Snorkel CEO: raw pre-training data will commoditize, refined post-training data margins will keep growing

ajratner · x · 2026-10-19

Snorkel AI founder Alex Ratner argues that amid the hype around AI data brokering, a basic distinction is being missed:

(A) Unrefined data for pre-training: raw company records (potentially de-identified). This is low-margin brokerage that will rapidly commoditize, especially as pre-training data curation techniques mature and sharpen buyers' discriminative capabilities.
(B) Refined data for post-training/eval: well-formed, high-quality tasks, environments, and rubrics built on private data or human traces. This is an incremental evolution of the sophisticated frontier data/envs that AI data labs already build — and margins and differentiation here will only increase as the frontier advances and raises the quality bar.

More from Venture