Inside the AI data grey market: token resellers make millions until a lab buys back its own data

vista8 · x · 2026-11-04

vista8 highlights a deep "elsewhere" interview with "Xiao Wu," a small-time player in the AI training-data grey market:

Diverse, ravenous demand: labs buy menus, drug inserts, handwritten medical records, game source code, console captures, discarded film footage; world-model teams want 3D game play data and modeling parameters — non-top-tier open-world studios are already selling it
Annotation underbelly: thousand-person video annotation factories pay under 300 RMB/day (roughly the cost of generating 20×15s clips with Seedance 2.0); some brokers get Surge crowdworkers to resell task data for double pay
Reseller flywheel: token relay stations subsidize customers with data-sales revenue, driving more usage and more data; an estimated second-tier internet company spends $20M/month on such stations, with reseller profits reaching tens of millions RMB this year
The turn: a major lab bought its own relay-station usage traces this summer, triggering a crackdown cascade — many dual-business resellers went quiet

Everything from individual sellers to top data brokers and the labs themselves operates in a grey zone.

More from AGI Musings