The AI Capacity Provider Standard

Organizations that consume AI capacity from more than one source face an immediate and largely unexamined measurement problem: the units in which that capacity is metered — tokens, requests, GPU-hours, invocations, retrieval calls — are defined by their issuers, priced by their issuers, and measured by mechanisms their issuers control. Because these meters look numeric, they invite arithmetic. Because they carry the same names across providers, they invite comparison. This paper argues that neither invitation should be accepted uncritically. AI capacity is not fungible merely because providers meter it. A meter reading stripped of its workload context, measurement method, capability boundary, rights terms, topology, threat evidence, uncertainty, and provenance is not a comparable quantity; it is a number wearing a unit's clothing. This paper proposes an AI Capacity Provider (ACP) standard: a definitional boundary for what counts as a governed supply endpoint of AI-relevant capacity, a taxonomy of seven capacity classes, and a fourteen-dimension minimum declaration that preserves the facts a defensible comparison requires. The standard's central mechanism — a versioned Capacity Work Package, introduced in Section 3.16 and developed in Section 4 — carries those declarations through acquisition, comparison, and consumption so that heterogeneous resources can be compared without manufacturing false equivalence. The counter-thesis deserves equal statement at the outset, not deferral to a rebuttal section. For a bounded case — one provider, one workload family, one contract — raw-unit comparison, direct routing, or ordinary contract review may reach the same defensible decision at lower burden. The proposed standard has not earned its overhead for such a case unless its additional fields change a decision, preserve a material audit fact, expose drift, or reduce material uncertainty. This paper therefore proposes the standard as a falsifiable claim with named evidence burdens, not as settled doctrine, and it states throughout where the standard should lose.