Allocate manufacturing impact without confusing it with avoidance

An embodied estimate should say which components and stages it covers: accelerator, memory, host, packaging, transport, and any included facility infrastructure or end-of-life treatment. Supply-chain energy mix, fabrication assumptions, and allocation among co-products can affect the result. A device label alone does not establish an inventory. Do not combine one supplier’s accelerator-only figure with another system’s full-node total and treat their difference as an efficiency ranking.

Amortization assigns a finite inventory across an assumed service lifetime. A calendar-capacity policy can spread it across all available capacity-hours, including idle capacity that needs a residual allocation. A productive-capacity policy assigns the inventory only across useful occupied capacity-hours, increasing the charge per used hour when utilization is low. Neither arithmetic choice changes the physical manufacturing that already happened. Choose a policy that fits the accounting question and reconcile its assigned and unassigned portions.

For a single asset under a simple policy, productive allocation capacity can be modeled as utilization times calendar service capacity. Lower assumed utilization raises the embodied allocation per productive hour. Doubling a presumed lifetime halves the assigned hourly amount, but that presumption needs a plausible support and reuse plan. A retrospective account can use observed service; a prospective decision needs sensitivity ranges because retirement date, demand, failure, and secondary use remain uncertain.

Hardware reuse can extend service, while replacement can reduce operating energy for matched work. The tradeoff depends on remaining lifetime, the actual workload response, grid assumptions, and the additional inventory caused by replacement. Avoid counting an old asset’s already-incurred manufacturing twice as a future saving. If disposal, resale, or another displaced purchase matters, model those alternatives explicitly rather than silently assuming that every reused machine displaces an identical new one.

A simplified replacement comparison can divide additional upfront embodied impact by an annual operating-impact saving to obtain a break-even duration. That duration is meaningful only if the saving is positive, service is matched, factors remain as assumed, and the equipment remains in use long enough. If future electricity factors or utilization change, integrate the time-varying difference instead. A nominal payback longer than the planning horizon is not evidence of a benefit within that horizon.

Report separate operational and embodied components before giving a combined figure. Include uncertainty ranges for inventory coverage, lifetime, utilization, and any avoided-purchase assumption. A per-job allocation is useful for budgets and comparisons under consistent conventions, but canceling one job does not automatically undo its allocated manufacturing share. Likewise, a lower allocated number from a longer spreadsheet lifetime does not itself establish an environmental improvement. The physical decision and the allocation policy require separate evidence.