Is AI data-labelling and RL-environment revenue as durable and high-quality as software ARR?

Andrew Feldman · published Oct 6, 2025 · hedged

Provisioning value-added tagged and evaluated data is clearly important now, but whether that market is durable in three years is genuinely unknowable

Machines may come to do the tagging and evaluation every bit as well as people, which would undercut the market

55:48 Automation of tagging could undercut durability of the data labeling market

Harry Stebbings · published Dec 1, 2025

The data labelling sector's reported revenue figures are dubious and its businesses are dangerously concentrated — several players claim $100M+ ARR but each derives over half its revenue from two customers, and 'revenue' is often used ambiguously versus GMV.

Seven players he knows of report over $100M ARR while all have more than 50% of revenue from two customers, and their revenue definitions are unclear.

0:18 Reported arr is inflated and concentration makes it fragile

Jonathan Siddharth · published Dec 1, 2025

Data-labelling/RL-environment revenue should be judged from first principles rather than as SaaS ARR — it is reoccurring project-based revenue that persists only if you keep performing

Lab projects for coding, multimodality, tool use or RL environments start and end, but demand is abundant as long as you consistently do a good job

33:27 Project based but reoccurring so judge it on its own terms

Erik Bernhardsson · published Nov 2, 2026 · hedged

Data/environment companies are fantastic cash businesses today but probably bad long-run venture businesses — founders should take the money rather than get on the venture treadmill

They will make a lot of money today, but he doubts they make a lot in the very long run; the cash flow today is what matters

45:53 Great cash businesses today but questionable long term venture quality take the money

Harry Stebbings · published Nov 2, 2026

Data providers can only expand margin and build a durable long-term business by going full lifecycle — deployment, implementation, evaluation, monitoring and enterprise handholding — rather than just selling acquired data

Pure data acquisition and resale offers no durable margin; being a real implementation partner, like providing FTEs, does

46:38 Full lifecycle deployment and enterprise handholding not data resale builds durable margin