Digital computation enforces stability physically, by thresholded restoration at every gate. Analog and wave computation executes directly in its substrate, so stability must instead be supplied informationally, through acquisition, identification and correction. We treat that obligation as a stability contract and establish its resource requirements. For independent Gaussian drift, holding n channels at a fixed mean-square distortion below the innovation variance requires feedback at order n bits per cycle, so a modality whose useful feedback capacity stays bounded fails beyond a computable channel count, whereas component-resolved observation delivers that information in one parallel acquisition. Information rate is not sequential depth. Maintenance separates into acquisition depth, feedback information and traffic, and host arithmetic and state. Within the declared class, a projected heavy-ball law driven by measured residuals achieves constant acquisition depth together with the linear information and arithmetic floors, at order-one auxiliary state per channel, reaching tolerance in four to five acquisitions on the tested monitor families. A routing-closure condition follows: a physical accelerator stays in the workload’s scaling class only when maintenance traffic does not cross the host boundary at a higher asymptotic rate than the workload interface. Evidence spans photonic, in-memory, large-scale simulation and quantum loop-closure studies.