Efficiency compounds when the feedback loop is owned end to end

The source frames efficiency as a stack-wide loop rather than a one-time model gain. Routing, scheduling, cache behavior, kernels, and repeated tool work are measured together so that a local improvement does not hide a larger cost elsewhere.

An efficiency claim becomes investable when the feedback loop has a clear owner and can be repeated. Cost reduction is more defensible when it also improves capacity, latency, and reliability instead of borrowing performance from one part of the system.