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

Mercor CEO Bets Token Spend Will Overtake Payroll

Brendan Foody argues that AI value is moving to the model labs and the infrastructure beneath them, because the model is the product. Applications face weak defensibility, and companies will spend more on tokens than on people. Mercor sits in the middle, supplying the expert judgment that trains the models.

Takeaways

• Application layer companies struggle to defend themselves. Brendan argues the model is the product, so frontier labs can expand into adjacent domains. Harry, who invests in Legora, says legal software needs separate product, sales and customer success teams. Brendan cites forward deployed services and network effects as defenses.
• Token spend will beat headcount within five years. Brendan predicts the average enterprise will spend more on compute than on people, because AI will automate services. Mercor already spends more on tokens for internal agents than on salaries, and its AI project manager finished a project end to end.
• Expert data demand rises as models automate work. The frontier model scores about 40% on the APEX productivity index, up from 1% a year earlier. Mercor pays experts over $3m a day. Models will clean data themselves, while humans codify tacit knowledge that is not written down.
• Mercor says the security incident did little harm. Brendan says Mercor engaged Mandiant quickly and added $300m of net new ARR in 60 days. OpenAI is closer than ever and Meta is paused. He denies the Handshake shift and the $13bn Amazon bid, and would not sell for $30bn.
• Mercor is profitable and cash rich. Mercor holds over $500m in cash and has been profitable since burning $500,000 after its seed round. Gross margins run 30 to 40%. Revenue is not GMV, since Mercor delivers end to end tasks powered by experts, as Uber is powered by drivers.
• Inference shifts toward open source and distilled models. Brendan predicts most inference in five years will run on open source, fine tuned or distilled models rather than frontier ones. He also sees at least one of OpenAI or Anthropic worth more than $10tn.

Mercor's funding path

1. Seed round in September 2023, when run rate revenue was about $1m.
2. Series A led by Benchmark at a $250m post money valuation.
3. Series B, which Brendan calls the most uncomfortable round, priced at about 100 times revenue.
4. A $10bn valuation when run rate revenue was about $400m.