Three post-2000 entrepreneurs, with a valuation of 70 billion.

What does it mean to drop out of school and start a business at 22, and build a unicorn valued at tens of billions of dollars at 24?

According to ZDXX on October 28th, today, Mercor, an AI recruitment unicorn in the United States, officially announced that it has secured $250 million (approximately RMB 1.8 billion) in new financing, reaching a valuation of $10 billion (approximately RMB 71 billion), which is five times its $2 billion (approximately RMB 14.2 billion) valuation in February this year.

This AI startup, founded in 2023, has now raised a total of $350 million (approximately RMB 2.5 billion) in financing. It has included the world's top five AI labs, such as OpenAI and Anthropic, in its customer list. Its 17 - month revenue run - rate has grown from $1 to $500 million (approximately RMB 3.6 billion).

The founders of this AI unicorn are three post - 2000s who dropped out of college in their sophomore year: CTO Adarsh Hiremath, CEO Brendan Foody, and COO Surya Midha. They dropped out of Harvard University and Georgetown University in 2023 to start their business together.

▲CTO Hiremath, CEO Foody, COO Midha (from left to right)

The business that helped them earn their first pot of gold is AI recruitment. It uses AI to screen resumes and quickly match candidates with positions. In February this year, based on this large professional talent network, Mercor launched data annotation and large - model evaluation services. That is, it signs contracts with existing expert talents to help large - model companies with data annotation and provide professional feedback in the short term. Now, the total number of experts it manages has reached 30,000, and the total daily salary of all experts exceeds $1.5 million (approximately RMB 10.65 million).

In February this year, Mercor's annual recurring revenue had reached $70 million (approximately RMB 497 million). With its new large - model evaluation business, this startup has found an "invisible gold mine" in the large - model evaluation track.

Mercor's new financing is led by venture capital firm Felicis, with participation from Benchmark, General Catalyst, Robinhood Ventures and other venture capital firms. The new financing will be used in three key areas: expanding the company's talent network, advancing the matching system and training opportunities among experts, and providing faster delivery.

It is worth mentioning that Scale AI, which had its shares acquired by Meta and its CEO poached, is a strong competitor of Mercor. However, after the turmoil, Scale AI's employees and customers have turned to Mercor, which has also doubled its revenue.

01. Dropping out of college in sophomore year to target AI recruitment, inadvertently building a huge high - quality talent network

The three founders of Mercor have very prominent labels: post - 2000s and dropped out of college in sophomore year.

Hiremath, Foody, and Midha were high - school classmates, all studying at Bellarmine College Preparatory in San Jose. They met in the school debate team and won the championship of the National Policy Debate Tournament in the United States as a team.

It is worth mentioning that Foody started his entrepreneurial journey in 2021. He founded Serosin, aiming to build the next - generation personal computer infrastructure in the cloud, and successfully reduced the cost of using high - performance computers by 90%.

In 2023, Hiremath, a sophomore at Harvard University, and Foody and Midha, sophomores at Georgetown University, all chose to drop out of school and focus on entrepreneurship. Mercor was established in the same year. At that time, Hiremath was majoring in computer science, while Foody and Midha were majoring in economics and diplomacy respectively.

In the early days of its establishment, Mercor's business scope was to use AI technology to screen resumes, match candidates with the most suitable positions, and conduct qualification reviews on candidates. Its services were mostly targeted at software engineers and technical positions related to mathematics.

Mercor's corporate customers describe job requirements and the desired candidates in natural language, such as "full - time Python developers with computer vision experience". Its AI tools can conduct in - depth semantic search queries on hundreds of thousands of resumes, personal portfolio websites, social platform X, AI interview records, and GitHub in a few seconds to find the best matches.

Then, customers can immediately watch the AI interview situation of candidates and add the matched candidates to the company with one click.

▲Job posting situation on Mercor's homepage

The startup's official website shows that in January 2024, Mercor's annual recurring revenue had reached the million - dollar level, and it had established a talent pool of 100,000 users in 25 countries and regions. Later, to meet the needs of talent recruitment, Mercor continued to expand its talent pool and helped human resources teams evaluate 468,000 applicants. India is its largest source of talent, followed by the United States, and the talent pools in Europe and South America are growing rapidly.

By February this year, in the process of promoting AI resume screening, Mercor inadvertently found that it had woven a large professional talent network, which is exactly what major AI companies are eager for. They hope to use these professional talents to train increasingly complex large models to improve their competitiveness.

This is because as the capabilities of models improve, they require professionals in specific fields to evaluate them in the short term, which means that AI companies need to quickly find corresponding talents and offer temporary positions.

After observing this trend, Mercor quickly expanded its scale and extended its business to large - model evaluation and data annotation. On the one hand, Mercor began to hire contractors who can evaluate the quality of chatbot answers and poached Sundeep Jain, the former chief product officer of Uber, to serve as its first president. On the other hand, it continued to expand the scale of its talent network and extended the scope of position screening to many industries such as lawyers, doctors, and journalists.

