AI Investors' 2026 Anxiety: When Models Devour Everything, What Moat Is Left for Startups?
In 2026, a wave of investor anxiety questions the defensibility of AI startups as models improve, fearing that most companies are just "thin wrappers" destined to be absorbed by foundation models or chipmakers. The author argues against this despair, positing that true moats lie not in benchmark performance but in areas models cannot easily reach.
The logic of despair is that if models excel at all measurable tasks, only compute and cutting-edge model weights hold lasting value. However, the essay contends that the most valuable work is inherently "untrainable." Benchmarks measure what can be measured and thus optimized for, but real-world correctness often resides in private, complex systems. Examples include legacy codebases, intricate legal transactions, or hospital workflows. This kind of correctness is proprietary, costly to establish, and cannot be validated quickly—it requires time and trust within an organization.
As models commodify visible, measurable tasks from both above (labs absorbing scaffolding) and below (saturation by cheaper models), value shifts to "untrainable ground." This encompasses work where correctness is a private truth, locked behind integration barriers, licenses, liability frameworks, and entrenched user habits. Trust and adoption are slow, human-centric processes that smarter models cannot accelerate.
Successful companies defend their position by embedding deeply into client operations, owning the definition of "good" within a specific domain (e.g., Harvey in law, OpenEvidence in medicine), and pricing on outcomes rather than tokens. While labs compete fiercely, they are incentivized to keep the application layer vibrant. The future belongs not to those competing on generic benchmarks but to those navigating unscoreable terrain, doing the "unsexy work" of translation between models and messy human realities. The most cited benchmark scores are thus maps of territory about to become worthless, signaling who will lose the right to define what counts as good.
marsbit06/11 03:34
