Mercor went from AI recruiting startup to one of the most talked-about names in training data in about two years. Its model is an expert marketplace: recruit and vet domain specialists at scale, then sell their work to AI labs as demonstrations, evaluations, and training data. Growth has been loud, and so has the news around it, from acquisitions to a reported valuation in the tens of billions. When a vendor gets that big that fast, buyers start comparison shopping, and "Mercor alternatives" becomes a real search.
This guide covers who competes with Mercor in each of its lanes, and when the better answer is not an expert network at all but the underlying data.
What Mercor is known for
Mercor's core asset is its expert network: tens of thousands of vetted specialists in fields like software engineering, medicine, law, and finance, matched to labs that need expert demonstrations, evaluations, and domain-specific training data. The recruiting DNA shows in how it competes, moving fast on sourcing and vetting people rather than building annotation tooling.
It has been expanding by acquisition. In early 2026 it bought Sepal AI, which built expert-graded benchmarks and evaluation data, and in July 2026 it acquired Deeptune, which builds reinforcement learning environments for training AI agents. That takes Mercor from expert data into evals and environments, the same full-stack direction Scale, Surge, and Centific are converging on. By press reports its annualized revenue has passed two billion dollars and it has been in talks to raise at a valuation around twenty billion.
Teams that go looking for alternatives usually have one of four reasons. Some want a second vendor for diversification, the lesson the market learned when Meta took its stake in Scale. Some were given pause by Mercor's rockier headlines, including a reported data breach in early 2026 and lawsuits from contract workers. Some need depth in a lane where a specialist beats a generalist. And some discover that what they actually need is training data that exists in the world, not experts producing new data to order.
Alternatives for expert data and demonstrations
If the job is expert-produced training data, the direct competitors are the other expert networks. Surge AI is the premium name for human feedback and expert labeling, selective and priced accordingly. Scale AI still runs some of the largest expert data programs, with the Meta caveat every buyer now weighs. Turing specializes in coding data and expert engineers. Micro1 and Handshake AI both run expert recruiting and vetting models similar to Mercor's, and both show up in the same vendor evaluations. If micro1 is on your shortlist, our guide to micro1 alternatives compares that field directly. Turing comes at the same market from a software engineering base; our guide to Turing alternatives covers how it stacks up.
We also compare that field in our guide to Handshake AI alternatives.
Toloka and Prolific come at it from managed crowds and vetted research participants respectively. Our guide to Prolific alternatives covers that side of the market.
For the human feedback side specifically, our guide to RLHF data providers maps that market in depth.
Alternatives for RL environments
The Deeptune acquisition puts Mercor in the RL environments business, where the competition is already crowded: Scale and Surge with their own environment offerings, environment-native startups like Mechanize, Fleet AI, and HUD, and open ecosystems like Prime Intellect. This market is young and definitions vary wildly between vendors, so scope carefully. We map the whole category in our guide to RL environment companies.
When the answer is data, not experts
Expert networks produce data to order: a doctor writes the ideal response, an engineer completes the task, a rater ranks the outputs. That is the right tool when the knowledge lives in people's heads.
It is the wrong tool when the knowledge lives in the world. A video model that needs real motion, a voice agent that needs natural conversation, a world model that needs gameplay and first-person footage, an agent that needs real workflow traces: no expert can write that data, it has to be captured from reality and licensed from whoever owns it. Commissioning experts to approximate it is slow and expensive, and the result still is not real. That is a different market, licensed data, and it is covered in our guides to AI training data marketplaces and where AI labs source training data.
The landscape
| Provider | Category | Best for |
|---|---|---|
| Mercor | Expert marketplace | Domain-expert demonstrations, evals, now environments |
| Surge AI | RLHF and evaluation | Frontier-grade preference data |
| Scale AI | Full-stack data services | Very large programs, Meta caveat applies |
| Turing | Expert data | Coding data and expert engineers |
| Micro1 | Expert recruiting | Vetted specialists on a similar model to Mercor |
| Handshake AI | Expert network | Expert data built on a recruiting platform |
| Toloka | Managed crowds to expert data | Global reach, lab-neutral positioning |
| Prolific | Vetted participant pool | Studies, evals, human baseline data |
| AfterQuery | Expert data and benchmarks | SFT and RLHF data with published benchmarks |
| Snorkel | Programmatic labeling | Reducing human labeling volume |
| Troveo | Licensed data marketplace | Rights-cleared video, audio, gameplay, and business data |
How to choose
Match the vendor to what the model is missing. If it lacks judgment or domain skill, an expert network is right, and the comparison is about vetting quality, domain depth, and price. If it lacks feedback, that is the RLHF lane. If it lacks practice on long tasks, that is environments. And if it lacks knowledge of the real world, no amount of expert output fills the gap, that is licensed data.
Then apply the diversification lens the market has learned: for anything mission-critical, know who owns your vendor, how concentrated your spend is, and what happens to your roadmap if that vendor stumbles. Companies like Mercor are impressive and busy; make sure your program gets the attention it is paying for.
Where Troveo fits
Troveo does not run an expert network. Troveo licenses the data experts cannot write: real-world video, audio, gameplay, and business data from more than 7,000 rights holders, cleared for AI training with documentation per asset. Labs use it for exactly the gaps expert data cannot fill, real motion, natural conversation, actual gameplay, genuine workflows, delivered training-ready. The two categories are complements: plenty of labs buy real-world data from Troveo and send it to an expert network for evaluation and feedback. Browse the catalog in Lens or talk to us about what your model is missing.