Inside the Expert-Data Boom: How Scale, Surge and Mercor Became Billion-Dollar Businesses

The early story of large language models was one of scale: more text, more parameters, more compute. As models matured, the bottleneck shifted. The general knowledge available on the public web had largely been absorbed, and further improvements depended on teaching models to reason like specialists, follow complex instructions and avoid subtle mistakes. That kind of improvement cannot be scraped. It has to be produced by people who know the subject.

Out of that need came a new kind of company. These firms recruit, manage and quality-check large networks of skilled contributors who write, rank, correct and evaluate model outputs. Their customers are the leading laboratories, and their product is high-quality human data: worked solutions to hard problems, expert critiques of model answers, and carefully labelled examples in fields from medicine to software engineering.

In the space of a few years, this segment has produced some of the most valuable private companies in technology. Three of them, Scale AI, Surge AI and Mercor, show how much the market is prepared to pay for human expertise delivered as data.

Scale AI began by labelling images for self-driving car companies and expanded into the data that powers language models. In May 2024, it raised $1 billion at a valuation of $14 billion, with Amazon and Meta among its backers, according to CNBC (opens in a new tab).

In June 2025, Scale announced (opens in a new tab) a significant investment from Meta that valued the company at more than $29 billion. The New York Times reported (opens in a new tab) the investment at $14.3 billion for a 49% stake. As part of the arrangement, founder Alexandr Wang joined Meta to work on its AI efforts while remaining on Scale’s board, and the company’s chief strategy officer, Jason Droege, became interim chief executive, CNBC reported (opens in a new tab).

The size of the investment is the clearest signal yet that access to high-quality training data is now a strategic priority for the largest technology companies. It is rare for a company of Meta’s scale to commit that much capital to a supplier, and the decision reflects the view that data quality is one of the main levers left for improving frontier models.

Surge AI, founded in 2020 by former Google and Meta engineer Edwin Chen, took a different route. Rather than raise venture capital, it grew from customer revenue alone. In July 2025, Reuters reported (opens in a new tab) that Surge had passed $1 billion in revenue in 2024, more than Scale over the same period, and that it was seeking to raise up to $1 billion in its first outside capital. Later that month, Bloomberg reported (opens in a new tab) that Surge was in talks for funding at a valuation of about $25 billion.

Chen has described the company’s focus as the most demanding end of the data market: highly skilled contributors working on complex tasks for the top laboratories. Inc. reported (opens in a new tab) that the business was profitable from early on, which is unusual in a sector where many companies are funded to grow ahead of revenue.

Surge’s results show that the market rewards quality over volume. The highest-value human data is not about having the most contributors, but about having the right ones, and being able to prove that their work is accurate.

Mercor started as an AI-powered recruiting platform and became one of the fastest-growing suppliers of expert data. In October 2025, it raised $350 million in a Series C round that valued the company at $10 billion, TechCrunch reported (opens in a new tab), five times its valuation earlier in the year. CNBC reported (opens in a new tab) that the round was led by Felicis Ventures, with participation from Benchmark, General Catalyst and Robinhood Ventures.

The company connects AI laboratories with professionals such as doctors, lawyers, bankers and engineers, who are paid to create and evaluate training data in their fields. According to TechCrunch, Mercor pays more than $1.5 million a day to its network of experts, which numbers in the tens of thousands.

That figure illustrates where the value in this market is flowing. A substantial share of the money laboratories spend on data ends up with the people and organisations who hold the underlying expertise. Business Insider’s survey (opens in a new tab) of the segment found a wave of companies raising large rounds on the same model.

The three companies differ in history, funding and approach, but they share a thesis. Frontier models improve most when they learn from skilled human judgement, and that judgement is scarce. The companies that can source it reliably, check its quality and deliver it in a form laboratories can use have been rewarded with valuations that put them among the most valuable private businesses in the world.

The expert-data companies mostly produce new data on demand, by paying skilled people to create it. But the expertise they are recruiting is already embodied in the records of many organisations: the case notes of a professional services firm, the technical documentation of an engineering business, the annotated archives of a research institution or the decision logs of a specialist operator.

That existing material can be valuable for the same reason that new expert work is. It captures how skilled people approach real problems, in context and at scale. When it is genuinely owned, properly de-identified and documented, it can be licensed to the same laboratories that pay billions for freshly created data.

There is also a practical lesson in how these companies win business. Laboratories pay for quality they can verify. Surge, Scale and Mercor have each invested heavily in quality control, because accuracy is what customers buy. Organisations considering licensing their own expert data will find buyers asking similar questions: how the data was created, who created it, how it was checked and what rights attach to it.

