TL;DR
The AI race is moving toward data control as public web text approaches projected limits for frontier training, according to Epoch AI estimates. Copyright settlements, publisher licensing and private corpora are making data access a barrier that smaller labs may struggle to clear.
AI Dispatch’s latest Control Series installment identifies data control as the emerging AI chokepoint, arguing that public web text is nearing projected training limits while copyrighted, enterprise, expert and sovereign datasets are being priced, fenced and litigated, a shift that could decide which labs, companies and governments retain leverage in advanced AI development.
Epoch AI estimates that the public internet contains about 300 trillion tokens of high-quality text, and its projections put full use of public human text between 2026 and 2032, with a median around 2028. The report says frontier training sets are already approaching that ceiling, making repeated web crawls less valuable than fresh, verified and private data.
The legal picture has also changed. Anthropic agreed to a $1.5 billion settlement with authors over claims tied to pirated books, at about $3,000 per work across roughly 500,000 titles, and agreed to destroy the pirated files. The settlement resolves past piracy claims but does not settle future training rights or model-output disputes, while The New York Times’ case against OpenAI remains in discovery and publishers including News Corp have pursued licensing deals.
Labs are also using synthetic data and specialist datasets to fill the gap. TechCrunch reported Nvidia’s $320 million purchase of Gretel, a synthetic-data company, and Microsoft has used hundreds of billions of synthetic tokens in training. Research cited in the International AI Safety Report 2026 warns that synthetic data can compound errors in domains where answers are hard to verify, increasing demand for human-made, expert-checked material.
Data: The One Thing You Can’t Rent
The free part of “all human knowledge” is running out. As compute and models commoditize, the corpus you can’t replicate becomes the moat — so data is being fenced, priced, and, in places, treated as a national asset.
Data was supposed to be the abundant input. It’s the scarce one. It’s also the chokepoint you can actually own — so guard your proprietary data, and don’t hand it to a provider who can become your competitor (the lesson everyone fled Scale to learn). Nations: license it like Ukraine — keep the model, keep the leverage.
Proprietary Corpora Become The Moat
The report’s central finding is that compute can be rented, but unique data cannot be copied on demand. If chips and model architectures become easier to access, the defensible advantage shifts toward datasets that competitors cannot buy in bulk: customer records, clinical notes, legal work product, robotics logs, autonomous-driving footage, intelligence feeds and expert evaluations.
That shift changes the balance of power. Large AI companies can pay for licenses and settlements that smaller developers may not be able to absorb, while enterprises and governments that control valuable data gain bargaining power. It also raises risk for customers that share proprietary records with a vendor that could later build competing products.
Synthetic Data Generation: A Beginner’s Guide
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Scraping Gave Way To Licensing
The first stage of large-model training depended heavily on public web scraping and low-cost labeling. As systems moved into reasoning, professional tasks and real-world decision support, the data need changed from broad text volume to expert judgments and hard-to-reproduce records.
The report points to several examples of the new data economy. Reported H100 rental rates have fallen 60% to 75% from their peak, reducing compute’s scarcity premium. At the same time, Meta’s reported $14.3 billion deal for a 49% stake in Scale AI prompted customer concerns about giving training pipelines to a company tied to a major AI competitor, according to CNBC and TechCrunch coverage cited by the report.
Data is also becoming a sovereign asset. The report cites Ukraine’s Ministry of Defence as treating combat data and related model rights as strategic leverage, including conditions meant to keep model control with Ukraine when outside partners use Ukrainian battlefield data.
“Data was supposed to be the abundant input. It’s the scarce one.”
— AI Dispatch
expert-verified training data datasets
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Licensing Costs Still Unsettled
Several points are still unresolved. Epoch AI’s public-text exhaustion date is a projection, not a measured deadline, and better algorithms could stretch available data. It is also unclear how far courts will let AI developers use legally acquired works for training, how output claims will be handled, and whether licensing prices will fall or harden into a long-term barrier.
Synthetic data remains an open question. It can lower costs in verifiable settings, but the safety report cited by AI Dispatch warns that repeated use of machine-made text can degrade model quality when mistakes are hard to detect. The durability of enterprise, expert and battlefield data as moats will depend on provenance, consent, contract terms and auditability.
Training Data for Machine Learning: Human Supervision from Annotation to Data Science
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Courts And Customers Set Terms
The next phase will be shaped by court rulings, licensing contracts and data-sharing agreements. Watch the Anthropic settlement administration, continued discovery in The New York Times v. OpenAI, new publisher licensing deals, and enterprise AI contracts that define whether customer data can train vendor models.
For companies and governments, the immediate step is contract discipline: know what data is being shared, whether it can be retained, whether it can train future models, and who controls any model produced from it.
private dataset storage solutions
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is the actual development in this report?
The report says AI’s bottleneck is shifting from compute to data. Public web text is projected to reach practical limits for frontier training, while valuable data is moving into paid licenses, private contracts and sovereign controls.
Is the public internet already used up for AI training?
No. Epoch AI projects full use of public human text between 2026 and 2032, with a median around 2028. That estimate means the easy supply is narrowing; it does not mean every useful public document has already been consumed.
Did the Anthropic settlement decide future AI copyright rules?
No. The settlement resolves past piracy claims involving downloaded books and requires destruction of the pirated files. It does not resolve future training, licensed datasets or model-output disputes.
Why does proprietary data matter more now?
Generic web text can be copied by many labs, but private enterprise records, expert review data and real-world operational logs are harder to replicate. Control of those datasets can affect model quality and bargaining power.
What should data owners watch in AI deals?
They should review whether the provider can retain prompts, files or outputs; whether shared data can train general models; and who owns improvements or models created from proprietary material.
Source: Thorsten Meyer AI
