Hook
Anthropic spent millions to buy and destroy physical books. The books were scanned, shredded, and discarded. The goal: clean training data for large language models. In crypto, we burn tokens to create scarcity. AI now burns books to create data scarcity. The parallel is precise — and terrifying.
Volatility is the tax on unverified assumptions. Here, the assumption is that destroying a physical artifact for a digital copy is legal and ethical. The market has not yet priced the risk of cultural backlash, regulatory reversal, or the irreversible loss of rare texts. This is an infrastructure problem disguised as a legal loophole.
Context
The process is called destructive scanning. ISBNdb, a data broker, offers a service: purchase books from publishers and wholesalers, cut the spines, scan every page, then destroy the physical copies. The 2025 U.S. court ruling on Google Books established a precedent: converting a lawfully owned physical book into a non-distributable digital copy is fair use, provided the original is destroyed to maintain a one-to-one count. This ruling gave AI companies a legal pathway to acquire high-quality, human-generated text that has never been exposed to AI-generated noise or data poisoning.
Anthropic, the AI safety company behind Claude, hired a former Google Books project lead and spent several million dollars to acquire millions of books. ISBNdb markets the service with promises of legal confidentiality and verifiable destruction. The rationale: pre-2022 physical books are pristine — no machine-generated filler, no adversarial contamination.
But this is not a technical innovation. It is a legal arbitrage. The same way DeFi exploits regulatory gaps to offer unregistered securities, this model exploits the physical-to-digital conversion loophole. And just as with DeFi, the risk is that the loophole closes — retroactively.
Core: The Infrastructure of Cultural Mining
From a macro perspective, this is a new form of mining. Instead of hashing algorithms, the work is done by industrial book scanners and shredders. The energy cost is not electricity but cultural heritage. The output: training data tokens with unique provenance.
Let me quantify this. A single high-resolution scan of a 300-page book produces approximately 150 MB of raw data. For 2 million books, that’s 300 petabytes of storage. Cloud storage at $0.02 per GB per month means $6 million per month just to hold the corpus. This is before OCR, metadata tagging, deduplication, and quality filtering. The true cost per token — the cost to acquire, digitize, store, and process — could be 10x the raw acquisition cost.
But the real insight is structural. This model creates a physical-world bottleneck on data supply. Unlike scraping the web, which is replicable at near-zero marginal cost, destroying books eliminates the original. First movers like Anthropic gain a permanent advantage. Competitors cannot re-acquire the same books because they no longer exist. This is akin to a 51% attack on the data supply chain: once a network controls a majority of hash power, others are locked out.
In crypto, we understand this as the “size premium” — the largest miners have lower average costs. Here, the largest book burners secure the highest quality data and destroy it, ensuring no one else can use it. This is not decentralization. This is the creation of a new monopoly on pure human text.
I witnessed a similar dynamic during the 2020 DeFi Summer. I reverse-engineered AMM pricing algorithms and found that liquidity depth was systematically fragmented by yield farmers. The inefficiency was 15% — a tax on unverified assumptions about capital efficiency. Here, the inefficiency is the destruction of cultural artifacts for a purely speculative benefit: a few points of improvement in model perplexity.
The Regulatory Precedent
The Tornado Cash sanctions taught us a brutal lesson: writing code can be a crime. The same reasoning applies here. If a future court decides that destroying a book to create a digital copy is not fair use — or that the digitization itself infringes on author moral rights — the entire corpus becomes illegal. The digital copies must be deleted. The physical books cannot be restored. The investment is lost. The model retrained.
This is the hidden leverage in the narrative. Investors are not pricing the legal tail risk. They see the cost as a one-time capex. They ignore the possibility of forced write-offs and retroactive liability.
Contrarian: Why This Model Will Fail
The contrarian view: the book-destruction model is not the future. It is a dead-end reflex born from fear — the fear of data contamination, the fear of copyright litigation, the fear of being left behind.
Code executes logic; humans execute fear. The logic says that destroying a physical copy to create a digital one is efficient. The fear says that if you don't do it, your competitor will. But both ignore the third path: decentralized provenance.
Blockchain-based storage networks — Arweave, IPFS, Filecoin — offer a superior alternative. Instead of destroying the original, anchor the digital copy with a cryptographic commitment. Prove that the training data came from a verified physical source without destroying that source. Use zero-knowledge proofs to show that the OCR output matches the original edition without exposing the full text. This preserves the cultural artifact while enabling legal compliance.
The market is wrong to accept the destructive model as optimal. The real infrastructure play is not in buying shredders. It is in building verifiable data markets where training data is permanently anchored on-chain. The company that solves data provenance without cultural vandalism will capture the next cycle.
Takeaway
The book burners are building a data monopoly on a foundation of ashes. The crypto-native response is not to emulate their destruction but to outperform it with trustless preservation. The next bull run will reward projects that prove data purity without destroying the past.
History doesn't repeat, but it rhymes. The 2017 ICOs promised code-as-law; most failed. The 2022 Terra collapse proved algorithm stability is an illusion. Now, the 2025 court ruling has given AI a license to burn. The question is not whether it is legal — but whether we will look back and call it civilized.