Data — Research Frontier

What's genuinely new and where the field is heading.

State of the rung as of 2026-09-24 (second compile, same day): 3 sources, 3 concepts, 4 entities. Two prices for training data — a court-set penalty and a negotiated licence — and one forecast of the public data stock. Synthetic data, retrieval corpora, contamination and data quality per token still have no source at all.

Active Frontiers

1. What training data costs

Status: Early stage Key papers: Bartz v. Anthropic — Authors Guild summary, Reddit S-1 Key players: Anthropic, The Authors Guild, Reddit

Two numbers now: ~$3,000 per registered book as a settlement for pirated copies (a ceiling), and $203.0M over two to three years for Reddit's January 2024 licences (at least $66.4M in 2024, from substantially one buyer). Different units — per work versus per year of access — and neither converts to tokens. See the price of training data.

Open problems:

• What does licensed data cost now? Reddit's contracts run out between January 2026 and January 2027; the renewal price is unrecorded, and its filings no longer give a licensing line (see the check below).
• How does a one-off data cost spread into a cost per served token?

2. Where the legal risk sits: acquisition or use

Status: Early stage Key papers: Bartz v. Anthropic — Authors Guild summary Key players: Anthropic

In this one case, training on lawfully bought books was fair use and the money attached to pirated acquisition. See acquisition versus training.

Open problems:

• Was the settlement finally approved, and on what date? Resolve from the docket.
• Does the split hold in other courts, and for scraped web data?

3. How much usable data is left

Status: Active — the forecast window has opened Key papers: Epoch AI, Can AI scaling continue through 2030? Key players: Epoch AI

Epoch's current window: the ~300T-token effective public text stock is fully used between 2026 and 2032, central 2028 if trained compute-optimally, 2027 at 5× overtraining. Its 2022 estimate had high-quality text gone before 2026; the move came from finding that filtered web text and repeated passes over the same data work better than assumed. The 2030 report ranks data behind power and chips as a limit on scaling. See the data stock.

Open problems:

• Has the window moved again since June 2024? Not checked in this KB.
• Does synthetic data work at frontier pretraining scale? Epoch excludes it from its estimate.

Recent Breakthroughs

Date	Breakthrough	By	Source
2025-09	First court-set per-unit price for training data: ~$3,000 per title	Bartz v. Anthropic settlement	Authors Guild
2024-06	Data-exhaustion window moved from "before 2026" to 2026–2032, central 2028	Epoch AI	Epoch AI
2024-02	First disclosed licence price for training data: $203.0M over 2–3 years	Reddit (S-1)	Reddit S-1

Predictions & Trends

• None offered. Two one-buyer prices and one forecaster are not a trend.

Checks made on 2026-09-24 (recorded, not ingested)

DATA-1 — Reddit's Q2 2026 10-Q — was checked against the filing and NOT ingested. The filing exists (period ended 30 June 2026, filed 31 July 2026, accession 0001713445-26-000100) and was read from EDGAR. The proposal said it shows "data licensing of about $43M, up 24%, about 5% of revenue, with OpenAI and Google named as the biggest buyers". What the filing actually shows:

• $43.28 million is "Other revenue", not licensing. Q2 2026 other revenue was $43,280 thousand against $34,842 thousand a year earlier (+24%), out of $804,905 thousand total revenue (5.4%). Reddit's revenue policy says it earns non-advertising revenue "from content licensing and products sold directly to users"; its S-1 defined the same line as licensing plus Premium subscriptions plus user-economy products. No licensing figure is broken out, so $43M cannot be cited as licensing revenue.
• OpenAI and Google are not named as licensees. They appear only as AI competitors. The filing says "substantially all of the contract value associated with our licensing revenue is derived from two of our partners", unnamed.
• One licensing-related figure is disclosed: $92.1 million of remaining performance obligations on contracts longer than a year, "primarily" long-term content licensing contracts — $62.0 million expected in the rest of 2026 and $30.1 million in 2027. That is contracted future revenue, not a quarterly licensing figure, and is recorded here only as the result of the check.

The S-1 (DATA-2), the named fallback, was ingested instead.

Knowledge Gaps

Each candidate still needs Connor's individual approval before ingest.

• A current licence price. The Reddit figure is 2024; a primary disclosure of a 2026 renewal, or of another seller's licence, would date the negotiated side properly.
• A cost per token for data. Nothing on the rung yet says how many tokens a licensed corpus or a book amounts to.
• Synthetic data, retrieval corpora, contamination, data quality per token — in scope, nothing proposed yet.