79% of Enterprises Now Have AI Agents Running in Production. The Genpact Survey Found Only 15% Can Tie the Spend to Profit. Here's the Accountability Gap That Separates Them.
With $614M in H1 2026 revenue — 90% from frontier AI labs — Mercor's pivot from hiring platform to training-data marketplace is the most important AI infrastructure story nobody is covering.
Mercor's $750 million Series D, announced November 10, 2026, at a $20 billion valuation, is the most important AI infrastructure funding event most people working in technology did not read about. The raise doubles the company's valuation from the $10 billion mark it hit just over a year earlier. The investors named in the round — Nvidia, Salesforce, General Catalyst, Benchmark, and Apollo-affiliated funds — are not making a bet on a startup. They are making a bet on a thesis: that the most defensible infrastructure position in AI over the next three years is not the model, not the compute, and not the application layer. It is the pipeline of human expertise that makes models good enough to sell.
That thesis is what Mercor is building. And the numbers support it. The Information reported $614 million in gross revenue for H1 2026, more than 70% above Mercor's full-year 2025 figure. Over 90% of that revenue came from frontier AI model companies. The gross annualized run-rate crossed $2 billion by mid-2026, according to multiple reporting sources. These are not typical startup numbers. They are the numbers of a company that has quietly become load-bearing infrastructure for the AI lab ecosystem.
From Interview Platform to AI Infrastructure: The Pivot That Changed Everything
Mercor was founded in 2022 by two founders who were 21 years old at the time. The original product was an AI-powered technical interview platform — a tool for automating the first-round evaluation of software engineering candidates. The company raised $30 million on the strength of that product, then another $100 million at a $2 billion valuation in early 2025, as reported by TechCrunch.
The pivot was not a strategy pivot in the traditional sense — it was a demand pivot. AI labs had begun requesting access to Mercor's expert network not to hire candidates but to source structured task completions. A model that was being trained for medical reasoning needed doctors who could grade its outputs against established clinical standards. A model being trained for legal analysis needed attorneys who could evaluate whether its interpretation of a contract matched what a competent lawyer would conclude. The tasks were different from interview evaluations in structure, but the underlying requirement was identical: find a domain expert, verify they actually have the expertise they claim, route them to a task that matches their knowledge level, and pay them reliably for their output.
Mercor's hiring platform had already built the infrastructure for all four steps. The pivot was largely a matter of redirecting that infrastructure toward a different customer — the AI lab — and a different task type — RLHF data and evaluation — rather than the employer and the candidate screening interview.
By 2024, the revenue from AI lab data work had eclipsed the hiring product. By H1 2026, the hiring product had been effectively superseded. In February 2026, Mercor formalized the commitment with the acquisition of Sepal AI, a startup that had been building evaluation benchmarks, reinforcement learning environments, and training data pipelines specifically for frontier language models. The acquisition gave Mercor the technical infrastructure to move up the stack from raw preference annotation toward structured evaluation and RL environment construction — the more complex and higher-margin work that the most advanced lab programs require.
What Expert Data Actually Is and Why the Post-Training Market Needed It
To understand why Mercor is worth $20 billion, you need to understand the structure of AI model development and where the most expensive bottleneck now sits.
Pre-training — the phase in which a model learns from massive quantities of text, code, and other data — is fundamentally a compute and data-volume problem. The frontier labs have solved for pre-training at a cost that remains high but that scales predictably with hardware investment. The AI training data war that defined 2024 and 2025 was largely a pre-training data war: who controls the highest-quality text corpus, and can internet data ever be exhausted as a training source.
Post-training is a different problem. Post-training — which encompasses fine-tuning, reinforcement learning from human feedback, constitutional AI training, and the construction of evaluation suites — is not a data-volume problem. It is a data-quality problem. A model that has completed pre-training knows a lot about the world in a statistical sense. Post-training is the process by which that model learns to be useful, accurate, helpful, and safe. And usefulness, accuracy, helpfulness, and safety in specialized domains cannot be evaluated by generalists.
A medical AI model that is wrong in plausible-sounding ways is not just useless — it is dangerous. Detecting the specific failure modes that make a medical model dangerous requires medical domain expertise. The model's training signal for avoiding those failure modes must come from people who can recognize them. That is what Mercor sells: the pipeline of expert recognition at scale.
The same logic applies across every high-stakes domain the frontier labs are targeting: legal analysis, financial reasoning, scientific research, software engineering at a senior level, and educational content designed to teach rather than merely inform. For each of these domains, the post-training bottleneck is not compute or data volume. It is the availability of domain experts who can evaluate model outputs accurately and whose evaluations can be trusted by the lab to produce reliable training signal.
