Access, Not Ownership: An Orthogonality Theorem for AI Governance Regimes

AI-governance debates often contrast public ownership of foundation-model infrastructure with private oligopoly. This paper develops a conditional alternative centered on a three-component access vector μ = (ω, π, ν): weight openness, compute portability, and verification unbundling. Under an ownership-invariant regulatory implementation, ownership type θ becomes welfare-second-order at fixed μ.
Five results develop the claim. T1 characterizes ownership-invariant regimes. T2 bounds the access-monopolization component of inequality. T3 derives the local effective-scale exponent Λeff = δ(1−ω) + (β+γ)r(ω,π,ν) and its stability condition. T4 shows that, under stated certification assumptions, credence-good rents can remain positive even with open weights and compute; a welfare gain further requires positive deadweight incidence. T5 identifies commitment and financing conditions that preserve the static result under capture.
A 2×2 case grid establishes institutional feasibility but does not identify the proposed mechanism against six rival explanations. Four prospective tests specify evidence that would reject the access–welfare bridge, fixed-access ownership irrelevance, the credence-good verification channel, or the estimated stability boundary. A companion Lean 4 formalization checks selected dependencies under stated assumptions; it does not establish empirical premises. The result is a falsifiable conditional theory rather than an empirically confirmed law.