, and search across them. But destructive scanning changes the moral equation because the physical object is sacrificed for a private technical benefit.

That raises several concerns:

Loss of material culture

Books are not just containers for text. They are artifacts. Editions, bindings, marginalia, paper quality, and printing history all have value to historians, collectors, and researchers. When a book is destroyed after scanning, that material history is lost.

Risk to endangered texts

Some books survive in very limited quantities. If a rare title is acquired in bulk and processed for training, there is a nonzero risk that the physical copies could vanish from circulation entirely.

Reduced public accountability

When procurement happens through intermediaries and anonymized buyers, it becomes difficult for the public to know what is being collected, where it came from, and whether it will be preserved.

The result is a strange tension: AI companies may be building the future of language technology by quietly dismantling pieces of the printed past.

What this means for authors and publishers

For writers, publishers, and rights holders, the implications are serious.

First, the use of books in AI training may influence future licensing debates. If large language models learn from copyrighted texts, authors may push for compensation, opt-out systems, or mandatory licensing frameworks.

Second, publishers may need to think more strategically about how their catalogs are distributed, digitized, and protected. Backlist titles, academic works, and niche nonfiction may have unexpected value in the AI economy.

Third, authors may need to consider how their work exists across formats. A book that lives only as a printed object can be harder to monitor than one distributed through controlled digital platforms with clear metadata and rights management.

For self-published authors especially, this may become part of the long-term business conversation. The same title that once seemed commercially modest could become relevant as training-data value increases.

The bigger story: AI is not disembodied

The most important lesson in all of this is that AI is not floating above the physical world. It depends on it.

AI models are trained on:

Electricity.
Chips and rare minerals.
Data centers and cooling systems.
Network infrastructure.
Human labor.
Printed books, scanned pages, and archived text.

That means the AI boom has a supply chain just like any other industrial system. The difference is that its raw materials are increasingly cultural rather than purely mineral or mechanical.

And once knowledge becomes supply chain input, every stage of acquisition becomes politically and economically important.

The future of book sourcing for AI

The demand for high-quality training data is unlikely to disappear soon. If anything, it may intensify as companies compete to build more capable models, reduce hallucinations, and improve performance on long-context and specialized tasks.

That could lead to several outcomes:

More formal licensing agreements between publishers and AI companies.
Stronger copyright enforcement and more litigation.
Growth in data brokers specializing in text acquisition.
New preservation standards for books bought for scanning.
Public backlash against destructive or opaque sourcing practices.

For now, the most likely scenario is continued tension. AI companies need books. Book owners want value, preservation, or both. Courts are still defining the boundaries. And the public is only beginning to understand how much of the AI revolution depends on physical objects that may be quietly disappearing.

Conclusion

The story of AI training data is often told as a software story, but it is really also a logistics story. Behind every polished chatbot may be a chain of old books, secondhand dealers, scanning machines, warehouse shipments, and legal disputes.

That is what makes the physical supply chain of AI so important. It reveals that the future of machine intelligence is being built not only in code, but in the material world—page by page, shipment by shipment, and sometimes book by book.

If we want to understand where AI is going, we need to pay attention not just to algorithms, but to the invisible infrastructure that feeds them.