Mercor spends 3X employee salaries on LLM inference, CEO says

Why it matters

Mercor's ratio shows how AI-native businesses may replace traditional software and staffing costs with a large variable compute bill, while keeping humans in the workflow as employees and paid experts.

Brendan Foody (@BrendanFoody), Mercor's co-founder and co-CEO, said Sunday that the AI data business spends three times as much on large-language-model inference as it spends on employee salaries. He argued that the compute produces enough economic value for Mercor to keep hiring rather than use AI as a reason to shrink its workforce.

The claim offers an unusually direct view of the cost structure inside an AI-native business, though the ratio comes without dollar figures, a measurement period or a breakdown of which model providers receive the spending. "Employee salaries" is also a narrow denominator for Mercor, which relies on a large network of independent experts to train and evaluate models.

Mercor says those experts are currently paid more than $4 million a day. In May, Mercor reported paying more than $2 million daily to over 30,000 weekly active contractors. Those payouts are outside the employee-salary figure Foody used, making the 3X ratio a measure of inference against internal payroll rather than Mercor's total spending on people.

Foody linked Mercor's experience to a much larger prediction: AI diffusion could help produce roughly 10% economic growth within five years. In a follow-up reply, he invoked the Jevons paradox, the idea that efficiency improvements can increase total consumption by making a resource cheaper and useful in more places.

The macroeconomic conclusion reaches far beyond the evidence in Mercor's internal spending ratio. Mercor's business sits directly inside the AI industry's spending cycle: Mercor recruits experts for model developers, automates the evaluation of those workers and sells infrastructure for training and testing AI systems. Heavy inference consumption is central to Mercor's product and revenue, rather than a representative pattern across the wider economy.

Where the tokens go

Mercor has built LLM calls into the machinery used to recruit and match its expert workforce. Mercor's AI interviewer, called Monty, asks role-specific questions, transcribes responses and evaluates candidates. In March, Mercor said about 10,000 interviews were taking place each day, with a new session beginning every nine seconds. Its engineering team said speech recognition, the LLM and text-to-speech systems run across a mixture of commercial APIs and open-source models.

The volume continues after the interview. Mercor uses assessments, interview results, skills and availability to identify candidates for projects, including offers sent to people who did not apply for a particular role. Mercor said in March that more than half of offers on its platform were being sent proactively.

Mercor is also pushing beyond recruiting and human-data operations. Its enterprise product uses AI-led employee interviews, workflow data, model routing and automated evaluations to identify and deploy workplace agents. In July, Mercor agreed to acquire Deeptune, which builds simulated environments where agents can practice tasks involving enterprise software. Financial terms were not announced.

Those products give Mercor several ways to turn inference into revenue: screening more experts, matching them to projects, evaluating model outputs and operating agents for enterprise customers. Foody's claim is that spending another dollar on model calls creates enough work and throughput to justify additional employees around those systems.

A founder who automated his own bottleneck

Foody founded Mercor in January 2023 with Adarsh Hiremath and Surya Midha, friends he had known since high school. The founders initially used LLMs to automate resume review and interviews while building a recruiting operation from their Harvard and Georgetown dorm rooms. Mercor's original product searched resumes, portfolios, interview transcripts and GitHub profiles to match candidates with jobs.

Mercor later concentrated on supplying domain experts to AI laboratories. The shift made Mercor both a seller of human intelligence and an aggressive buyer of machine intelligence: contractors create and judge specialized work, while models help Mercor recruit, route and evaluate those contractors at scale.

Mercor said it was profitable and had crossed a $1 billion annualized revenue run rate earlier in 2026. Forbes reported in July that Mercor claimed to have reached $2 billion in annualized revenue in June. Mercor was valued at $10 billion in an October 2025 funding round led by Felicis, with Benchmark, General Catalyst and Robinhood Ventures participating. Reports this summer said investors were discussing another round at a $20 billion valuation, though Mercor has not announced one.

That growth makes the missing dollar amount behind Foody's 3X ratio consequential. If the comparison covers a substantial internal payroll, Mercor has become a major inference customer in its own right. If Mercor maintains a small employee base relative to billions of dollars in gross revenue and contractor payouts, the ratio says more about organizational design: a thin internal workforce coordinating large pools of models and outside experts.

Either reading captures the bet Foody is making. Mercor is treating inference as a variable input that should rise when each additional model call produces more interviews, evaluations, matches or customer work. Salaries become the smaller line item because Mercor's software is expected to multiply what each employee can operate.