Mercor Raises $750M at $20B Valuation | CEO Brendan Foody
Interview, Fireside Chat
- The company currently disburses $5 million daily, a significant increase from $1 million annually 24 months prior, with plans to hire experts and full-time staff rapidly despite AI displacement narratives.
- Market demand for human trainers is predicted to expand exponentially driven by the fact that 99% of knowledge resides in human heads and top researchers expect models to require human feedback for decades.
- Inference spending by enterprises is expected to ramp materially, triggering an adoption curve where companies allocate roughly 10% of budgets to evaluation datasets upon reaching the $10 million to $100 million range, and begin considering post-training for their own models once surpassing $100 million.
- Future compute spend on training is forecasted to shift toward Reinforcement Learning (RL), requiring subject matter experts to create rubrics and tasks while companies build "apps" or clones of software like Salesforce for RL environments.
- By the time inference spending exceeds $100 million, the speaker expects a serious shift toward sovereign AI stacks and proprietary model training to avoid zero switching costs, with the belief that application-layer software will eventually be commoditized unless companies build their own frontier models to create data moats.
- The speaker forecasts that every Fortune 500 company will eventually resemble a "Neolab" mobilizing employees to curate high-quality tasks, with non-technology firms facing existential threats within the next 12 months serving as a significant catalyst.
- Agents are reported to be three to five times more productive for creating data and training agents, which the speaker expects will trigger Jevons' paradox leading to expanded demand for experts as productivity rises.
- Anthropic and OpenAI are predicted to potentially reach valuations exceeding $10 trillion, while the speaker believes the company itself will not conduct an IPO in the next 12 months, though it remains on the radar.
- Strategic positioning relies on the "time travelers" concept of working backward from a 5 to 10-year future, focusing on defining legendary category-defining categories rather than following competitors.
- Core values including customer obsession, high standards, and intrinsic motivation are expected to guide the organization, with the "emotional pull" of building extraordinary companies serving as the primary driver for founders and executives.
- The speaker anticipates that the "lump of labor fallacy" will become consensus as enterprises deploy agents, while M&A will remain a high-leverage tool to explore categories with high positive impact and solve app clone bottlenecks, exemplified by the DeepTune acquisition.
- Durable competitive advantages will vary by sector: economy of scale and recursive self-improvement for frontier labs, data and brand moats for application-layer companies like Harvey, and custom models for inference companies to mitigate low switching costs associated with APIs.
- Risks include the potential for investors to overrate revenue quality of API businesses due to low switching costs, the expectation that software layers will become commoditized as building software gets cheaper, and the necessity of a "shock absorber" dynamic for employees during rough operational patches.
- The market for human data creation is not expected to dissipate within a couple of years, and the full distribution of RL environments mirroring the global economy's apps, worlds, and tasks will require the business to cover diverse capabilities as long as superintelligence has not been achieved.