Mercor CEO Brendan Foody puts a number-shaped hole in the AI agent story
A 20VC interview frames Mercor as spending more on AI agent tokens than salaries, but the exact cost comparison remains unclear.
By Ryan Merket ยท Published
Why it matters
Mercor is being used as a live example of the AI-native cost structure founders and investors keep predicting: fewer humans per dollar of output, more spend on agents and model usage, and a new set of margin questions.
Brendan Foody (@BrendanFoody), founder and CEO of Mercor (@mercor_ai), has turned a familiar AI operating thesis into a concrete test case: in a post on X linking to a new 20VC interview, Mercor is framed as spending more on AI agent tokens than on human salaries.
That claim is striking because Mercor is not being pitched as an ordinary software company. The 20VC with Harry Stebbings episode describes Mercor as a data provider to major AI labs, including OpenAI, and says Foody has scaled Mercor in the last two years to $1.5 billion in ARR and a $10 billion valuation. Those figures come from the episode description, not from filings or company financials available in the source material.
Foody's broader argument in the interview is less about a single spending line and more about what kind of company survives as models get better. The visible transcript opens with him saying, "Building defensibility in the software layer on top of the models is going to be incredibly difficult." The episode's own chapter titles point to the same theme: "Infrastructure Will Win Over Application Layer" and "Is SaaS Dead? When Network Effects Are the Only True Moat."
The claim is powerful, but imprecise
The cleanest version of the story is that Mercor now spends more on AI agent tokens than on human salaries. That is how the linked episode title and summary present it.
But the underlying source material is not perfectly consistent. The YouTube title says "Token Spend Exceeds Salaries," while the episode timestamp labels the relevant section "Token Spend on Agents Now Exceeds Employee Headcount." Salaries and headcount are not the same measure. One is a dollar cost. The other is a count of people. Without the full quote or a disclosed token-spend figure, the safest reading is that Foody is describing Mercor as an AI-native company whose operating costs are shifting toward model usage and agent infrastructure, not simply that a verified line item has overtaken payroll.
That distinction matters. A company can have high token spend because agents are replacing work, because revenue is growing quickly, because workflows are inefficient, or because it is subsidizing automation ahead of margin improvement. The source material does not disclose Mercor's employee count, salary expense, gross margin, token bill, burn, or cash balance.
Why Foody is making this argument now
Mercor sits inside one of the most consequential fights in AI: whether durable value accrues to model labs, infrastructure providers, labor-data networks, or application software. If Foody is right that application-layer companies have thin defensibility, Mercor's pitch becomes clearer. Mercor is presenting itself less as another SaaS layer on top of models and more as part of the supply chain that makes frontier AI work.
That also explains the emphasis on tokens. In a conventional startup, employee count is a proxy for scale. In an AI-native company, Foody is suggesting the more important operating question may become how much autonomous work can be routed through agents, and at what cost. The comparison flatters companies that can grow revenue without growing headcount at the same rate.
The incentives cut both ways. Investors like the idea of software businesses with lower human labor intensity. Founders like the idea of smaller teams moving faster. But customers and analysts still need to know whether token-heavy operations produce better margins, better service levels, or just a new category of cloud-like expense.
The unanswered parts are the story
The 20VC description also tees up sensitive questions around Mercor's customer base and growth. One chapter asks, "True or False: Mercor lost Meta & OpenAI as a customer with the hack?" That is phrased as a question in the source material, so it should not be read as confirmation that Mercor lost either customer. Another chapter refers to "Rejecting a $30B Acquisition," but the scrape does not identify a buyer, timing, structure, or whether a formal offer existed.
What is verifiable from the available material is narrower but still important: Foody is publicly arguing that the software layer is hard to defend, that Mercor's model is tied to the AI infrastructure and labor supply chain, and that AI agent usage has become a central operating cost inside Mercor.
For founders, that is the useful signal. The next AI-native company may not look like a 2010s SaaS company with a sales team, a support team, and a steadily rising employee graph. It may look more like a small human organization wrapped around a very large and very measurable token budget. Whether that is a margin breakthrough or just a new dependency on model providers is the question Foody's own example now puts on the table.