From Signal to Substrate: Architecture, Irreversibility, and the Capital Economics of Custom Silicon

Hardware partitioning is commonly treated as an engineering decision: identify a computational bottleneck, estimate the performance or energy benefit of specialization, and determine whether the function belongs in software, programmable logic, or fixed silicon. That framing is incomplete. Moving computation toward dedicated hardware converts flexibility into commitment. Algorithms, dataflows, security assumptions, workload forecasts, interfaces, manufacturing dependencies, and supplier choices begin accumulating non-recurring engineering cost, verification effort, package constraints, capacity reservations, inventory exposure, and lifecycle obligations. Architecture is therefore also a capital-allocation discipline. This paper proposes a framework for deciding what deserves to become hardware by examining four questions: whether specialization creates measured technical advantage; whether the underlying computation is sufficiently persistent; whether risk-adjusted lifetime economics justify greater irreversibility; and whether the resulting implementation is manufacturable, supplyable, trustworthy, and deployable. The resulting question is not simply what can become silicon. It is what deserves to become irreversible.