For a decade, AI training data was something you scraped. GitHub's public repos, Stack Overflow, documentation, the open web. That era is ending for a simple reason: the frontier labs have already used most of it.
The public well is running dry
Researchers at Epoch AI estimated that the stock of high-quality public human-generated text will be effectively exhausted by frontier training runs before 2030 — and high-quality public code is a small slice of that stock. Meanwhile the legal ground under scraping shifted: publishers sued, licensing deals replaced silent crawling, and every major lab started caring about provenance — being able to prove where training data came from and that they had the right to use it.
At the same time, model quality became less about raw pretraining volume and more about post-training: fine-tuning and reinforcement learning on carefully chosen, verified data. Post-training doesn't need a trillion mediocre tokens. It needs the right ones.
What's special about private production code
Public GitHub skews toward tutorials, forks, abandoned side projects, and single-commit dumps. What it underrepresents is exactly what working software engineers do all day: maintain large living systems, fix production incidents, review each other's changes, migrate schemas, integrate vendors, and make tradeoffs under constraints. Private production repositories contain that entire texture:
- Multi-year commit histories showing how real changes land, get reviewed, and get reverted
- Tests and CI that define what "correct" means for the system
- Bug fixes paired with the bugs they fixed — natural training pairs
- Domain logic that simply does not exist in any public corpus
There's a second, newer use. Labs now train agents with reinforcement learning inside working software — repositories turned into environments where an agent attempts real tasks (fix this bug, add this feature, make the tests pass) and gets scored. A runnable private repo with a good test suite is raw material for exactly that, which is why the environment companies — a sector that attracted hundreds of millions in venture funding in 2025–2026 — are themselves buyers of code.
Why buy instead of generate?
Labs do generate synthetic code data at scale, but synthetic data inherits the blind spots of the model that generated it. Research on benchmark contamination and eval gaming — for example the SWE-bench Illusion line of work, which found models substantially overperforming on memorized public repos versus unseen ones — pushed labs toward fresh, never-scraped code as both training material and honest evaluation ground. Your private repo has one property no synthetic generator and no public scrape can offer: the model has genuinely never seen it.
The market that formed around this
By mid-2026 the demand side is crowded: frontier labs buying directly, AI-data vendors buying repositories outright, and platforms listing per-codebase pricing. The supply side — people who own good code and know it's sellable — is still catching up. That information gap is exactly why we publish pricing: here's what codebases actually go for.