The CEO of Mercor argues that defensibility in the AI application layer is incredibly difficult to build. As foundation models like Claude improve, they will natively absorb the functionality of vertical-specific applications (e.g., for law, finance), making the underlying model the true, defensible product.
While data cleanliness is a challenge, AI models will become proficient at structuring data themselves. The true bottleneck for enterprise AI is codifying the vast amount of tacit knowledge that exists only in employees' heads. The new job of employees will be to translate this context for AI agents to perform effectively.
As enterprise spend on AI workflows explodes, companies will create custom evaluation benchmarks (evals) for each specific use case. These evals act as a system of record to hot-swap between different models based on price-performance, enabling perfect competition and ultimately commoditizing the API layer.
In the age of AI, a strong go-to-market team is not enough. The real defensibility comes from a "forward deployed" motion—a post-sales services layer that deeply embeds with customers to train agents on their specific, tacit internal knowledge. This is incredibly hard for competitors or foundation models to replicate.
AI has armed cyber attackers with a new weapon: swarms of coding agents. Unlike human attackers, these agents can exhaustively and rapidly review an entire codebase to find vulnerabilities, dramatically increasing the speed and scale of cyber threats. This necessitates a boom in AI-powered defensive tools.
Counterintuitively, a public security incident did not slow down Mercor's growth. By handling the crisis quickly, communicating proactively with customers, and engaging security experts, the company strengthened its relationships and added $300 million in net new ARR in the two months immediately following the event.
Software moats are diminishing rapidly. In the next 12-24 months, foundation models will gain the capability to build entire SaaS applications, like cloning Slack, from a simple prompt. This will severely challenge software companies that lack strong network effects, as their core product can be replicated with ease.
Mercor's Series B valuation of $2B on $20M ARR (a 100x multiple) seemed high but was justified by their track record of hypergrowth. They had consistently grown 50% month-over-month and accurately projected massive future revenue milestones, giving investors confidence in a valuation that priced in future performance.
Illustrating a dramatic shift in operational expenses, AI company Mercor now spends more on API tokens for its internal agents than on employee salaries. This is a leading indicator for how most enterprises will operate within five years, where compute costs will eclipse human capital costs.
The intense competition for elite AI talent has driven compensation to staggering levels. High-quality AI researchers now often receive offers valued in the tens of millions of dollars in stock per year, with one anecdote citing a $20 million cash-equivalent offer, highlighting a major challenge for startups.
While the cost-per-token is decreasing as models become more efficient, this efficiency gain drives a massive increase in new use cases and overall consumption. This economic principle, Jevons Paradox, explains why total enterprise spending on model inference is skyrocketing, even as the unit cost falls.
The future of knowledge work isn't about humans performing tasks, but about training an AI agent to perform them once. This is a structurally more efficient model, amortizing the initial training effort over the agent's entire lifecycle, which will create a new job category centered on agent management and training.