Why modern artificial‑intelligence systems exhibit chronic financial instability despite unprecedented investment. The paper models the AI ecosystem as two orthogonal economic regimes: (1) a capital‑intensive, burst‑driven training regime requiring hyperscale infrastructure, steady‑state energy draw, and multi‑billion‑dollar depreciation cycles; and (2) a low‑margin, intermittent inference regime characterized by stochastic demand, low utilization (typically 5–15%), and limited revenue elasticity. When vertically integrated, these regimes form a structurally disequilibrated system in which fixed costs from training overwhelm the throughput economics of inference, producing negative‑return scaling even as total capacity increases. The title, "The Economics of AI — Sailing the Insolvent Seas," is a cheeky nod to Monty Python's classic "Accountancy Shanty" from The Meaning of Life—where accountants-turned-pirates gleefully "sail the wide accountancy" to skirt the shoals of bankruptcy—this paper analyzes the precarious economics of artificial intelligence. The analysis introduces a geometric framework for understanding this imbalance, drawing from utilization asymmetry, power and cooling irreversibility, reflexive capital expansion, and constraint topology. Training is modeled as a high‑fixed‑cost vector with long cycles, while inference is modeled as a demand‑elastic curve constrained by thermodynamic and synchronization ceilings. Their integration yields an unstable manifold that cannot be equilibrated through scale alone. The paper argues that sustainable equilibrium requires bifurcation: training consolidating into utility‑like providers optimized for depreciation, yield, and throughput; inference maturing into a competitive service economy optimized for latency, cost per query, and user‑facing value. Historical analogs — particularly the semiconductor transition from vertically integrated IDMs to the fabless‑foundry model — reinforce the structural inevitability of this shift. The work is conceptual rather than predictive, offering primitives for analyzing AI economic cycles, negative‑return scaling, and the transition from expansion to consolidation. This paper is part of an emerging structural framework for understanding AI through the lenses of physics, economics, systems theory, and thermodynamic constraint.