02. Part - time experts work 20 hours a week, and 30,000 experts can earn $16 million a day

Now, Mercor's business system for evaluating large - model capabilities has gradually matured.

Mercor currently manages 30,000 experts globally. These experts are responsible for tasks such as image annotation, sentence writing, and providing professional feedback, helping chatbots master human - like thinking and expression abilities. These experts can earn more than $1.5 million (approximately RMB 10.65 million) in total every day.

According to the company's contract list obtained by The Wall Street Journal, doctors working part - time as data labelers are tasked with evaluating AI's medical - related answers and reviewing AI - generated medical research. They can earn up to $170 per hour (approximately RMB 1,207) and must work at least 20 hours a week in a six - week contract. Based on a five - day workweek, experts need to work more than 4 hours a day on average, which means that doctors working part - time can earn at least $680 a day (approximately RMB 4,828).

In addition, if a customer pays Mercor $100 per hour (approximately RMB 710) for data labeling, Mercor will retain about 30% to 35%, and the rest will be passed on to the contractors. The average hourly wage in their contracts is about $85 (approximately RMB 603).

At the beginning of this month, Mercor officially announced its first - of - its - kind AI Productivity Index (APEX), which can evaluate AI models based on their ability to perform economically valuable knowledge work. Currently, APEX includes tasks representing four professional jobs: investment banking assistant, large - scale legal assistant, strategic consulting assistant, and general practitioner (MD).

APEX v1.0 consists of 200 cases, evenly distributed among investment banking, law, consulting, and healthcare. Each case consists of a prompt (task description), a source (information required to complete the task), and a scoring standard (criteria for scoring the model's response).

Its construction involves five steps: assembling a team of about 100 experts with top - level experience covering four professions; experts generating task descriptions or prompts to describe common workflows in each field; experts generating source documents containing relevant evidence required to respond to the prompts; experts generating scoring standards specific to the prompts; after experts generate prompts, sources, and scoring standards, they are reviewed by separate experts to ensure quality control.

Its blog mentions that professionals need 1 to 8 hours to complete tasks in APEX, with an average of 3.5 hours.

In May this year, HealthBench, a medical large - model test evaluation set released by OpenAI, also adopted this APEX system. Based on the evaluation results of APEX, GPT - 5 scored the highest at 64.2%, and the best - performing open - source model was Qwen3, ranking 7th with a score of 59.8%.

03. The turmoil at Scale AI boosts Mercor's revenue surge and leads to a commercial lawsuit

In addition to the revenue brought by its large talent network, the recent turmoil at data annotation startup Scale AI has also caused Mercor's revenue to surge.

In June this year, Meta acquired a 49% stake in Scale AI for $14 billion (approximately RMB 99.4 billion), pushing Scale's valuation up to a staggering $29 billion (approximately RMB 205.9 billion). Subsequently, as part of the deal, the company's co - founder and CEO, Alexandr Wang, moved to Meta to lead its AI work.

This has led some customers and competitors of Scale AI to express concerns about its ability to remain neutral and protect customer data after Meta's investment.

As a result, this deal has instead increased Mercor's revenue. According to The Wall Street Journal, citing people familiar with the matter, Mercor's revenue has quadrupled since Meta's investment in Scale.

Meanwhile, Mercor has also recruited many former Scale employees. Last month, Scale sued and accused Mercor of stealing trade secrets and sued former Scale employee Eugene Ling for breach of contract. The lawsuit revealed that this employee tried to promote Mercor to one of Scale's largest customers before officially leaving Scale. However, this lawsuit has not yet reached a conclusion.

In addition, there is a major debate surrounding Mercor that AI progress may accelerate the loss of recruitment jobs. However, Foody believes that Mercor is not replacing humans but automating most of the economy, making humans more valuable in areas where they are still needed.

He told foreign media TechCrunch: "If AI automates 90% of the economy, then humans will become the bottleneck for the remaining 10%. Therefore, each unit of economic output contributed by humans has a 10 - fold leverage effect because the rest has been automated. This means that as we shift to a more fragmented, gig - like work model, the way people work is changing. More and more companies are now hiring experts for short - term projects instead of relying on full - time employees."

04. Conclusion: Using AI recruitment to accumulate a large talent pool and fill the gap in large - model evaluation

Mercor automatically screens resumes and matches candidates, and provides AI - driven interviews and salary management. Enterprises upload job descriptions in natural language, and the system will recommend the best candidates. Relying on the large high - quality talent pool accumulated through this model, Mercor has unexpectedly become an "invisible winner" in the large - model evaluation track.

The iteration of large models relies on high - quality data and professional feedback. The large expert talent network built by Mercor exactly fills this industry pain point, making it a winner in the large - model track. This also shows that there are still many new possibilities for entrepreneurial opportunities in the AI era.

This article is from the WeChat official account “ZDXX” (ID: zhidxcom), author: Cheng Qian, published by 36Kr with authorization.