Inside the $100B market for AI training data
In July 2026, Menlo Ventures investor Deedy Das published an "AI Training Data & RL Environments" map. More than 50 companies now sell training data, RL environments and evaluation services to model labs, with about $8.5 billion in combined annual revenue and total valuations near $100 billion. Over 75% of that revenue goes to fewer than 10 companies; the remaining 40-plus companies share less than a quarter.
01Three leaders, three different bets
Scale AI was the market's acknowledged leader. On June 13, 2025, Meta invested $14.3 billion for 49% of the company, pushing its valuation to $29 billion, and founder Wang moved to Meta's superintelligence lab. Neutrality died with the deal. Google, Scale's largest customer, had planned to pay it about $200 million in 2025 and cut ties once the deal closed. OpenAI had already started reducing cooperation before the deal. In July, right after taking the money, Scale AI laid off 14% of its staff.
Surge AI took the opposite path: no VC money, no publicity push, heads down on expert alignment data for hard code, math and logic. Its reported annual revenue exceeds $1 billion, and on Das's map it carries the highest valuation — over $20 billion from investors.
Mercor is the third leader. It uses AI interviews to match top PhDs, lawyers and programmers to write data for model labs. On February 20, 2025, it raised a $100 million Series B led by Felicis at a $2 billion valuation. By October 2025, its Series C was $350 million at $10 billion — $8 billion added in eight months. Its CEO then declared that "the era of data labeling companies is over."
02Data labeling is no longer a category
The interesting part of the map is not how many companies it lists. It is that "data labeling" has disappeared from the table. The table now reads: RL Training Environment, Expert Data, Agent Evaluation, Physical AI Data, World Model Data.
The plain version: models moved from student level to PhD level. Part-time annotators drawing boxes and scoring text can no longer teach them. Labs need real experts to set tasks (Expert Data) and real sandboxes where agents can run code and fail (RL Environments). The species changed, and valuations followed. Per The Information, the environment company Matrices reached a valuation of about $750 million in April 2026 — on this map, that still counts only as a "challenger."
03Why the money moved from models to data
The industry's underlying logic changed: AI has shifted from lacking models to lacking data, and the shift is not reversing. The past two years went to scrambling for GPUs; compute alone no longer decides anything. Synthetic data bootstrapped from itself stalls, and public pretraining corpora have been scraped clean. The coming contest is over high-quality data, real environments and effective feedback.
04What buyers should weigh
First, neutrality. Scale AI's lesson is on the record: Google's $200 million budget was cut as soon as Meta's stake closed. Few labs want to hand their most core RL environment and data needs to a supplier backed by a direct rival.
Second, stop buying data with the old outsourcing-labeling logic. Simple Q&A pairs written by non-experts are empty calories for current frontier models. What you buy today is data that makes your model fail: hard, real environments that expose the model's limits. If a vendor's data lets your model score full marks, the money is wasted — the model already knew it all. The valuable data is the expert-level data and complex environments that beat top models.
05DataMining: what hard data looks like
This is why leading Chinese labs are quietly reworking their data supply chains. DataMining was built by a PhD team from Peking University's School of Computer Science (Software Institute) and has been deeply serving ByteDance, Alibaba and Tencent. We do not do headcount labeling; we build hard Deep SWE-class data and environments.
The hardness is measurable. Run GPT-5.6-sol on our Deep SWE-class data and the accuracy (ACC) comes out 15 percentage points lower than on data from comparable vendors, with interaction trajectories about 20 rounds longer. Lower accuracy and longer trajectories mean higher task realism and complexity: the model fails harder here, and RL can extract more capability from those failures.
06Compute is only the ticket
While the crowd watches launch events and argues over 0.1-point leaderboard gaps, money has already moved down Das's list into a market taking in $8.5 billion a year at near-$100B valuations. In the new phase where data, not models, is the constraint, compute is only the entry ticket. Whoever holds the real environments and the top experts that can push models to their limits holds the actual path to the next generation of models.
Sources: