Here is the frontier, and it is not energy. A generative workload amortises its input cost: load the model once and draw many samples. A closed-loop controller does not. Every control tick it imposes fresh sensor state on the fabric as new boundary conditions and reads an action back. So the governing quantity for embodied use is conditioning bandwidth, how fast new boundary conditions can be imposed and results read. Two operations are easy to read as one here, and the distinction sets the number. Restreaming loads a whole model onto a fabric and runs on the order of a problem per second. Clamping imposes new boundary conditions on a model already resident and is much faster. A controller does the second. The optimisation community has described the same structure from its own side, naming model transfer latency, per-problem parameter tuning, and the fact that the next problem cannot be determined until the current one is solved, which is a closed loop in other words. And they have moved it: thirty-two-fold coupling compression and a learned estimator in place of tuning brought end-to-end latency to about one hundred milliseconds, independent of problem size. Measured against a control period, one to two orders remain, and each is an interface and memory-hierarchy problem rather than a physics one.