A General Theory of Intelligence Infrastructure: The Semantic-to-Physical Continuum and the Dynamic Infrastructure Control Law

Contemporary artificial intelligence strategy evaluates capability almost exclusively through single-node expansion—parameter counts, training compute, and dataset scale. This paper argues that such an approach constitutes a local optimization within a fixed infrastructure regime. We propose a General Theory of Intelligence Infrastructure that treats intelligence as an emergent, closed-loop property of a continuous Semantic-to-Physical Continuum, spanning semantic observation, logical inference, edge processing, analog conversion, physical actuation, and material-thermodynamic substrates. We formalize a six-layer architecture—the L0–L5 Functional Intelligence Stack—defined by operational boundary conditions rather than technological taxonomy. Within this architecture, we establish the Intelligence Bottleneck Principle: effective civilizational intelligence I_effective is governed by a production function subject to a non-linear threshold penalty Phi(B), where the systemic bottleneck ratio B(t) = min_i (L_i(t) / L_i*(t)) strictly bounds performance. Asymmetric investment in non-binding layers yields vanishing marginal returns once any required functional boundary collapses. The static principle is extended into a continuous-time dynamic control framework that governs Dynamic Bottleneck Migration, Lock-In Phase Transitions, and State-Based Closed-Loop Allocation. Capital allocation is structured as a five-level decision hierarchy: Where -> Why -> How Much -> How Sharply -> When to Adapt. Here "How Sharply" is operationalized by an investment temperature parameter tau(t) (with targeting sensitivity gamma(t) = 1/tau(t)) that continuously interpolates between hard bottleneck focus and diffuse investment; "When to Adapt" is realized through state feedback on ratio spread S(t), mean ratio deficit R(t), and inter-layer gap G(t). Numerical trajectories demonstrate that soft, state-responsive allocation systematically outperforms uniform policies under differential threshold growth. Finally, the framework is applied to Japan. Genten Kaiki (原点回帰) is reinterpreted as a strategic re-orientation toward the un-bypassable physical and semantic boundary conditions that ultimately govern scalable intelligence.