The Expert Marketplace Model: Vetting, Routing, and the Data Flywheel
Mercor's operational model has three layers that are more complex than a simple freelancer marketplace. The first is expert recruitment and vetting. The company's AI-driven vetting system evaluates candidates' claimed expertise against task-completion performance on structured samples. This is a capability Mercor developed for hiring interviews and adapted directly to the expert data use case: an annotator who claims cardiology expertise but whose evaluations of model outputs diverge systematically from the consensus of a panel of verified cardiologists is flagged and either re-routed to lower-complexity tasks or removed from the network. The vetting system makes the network self-improving over time: as more experts complete more tasks, the system builds a richer model of which experts produce reliable signal on which task types.
The second layer is task routing. Mercor does not simply match any expert to any task. It uses a matching algorithm that optimizes for expertise alignment, historical performance on similar tasks, and turnaround time. For a lab running a large-scale RLHF evaluation sprint, the difference between a well-routed expert pipeline and a poorly-routed one is the difference between a training dataset that produces reliable improvement and one that produces noise. The routing quality is a direct function of the size and diversity of the expert network — a network of five million experts across dozens of domains allows for routing precision that a network of fifty thousand cannot match.
The third layer is payment infrastructure. Mercor pays over 30,000 active expert contractors at scale, managing tax documents, payment schedules, and compliance across multiple jurisdictions. This is operationally complex and is a moat in its own right: labs that want to manage their own expert data pipelines would need to build this infrastructure themselves or contract with a payroll provider that lacks the vetting and routing layers. The bundled offering — expert network plus vetting plus routing plus payment — is what makes Mercor's product defensible against disaggregation.
Unit Economics: The Take Rate and What It Means at Scale
Mercor's economics are those of a high-quality services marketplace. Expert contractors on the platform typically bill at $80 to $170 per hour for domain-specific annotation and evaluation tasks, with the higher rates reserved for recognized specialists in fields like medicine, law, and senior software engineering. Reports from industry analysis suggest Mercor retains approximately 35 to 40% of the billing rate as its take, paying experts the remaining 60 to 65%.
| Metric | Estimate | Source / Basis |
|---|---|---|
| Gross revenue run-rate (mid-2026) | ~$2B | Multiple reporting sources |
| H1 2026 gross revenue | $614M | The Information |
| Revenue from AI labs (% of total) | 90%+ | Cryptobriefing, Kucoin reports |
| Expert network size | 5M+ | Company statements |
| Active contractors (est.) | 30,000+ | Industry analysis |
| Expert pay rate range | $30–$170/hr | Public job listings |
| Estimated take rate | 35–40% | 36kr analysis of contractor pay reports |
| Series D valuation | $20B | Multiple news reports, Nov 10, 2026 |
| Prior round valuation (Oct 2025) | $10B | Historical reporting |
At a $2 billion gross run-rate and a 37% take rate, Mercor's net revenue would be approximately $740 million annualized. At a $20 billion valuation, that implies a revenue multiple of roughly 27x on net revenue — elevated, but consistent with infrastructure businesses that have demonstrated they are on the critical path for a large and growing market.
The AI inference economics debate has focused on how to reduce the cost of running models at scale. Mercor's business is on the other side of the same equation: reducing the cost of making models good enough to run at scale in the first place. If inference efficiency is the story of AI economics in 2025, post-training data quality is the story of 2026 and 2027.
The Concentration Risk: Why 90% Is Both the Strength and the Vulnerability
Mercor's 90%-plus revenue concentration from frontier AI labs is simultaneously its greatest strength and its most visible vulnerability. The strength is obvious: being the dominant supplier to Anthropic, OpenAI, Google DeepMind, and Meta AI for post-training data work means being deeply embedded in the most important AI programs in the world. These are not customers who switch suppliers casually. The switching cost is high — rebuilding an equivalent expert network, vetting infrastructure, and routing system is a multi-year investment — and the risk of degrading the quality of a critical training run is too high for labs to accept lightly.
The vulnerability is equally obvious. If the top three labs, which likely account for the majority of Mercor's $614M H1 revenue, all shifted strategy simultaneously — moving toward synthetic training data generated by other models, automating evaluation tasks that currently require human experts, or reducing post-training investment in favor of longer pre-training runs — Mercor's revenue profile could deteriorate faster than its cost structure. This is not a theoretical scenario: the history of AI tooling is full of companies that were indispensable to one phase of the development cycle and became less critical as the cycle evolved.
The AI evaluation infrastructure market is addressing this same tension from a different angle: building evaluation systems that reduce the per-unit cost of human evaluation while maintaining its quality. If those systems succeed, the aggregate demand for expert annotation may not fall — but the premium per expert hour could compress, which would reduce Mercor's take rate and revenue even as the number of tasks grows.
Mercor's disclosed response to the concentration risk is diversification: a new enterprise product that turns employee workflows into agent specifications, an expanded domain mix beyond pure language model RLHF, and a broader geographic distribution of the expert network. As of H1 2026, that diversification is early-stage. The $750M raise gives it the runway to prove out the diversification thesis, but investors pricing the $20B valuation are largely betting on the existing concentration holding and the total post-training market growing — not on the diversification story being enough to protect against a lab budget shift.
Amazon's Mechanical Turk Is Fading. Mercor Is What Comes Next.
The comparison to Amazon's Mechanical Turk is instructive precisely because of how different the two businesses are. Mechanical Turk is a commodity annotation marketplace: tasks are simple (image labeling, text classification, transcription), pay is low ($1 to $10 per task), quality is variable, and the value proposition is volume at minimum cost. Amazon has not invested significantly in Mechanical Turk's development for years, and reporting from multiple industry sources describes it as a product that Amazon is allowing to wind down as AI automates the simple tasks it historically handled.
Mercor is the inverse: high-expertise tasks, premium pay, AI-driven quality filtering, and a value proposition built on accuracy and trust rather than volume and minimum cost. The tasks Mercor sources — evaluating a model's reasoning about a complex legal brief, grading a model's understanding of a rare medical condition, assessing a model's ability to debug an intricate piece of enterprise software — are not automatable in the near term because automating them would require the AI model to already be as good as the expert doing the evaluation. The irony is that the better frontier models become, the more important it is to have experts capable of evaluating the edge cases the models are still getting wrong.
The shift from pre-training data moats to post-training quality moats is the defining transition in AI infrastructure economics for 2026. Mercor is positioned at the center of that transition.
Five Questions for Evaluating AI Training Data Marketplace Businesses
The Mercor story is likely not unique. As post-training becomes the dominant investment thesis in AI development, additional expert marketplace businesses will emerge, raise large rounds, and claim defensible moats. The following framework distinguishes the durable ones from the momentum plays.
1. Does the vetting quality improve with scale? The defining characteristic of a durable AI data marketplace is a vetting system that gets better as the network grows. If expert quality assessment is manual and linear, scale is a cost driver, not a moat driver. If the vetting system is data-driven and self-improving, scale creates compounding quality advantages that are difficult to replicate from a standing start.
2. What is the domain specificity of the task portfolio? Commodity annotation tasks — image labeling, simple preference ranking, text classification — are automatable by definition: when AI gets good enough at a domain, it can perform the annotation itself. The durable portion of the expert data market is the tasks that cannot be automated because automating them requires the AI model to already have the capability the annotation is meant to build.
3. What is the revenue concentration profile? A marketplace that derives 90% of revenue from three customers is structurally different from one with twenty customers at 5% each. Neither is inherently bad — concentrated customers often indicate deep integration and high switching costs — but it requires explicit analysis of what happens if the top customers shift strategy.
4. Can the platform expand beyond its founding use case? Mercor started in hiring and pivoted to RLHF. The question for the next five years is whether its expert network and routing infrastructure can serve enterprise use cases — employee workflow documentation, audit trails for AI decisions, domain expert consultations — that are structurally different from lab post-training pipelines but require the same underlying capability.
5. What is the competitive moat against in-house building by the largest customers? The largest AI labs have the engineering capacity to build their own expert annotation platforms. The question is whether the switching cost from an external marketplace is high enough that it is not worth doing. Mercor's vetting infrastructure, five-million-expert network, and routing system represent multiple years of development that would be expensive to replicate. The labs' own experience attempting to source expert data at scale before Mercor existed is the strongest evidence for the moat — they tried, found it expensive and slow, and bought from Mercor instead.
Takeaway: Mercor's $750 million Series D is the clearest signal yet that post-training data quality is the defining infrastructure bottleneck in frontier AI development. The company has built, through a demand-driven pivot from hiring automation, the expert marketplace that makes AI labs' most critical training investments possible. Its $20 billion valuation reflects the bet that the labs will continue to need human expertise at the edge of model capability for as long as models are being pushed into new specialized domains — and that the infrastructure for sourcing, vetting, and routing that expertise at scale is worth more than any single lab is willing to build internally. The investors who led the round have seen the lab training budgets from the inside. They have made their bet.
Frequently Asked Questions
What does Mercor actually do in 2026?
Mercor operates as a two-sided expert marketplace for AI training data. On one side, it maintains a network of more than five million domain experts — doctors, lawyers, engineers, financial analysts, software developers — who complete structured tasks for AI labs: ranking model responses, grading outputs against criteria, writing demonstrations, and running adversarial evaluations. On the other side, it sells access to these expert pipelines to frontier AI labs, primarily as post-training infrastructure for reinforcement learning from human feedback (RLHF), evaluation suite construction, and reinforcement learning environment design. Mercor's AI-powered vetting system screens and routes experts to tasks that match their knowledge level, which is the core of its operational moat. The company started in 2022 as an AI-powered hiring platform automating the interview process for software engineers. By 2024, demand from AI labs for structured expert data had grown faster than its hiring business, and the company pivoted. By H1 2026, over 90% of its revenue came from frontier AI model companies, and its gross revenue run-rate had reached approximately $2 billion.
Why did Mercor pivot from AI hiring to AI training data?
The pivot was demand-driven. Mercor's core capability — AI-accelerated vetting of human expertise at scale — turned out to be more valuable to AI labs building post-training pipelines than to employers conducting hiring interviews. The post-training market for RLHF and evaluation data exploded as frontier labs moved from pre-training (where the bottleneck is compute and raw data) to post-training (where the bottleneck is high-quality, expert-graded preference data and hard evaluation tasks). Mercor's existing infrastructure for recruiting, vetting, and paying knowledge workers at scale mapped almost directly onto the needs of AI labs sourcing expert annotators and evaluators. The company accelerated the pivot by acquiring Sepal AI in February 2026, a startup specializing in training data, evaluation benchmarks, and reinforcement learning environments for frontier language models. The result is a business where the same underlying infrastructure produces fundamentally different revenue depending on which side of the market is growing faster.
How does Mercor's business model work, and what are the unit economics?
Mercor operates as a marketplace intermediary, billing AI labs for expert time and retaining a portion of that billing. Public job listings and contractor reports indicate domain experts on the platform typically earn $80 to $170 per hour depending on expertise level. Reports suggest Mercor bills labs at a gross rate and pays experts approximately 60 to 65% of that, implying a take rate of roughly 35 to 40% — consistent with high-quality professional services marketplaces. At a $2 billion gross run-rate and that take rate, Mercor's net revenue would be approximately $700M to $800M annualized. The business benefits from a structural moat: the expert vetting system improves as more experts complete more tasks, creating a data flywheel that makes routing more accurate over time. The primary risk is revenue concentration: over 90% of Mercor's gross revenue comes from a small number of frontier AI labs, which means its financials track lab post-training budgets almost directly.
What is the biggest risk in Mercor's $20B valuation?
The 90%-plus revenue concentration from frontier AI labs is the central risk in Mercor's $20 billion valuation. If a major lab slows its post-training investment — because it has completed a training run, because a new training technique reduces the need for expert-generated data, or because it builds competing in-house capability — Mercor's revenue can decline sharply with minimal warning. This is a known risk in marketplace businesses that serve enterprise customers: the platform is structurally exposed to its top customers' budget cycles and technology choices. The analogous case is any professional services firm that derives a majority of revenue from a handful of large clients. Mercor's response to this risk appears to be diversification — expanding the expert network to new domains, building toward an enterprise product that turns employee workflows into agent specifications, and growing the share of revenue from non-lab customers. But as of H1 2026, the concentration risk remains the defining feature of the financial profile, and any investor pricing the $20B valuation is implicitly making a bet that frontier lab post-training budgets will continue to grow.
What does Mercor's $750M raise signal about the AI training data market?
The $750 million Series D at a $20 billion valuation — doubling from the $10 billion mark in October 2025 — signals that institutional investors with deep AI knowledge believe frontier lab post-training budgets will continue to grow substantially, and that Mercor has built a defensible position in the supply chain for the expert data those budgets fund. It also signals that the AI training data market is bifurcating: commodity annotation at scale is declining in value as AI automates simple labeling tasks, while high-expertise annotation — the kind that requires a doctor, a securities attorney, or a senior software engineer — is growing in value as labs push their models further into specialized domains. Mercor's $750M raise is a bet that the most valuable AI training data in 2027 and 2028 will require genuine domain expertise, and that the marketplace infrastructure to source, vet, and deploy that expertise at scale will be worth more than what any single AI lab is willing to build internally.
How does Mercor compare to Scale AI and Surge AI for enterprise AI teams?
Scale AI, through its Outlier product, is Mercor's closest direct competitor for high-expertise RLHF and evaluation tasks. Scale AI has a larger enterprise footprint and a broader product portfolio that includes data annotation for computer vision and autonomous systems, not just language models. It also has established relationships with the US Department of Defense and commercial enterprises, giving it a more diversified revenue base. Surge AI focuses on a smaller, highly curated pool of annotators at the premium end of the market. Mercor's differentiation versus both is its AI-driven vetting and routing infrastructure, which it claims produces higher task-match quality than manual expert assignment, and the breadth of its expert network — five million experts across domains is an order of magnitude larger than most competitors. For enterprise AI teams evaluating these platforms, the key questions are task complexity, required expertise level, and speed versus quality tradeoffs. For large-volume commodity tasks, Scale AI has the infrastructure advantage. For specialized domain evaluations, Mercor's routing capabilities and network breadth are the strongest claims.