Insights · 6 min read

What Makes Data Valuable for AI Training? Six Real Drivers

Volume is the least interesting thing about your data. Here is what AI labs actually look for.

Bearish AI · October 21, 2026

Every company with years of code, chat and tickets eventually asks the same question: is any of this worth money to an AI lab? The honest answer is sometimes a lot and sometimes almost nothing. The difference rarely comes down to size. It comes down to a handful of properties that buyers check, often in this order.

Start with what is scarce

Epoch AI, a research group that tracks AI inputs, estimates the effective stock of quality-adjusted, human-written public text at about 300 trillion tokens, with a wide 90 percent range of 100 trillion to 1,000 trillion. It expects that stock to be fully used somewhere between 2026 and 2032, depending on how models are scaled.

The same analysis deliberately leaves out private material such as instant messages. It says non-public data seems unlikely to be used at scale because of legal issues and because it sits in fragmented systems controlled by different parties. Read that as a description of the opportunity and the obstacle at once. Public text is finite. Private work records are a different category, but they only count if someone can lawfully and cleanly deliver them.

Quality can beat quantity

Two well-known research results show how far careful selection goes. The LIMA paper fine-tuned a 65-billion-parameter model on only 1,000 carefully curated prompts and responses, and concluded that almost all of a model’s knowledge is learned in pretraining, so a small amount of well-chosen instruction data can be enough to shape its behavior. A separate coding study trained a small model on roughly 6 billion tokens of web data filtered for “textbook quality” plus about 1 billion tokens of generated textbooks and exercises, and reported 50.6 percent pass@1 on a standard coding test.

Neither result means your archive replaces pretraining. They mean the buyer’s question is rarely “how much do you have?” and more often “how much of it teaches something specific?”

The six drivers

Why human origin matters more over time

A widely cited study of recursive training found that when models learn from model-generated data, rare content at the edges of the original distribution disappears, an effect the authors call model collapse. The authors suggest that data reflecting genuine human interaction with systems becomes more valuable as generated text spreads. A record of people actually debating, deciding and correcting each other is exactly that kind of data.

Why checkable outcomes set the price

Agent training works by trial and grading. Epoch’s interviews with people in the reinforcement learning market put the most important quality criterion as robustness against reward hacking: a high score must mean the task was really solved. They also describe difficulty calibration. A task every model fails teaches nothing, and so does one every model passes. Buyers look for tasks that a model solves at least occasionally, and they may drop tasks that are already solved around 70 percent of the time.

That is why graded tasks have a visible price: most are reported at $200 to $2,000 each, with about $20,000 possible but rare. Exclusive rights were described by two founders as costing roughly four to five times the non-exclusive price. A raw archive has no such unit price, which is why it is usually valued as a lot, and why preparing it into checkable tasks changes the outcome.

A hypothetical comparison

Picture two small software companies, each with five years of records. The first exported its repository once, as a single folder of files, and its chat is a handful of channels about lunch plans and releases. The second kept its full commit history, its tickets reference commit numbers, its chat threads link to the tickets they discuss, and a test suite ran on every change. The first has volume. The second has tasks: each fixed bug is a problem, an attempt, a review and a check that can be run again. Even if the first archive is twice as large, a buyer building agent training would take the second.

Preparation can move an archive from the first group toward the second, but only if the raw material survived. Links that were never recorded cannot be invented later.

Price signals worth knowing

Reporting on the resale of startup archives puts typical payouts at $10,000 to $100,000 per company, with some reports of more. At the other end, Epoch’s interviewees describe replicas of websites built as practice environments at about $20,000 each, and far more complex products, such as a faithful replica of a team-chat product, at around $300,000. The gap between these shows how much the form of the data matters. A lot of raw records is one product. A finished, graded environment is another.

What to do with this

• Keep the links. Do not strip ticket numbers, review threads or commit messages. Context is what turns a file into a lesson. See /insights/linked-data-code-tickets-chat.
• Keep the history. A full record of changes is worth more than the final files. See /insights/why-version-history-worth-more-than-code-snapshot.
• Settle rights early. Know who owns what before a buyer asks. See /insights/data-ownership-rights-before-selling-data-to-ai-companies.
• Be honest about weak spots. Duplicates, generated code and thin archives lower value. See /insights/what-makes-data-worth-less-to-ai-labs.

If you want a first sense of where your own data sits on these six drivers, the free estimate at /estimate asks about the sources you hold, how long they cover, how many people made them, and how clear your rights are. It is deliberately conservative.