Data Marketplaces & Intermediaries

Four vendors now control three-quarters of the training-data market — concentration is the real story in AI's supply chain

Source: Pebblous · Oct 1, 2026

A market map circulating this cycle made the shape of the training-data business hard to miss: four vendors — Scale, Surge, Mercor, and Handshake — reportedly account for more than three-quarters of a roughly $8.5 billion market, with combined private valuations across the sector approaching $100 billion and a hyperscaler holding a commanding stake in the leader. The data-supply layer of AI, in other words, already looks like an oligopoly.

The strategic read is that scale begets scale. The biggest labelers won early contracts, used the revenue to hire and tool up, and now sit between the frontier labs and the human workforce that produces the data. Newer entrants are mostly not trying to out-label the incumbents head-on; they are moving to a different surface — interactive reinforcement-learning environments — where the leaders' head start matters less.

But concentration this steep is also a tell about where the leverage actually is. When a handful of vendors own the commodity supply, the thing that holds value is data they do not already have: differentiated, domain-specific, defensibly-sourced material that cannot simply be relabeled from the same crowdsourced pool everyone else is drawing from. A larger share of an undifferentiated market is still an undifferentiated market.

That is the opening. The incumbents are extraordinary at volume and throughput; they are not automatically the owners of consented, provenance-tracked data in the domains that matter most — high-stakes conversations, real human interaction, the modalities that a transcript flattens. As the supply of generic labeled data commoditizes under four big sellers, the premium migrates to data that is scarce because it was collected carefully, with the rights to use it. Concentration at the commodity layer and scarcity at the defensible layer are the same story told from two ends — and the second end is the one still open to build on.

Key Points

• An industry tally puts four vendors — Scale, Surge, Mercor, Handshake — at more than 75% of a roughly $8.5B training-data market
• Combined private-market valuations across the sector run near $100B, with a hyperscaler holding a large stake in the leader
• Challengers are clustering around interactive RL 'environments' rather than static datasets
• Concentration raises the value of differentiated, defensibly-sourced data that the incumbents don't already own