Executive Summary
Behind the frontier model race sits a less visible market: the data platforms that supply what models learn. In June 2025, Meta invested $14.3B for a 49% stake in Scale AI and hired away its founder — and the market convulsed. OpenAI, Google, and Microsoft halted or scaled back work with Scale, and Surge AI and Mercor absorbed the vacuum, flipping the leaderboard within a year. The deeper change is in the value chain itself: from low-wage labeling to PhD-level expert data, evaluations, and reinforcement learning environments.
1. The Meta-Scale Deal: One Investment That Reshaped a Market
Scale AI defined industrial data labeling with a global contractor workforce. With 2024 revenue of $870M and projections of $2B for 2025, Meta invested $14.3B at a ~$29B valuation and brought founder Alexandr Wang in to co-lead Meta Superintelligence Labs.
The problem: Scale's biggest customers were Meta's competitors. OpenAI, Google, and Microsoft cut ties — no lab could accept its training data pipeline being visible to a rival — and Scale cut 14% of staff. Revenue is recovering under new leadership (projected $1B+ in 2026), but the lesson stands: the neutrality of a data pipeline is itself an asset.
2. The Reshaped Landscape
| Player | Position | Scale (as reported) | Differentiation |
|---|---|---|---|
| Scale AI | Labeling pioneer; Meta 49% | $870M (2024) → $1B+ recovery projected (2026) | Scale and operations; neutrality lost |
| Surge AI | High-quality RLHF and expert data | Bootstrapped, ~110 staff; reported to have passed Scale in revenue | Small, elite, frontier-lab focused |
| Mercor | Expert talent marketplace | $2B annualized revenue (June 2026); reported $20B valuation talks | PhD/lawyer/doctor networks; 90% of revenue from foundation model companies |
| Arena & eval platforms | Model evaluation and rankings | ~$100M ARR (est. June 2026) | "The Nielsen of LLMs" — evaluation data as training signal |
Sources: Forbes, Business Insider, Sacra, HeroHunt (2026). Private-company figures are reported estimates with variance.
The essence is not "who won" but which data got expensive. The center of gravity moved from drawing boxes on images to PhD-level experts evaluating outputs, writing preference data, red-teaming, and designing RL environments — from dollars per hour to tens or hundreds of dollars per hour.
3. The Value Chain Shift: Labeling → Expert Data → Environments
Pre-training data and labeling (2016–2023)
Web data curation and mass low-wage annotation — the market Scale defined. Growth slowed as web data was exhausted and synthetic data rose.
Alignment and expert data (2023–)
RLHF, preference data, and domain-expert evaluation became the bottleneck. Math, coding, legal, and medical experts serve as the model's graders and teachers — the territory Surge and Mercor captured.
RL environments and eval infrastructure (2025–)
Agent-era training data is not static text but environments to act in and get feedback from. RL environment construction, simulators, and agent task evals are the new battleground.
Strategic Implications
For enterprises — companies that can convert domain expertise into data assets can become suppliers in this value chain. Before relying on external platforms, systematize your internal experts' judgment data: evaluation criteria, preferences, edge-case handling.
For investors — multiples here are priced on the "essential frontier-lab infrastructure" frame, not labeling outsourcing. Watch customer concentration, labs' in-housing risk, and the speed of transition to Gen 3 (environments and evals).
Need a talk, article, or advisory on this topic?
I deliver executive keynotes, contribute industry reports, and advise on AI strategy.
Get in Touch