20VC: Mercor CEO on Why Application Layer Companies Have No Defensibility, The Model is the Product | Token Spend Will Exceed Headcount Spend in 5 Years | The True Cost of Hiring AI Researchers in the Valley Today with Brendan Foody

What It Covers

Mercor CEO Brendan Foody discusses why application layer AI companies lack defensibility, how the foundation model layer will capture outsized value, and why token spend will surpass headcount costs within five years. Mercor operates at over $1B revenue, is profitable, and pays out $3M daily to its 5M-person talent network training frontier models.

Key Questions Answered

• •Application Layer Defensibility: Companies building software abstractions on top of foundation models face a structural threat: Claude and GPT can replicate vertical SaaS workflows within 12 months. The only durable moats exist where network effects operate — Salesforce's integration marketplace, Slack Connect, or Carta's cross-company data. Pure software layers without network effects will lose pricing power rapidly as model capabilities expand into their core use cases.
• •Token Spend Exceeding Headcount: Mercor currently spends more on inference tokens for internal AI agents than on employee salaries. Foody projects that within five years, the average Fortune 500 company will spend more on compute than total headcount. Enterprises should begin building workflow-specific evaluation frameworks now to benchmark models, enable hot-swapping between providers, and distill open-source models that match frontier performance at dramatically lower cost.
• •Agent Training as the Dominant Job Category: The fastest-growing job category is training AI agents to replace redundant knowledge work. Instead of a lawyer repeatedly redlining similar contracts, they train an agent once and amortize that effort across its lifecycle. Mercor pays $3M daily to workers performing this function and projects that figure to triple within 12 months, making agent training the defining labor market shift of the next decade.
• •Data Quality Power Law: Within any dataset of 10,000 tasks, the top 2,000 tasks generate the majority of model improvement value. High-quality, long-horizon tasks — multi-week financial modeling projects, end-to-end legal workflows coordinating multiple colleagues — drive disproportionate frontier model gains. Labs pay premium rates for experts who combine domain expertise (medicine, law, finance) with hands-on frontier model usage, as that combination identifies failure modes humans alone cannot surface.
• •Foundation Model Valuation Trajectory: Foody predicts at least one of OpenAI or Anthropic reaches $10T in valuation, driven by their position as teacher models that enable distillation of superior smaller models across every enterprise workflow. The majority of inference in five years will run on fine-tuned open-source or distilled models, but frontier labs capture value by setting the capability ceiling from which all downstream distillation derives its performance baseline.
• •Eval Frameworks as Enterprise Infrastructure: Academic benchmarks like GPQA and Humanity's Last Exam are being replaced by end-to-end workflow evals — can the model build a complete SaaS application, or coordinate a multi-week financial deliverable? Enterprises that build proprietary eval sets for specific workflows gain a 10x price-performance advantage by enabling precise model selection and distillation. This eval infrastructure becomes the system of record for all agent deployment decisions across the organization.