Large language models learn general knowledge from broad text. Making them reliable in specialist work requires something different: examples of how experts think through problems, where they spot errors and how they explain their reasoning. Techniques such as reinforcement learning from human feedback depend on skilled people ranking and correcting model outputs, and the quality of the resulting model depends heavily on the quality of those judgements.

As models have moved into demanding fields such as medicine, law, finance and software engineering, the bar for useful feedback has risen. A general contributor can tell whether an answer reads well. Only a specialist can tell whether a complex tax analysis, a medical differential or a piece of production code is actually right. That is why the leading suppliers have shifted their recruitment toward highly qualified professionals, and why the price of their work has risen accordingly.

The same logic explains why evaluation has become a business in its own right. Laboratories need to measure how well their models perform on hard, realistic tasks, and building those evaluations requires the same expert input as building training data. Suppliers that can do both are deeply embedded in their customers’ development cycles.

The economics of the expert-data business rest on three features. First, demand is recurring. Each new model generation needs fresh, high-quality examples, and laboratories retrain and refine their models continually. That turns customers into long-term relationships rather than one-off buyers.

Second, the work commands high prices because the inputs are scarce. Qualified professionals are expensive and in limited supply, and laboratories are willing to pay for their time because a small amount of excellent data can improve a model more than a large amount of mediocre data. Mercor’s reported payments of more than $1.5 million a day to its experts give a sense of the scale of the market for this work.

Third, quality is the main competitive differentiator. Suppliers compete on their ability to recruit the right people, check their work and deliver consistent results. Surge AI’s reported growth to more than $1 billion in revenue without outside funding suggests that customers will pay a premium for suppliers they trust to get quality right.

Together, these features explain the valuations. Investors are paying for businesses with recurring revenue, pricing power and a direct role in improving the most important technology of the decade.

The expert-data market is shaped by a relatively small number of very large buyers. The frontier laboratories and the technology groups that back them account for a large share of spending, and their requirements set the standard for everyone else. When those buyers raise their bar for quality, provenance or security, suppliers across the market adjust.

Meta’s investment in Scale AI illustrates the point. A buyer of that size taking a large stake in a supplier signals how strategic the supply of training data has become. Other laboratories have responded by diversifying across several suppliers, which has helped companies such as Surge and Mercor grow quickly alongside Scale.

For organisations holding expert material, this concentration is useful context. The standards set by the largest buyers, around documentation, quality control and data handling, are the standards any serious dataset will be measured against. Meeting them opens the door to the most valuable part of the market.

Scale, Surge and Mercor are the most prominent names, but they are not alone. Business Insider’s survey (opens in a new tab) of the segment in September 2025 described a wave of AI training companies raising large rounds on the same thesis: that skilled human input is essential to improving models, and that the businesses able to supply it at quality will be valuable.

The breadth of investment is itself a signal. When capital flows into many companies in a segment, rather than a single leader, it suggests investors expect demand to remain strong and to grow. For the expert-data market, that expectation rests on the continuing need for each new generation of models to be trained and evaluated on harder, more realistic tasks.

The competition among suppliers also benefits the people and organisations who hold expertise. As more companies compete to recruit skilled contributors and source specialist material, the value placed on that expertise rises. The professionals and organisations at the start of the supply chain are well placed to capture a share of that value.

It is notable, too, how quickly these companies have scaled. Mercor’s valuation rose fivefold within a single year, and Surge reached more than $1 billion in revenue within roughly four years of its founding. Growth at that pace is rare in any industry, and it reflects how urgently the leading laboratories need the work these companies supply.

The speed also says something about the size of the opportunity still ahead. Each of these companies built its business largely by producing new expert data on demand, paying skilled people for their time. A great deal of equally valuable expertise already exists in recorded form inside professional firms, research institutions and specialist operators. As laboratories look for more efficient ways to source high-quality material, organisations that can offer existing expert records, properly prepared and documented, are well placed to become an important part of the supply chain alongside the specialist firms.

The next phase of this market is likely to be defined by specialisation. As laboratories build models for medicine, law, finance, science and engineering, they need data that reflects the depth of each field. That favours suppliers, and data owners, with genuine domain expertise over generalists.

Verification will matter more as well. Buyers increasingly want evidence that data is accurate, original and properly licensed. Suppliers that can demonstrate quality and provenance will command higher prices, while those that cannot will compete on cost.

Finally, recurring supply is becoming the norm. Models are retrained and updated continually, and each new generation needs fresh, high-quality examples. That turns expert data from a one-off purchase into an ongoing relationship, which is what has allowed companies like Surge and Mercor to grow revenue so quickly.

The expert-data boom is the clearest evidence yet that, in the current phase of AI, skilled human judgement is one of the most valuable inputs in technology. The organisations that hold it are in a strong position.