The prevailing assumption in foundation-model robotics is that once general manipulation is solved, the robotics market is solved. But solving general manipulation is not solving the market. Generalism on its own does not capture the edge.

Work in the physical world requires managing a fundamental trade-off between speed and fidelity while acting with incomplete information. To act swiftly, we default to low-resolution representations of our environment. But where performance, safety, and precision truly matter, we cannot remain at low resolution. We are forced to zoom in, increasing the fidelity of our internal model until it matches the exact constraints of the physical operation.

Computing already ran this experiment. General-purpose CPUs were sufficient for broad, low-resolution workloads like sending an email. But when raw throughput and energy efficiency became non-negotiable, computing did not double down on generalism — it shifted to specialized silicon accelerators designed for a high-resolution fit.

Robotics will follow the same pattern. Foundation models will set an impressive floor for basic manipulation. But where line uptime, millimeter accuracy, and cycle times dictate profit, a generalized model cannot outcompete a domain-optimized fit.