What happens when a company spends more on AI than on its workers? Mercor’s CEO says his startup is already finding out.
“Right now we’re spending more on tokens for our internal agents than we are on employee head count,” Foody said during an appearance on the “20VC” podcast on Monday.
When host Harry Stebbings asked if Mercor’s token spending on AI agents exceeded salaries, Foody replied: “That’s correct. It’s pretty incredible.”
Mercor — a $10 billion startup that helps companies such as OpenAI and Anthropic train AI models through a network of its human experts — has become one of the fastest-growing companies in the AI ecosystem since its 2023 launch.
As of October 2025, per PitchBook, it had around 300 employees. The company did not respond to a request for comment.
Foody said Mercor uses AI agents across a wide range of functions, including project management, recruiting, accounting, fraud detection, and candidate evaluation. The company has conducted more than 5 million AI-assisted interviews, he said.
The executive believes Mercor’s spending patterns foreshadow a broader shift across corporate America.
“I would bet that in five years the average enterprise spends more on compute than headcount,” Foody said.
When AI costs more than employees
Foody’s comments come amid a broader debate among executives over whether rising AI spending is translating into meaningful business returns.
Uber COO Andrew Macdonald recently said he has yet to see a clear link between rising AI spending and proportional productivity gains.
Foody said that falling costs and rapidly improving model capabilities are driving a Jevons paradox-style effect, where cheaper AI leads to significantly more consumption rather than less.
He said Mercor measures the performance of different AI models for specific business tasks and evaluates whether newer models offer better value.
The result, he said, is a future in which AI becomes a core operating expense for companies, potentially rivaling or surpassing the cost of human labor itself.
“Humans will still play an important role at the things models can’t do,” he said. “But I expect that cost of inference, cost of compute will exceed that.”
