The AI Industry Is Rebuilding Its Moat Out of Paper, Silicon, and Power

In a warehouse somewhere, industrial hydraulic cutters are feeding millions of printed books into a scanner that destroys them as it reads. The pages are pulped; what survives is the text, digitized and fed to a model. This is Anthropic's Project Panama, and it is the clearest sign yet of where the AI industry's advantage now lives ^[1]. The most digital industry on earth is mechanically destroying paper to build its models. The reason is a three-way squeeze on the digital data training used to run on. First, the sources are being walled off. 241 news sites across nine countries now block the Internet Archive's Wayback Machine to stop AI scraping ^[2]. Cloudflare will block AI training crawlers by default on ad-supported pages starting in September, turning what was free scraped data into a paid commodity ^[3]. And MediaNews Group is suing OpenAI for $10 billion with an argument that names the asymmetry plainly ^[4].

OpenAI pays for its chips. It pays for its computers. It pays its programmers. But it steals the raw material for its GAI products — valuable well-written content — from hard-working journalists without payment and without permission. — Steven Lieberman

Second, the digital data that remains is contaminated. AI-generated ebooks on Amazon have increased 19-fold, displacing human-authored text in the very channels models were trained on ^[5]. DeepMind's own researchers have identified the problem directly.

So you essentially lose all those tokens inside of that document, even if it was a small single line that contained some personally identifying information. — Minqi Jiang

The EU's watermark rules make AI text easier to distinguish from human writing, which quietly raises the value of unmarked, pre-AI content ^[6]. Third, the model layer itself is commoditizing: DeepSeek's V4-Pro costs 12 to 19 times less than OpenAI or Anthropic for equivalent tasks ^[7], and OpenAI has cut Luna prices 80% ^[8]. No lab has said it in so many words, but the logic is hard to miss: when the algorithm is cheap, the advantage has to come from somewhere else. That somewhere else is physical. AI companies are now buying bulk secondhand books from libraries and estate sales, targeting pre-2022 texts specifically to avoid chatbot-contaminated data ^[9]. The book database ISBNdb brokers these purchases anonymously, and its pitch is blunt.

the world's best AI training data is sitting on a shelf. — ISBNdb

The same turn shows up in infrastructure. Anthropic's IPO filing commits more than $100 billion to AWS over the next decade ^[10], and it is in talks to acquire Decart AI, a chip-efficiency company, for $6 billion ^[11]. Zhipu AI is building a 1-gigawatt data center on domestic Chinese silicon ^[12]. The financial verdict is already in: AI capital spending has grown from $300 billion to $650 billion in a year, and cloud revenue backlog has reached $2.4 trillion ^[13]. Markets are pricing the industry's future as a function of physical deployment, not algorithmic advantage. The counter-arguments don't break the pattern; they confirm it. Perplexity's CEO argues on-device AI will replace centralized data centers ^[14] — but that addresses where inference runs, not where training data comes from. DeepMind's Generative Data Refinement exists to clean contaminated digital data rather than discard it ^[15] — a rescue method that only makes sense because the digital pipeline has degraded so far. The objections don't weaken the turn; they show how far it has already gone.