Inside the training-data business: $2k-5k per problem, batches priced in millions
teortaxesTex · x · 2026-11-17
An insider breakdown of the training-data vendor ("数商") business:
- Per-problem pricing varies wildly: domestic vendors quote 2k-5k CNY, while vendors following overseas rates charge up to $2k per problem
- Pricier problems require larger minimum orders; batch deals are priced in the millions, and most labs partner with data vendors for bulk procurement
- Quality is hit-or-miss: some datasets are genuinely "human-expert annotated", others are actually "agent-expert annotated"
- A vendor's moat is knowhow/resources/scenarios the buyer can't cover internally — but if a problem costs x to produce, how many times that is a fair price?
The author predicts buyers' push for cost-effectiveness will squeeze out the huge margins, bringing demand and fundraising back to rationality.
More from Venture
- Study: 100% of Profound's "AEO" recommendations map to existing SEO advice — gaganghotra_ · 2026-11-17
- Nvidia's problem: OpenAI is "90 plus percent Nvidia" and burning the most cash — ThereWas · 2026-11-17
- Microsoft overtakes Nvidia to become world's second most valuable company, behind Apple — Polymarket · 2026-11-17
- RoboStrategy CEO says humanoid labor market could end up bigger than the AI market — Rewkang · 2026-11-17
- Anthropic's annualized revenue hits ~$81B ahead of IPO, with $895B compute commitments — lennysan · 2026-11-17
- World 2030: AI-simulated US economy run shows seat-license SaaS debt defaults in every branch — sebkrier · 2026-11-17