The struggle over Large Language Models is not a technical debate; it is a political war over who governs the infrastructure of meaning and for whose benefit. Your next query delivers a seamless answer—a metabolic subsidy for your nervous system. The bill is paid elsewhere: as PTSD in a Nairobi content moderator, as carbon debt in a heating sky, as the quiet erasure of a thousand ways of knowing that could not be scraped. The interface glows softly, apologizing for the war it conceals. Every token cuts both ways. This paper analyzes this metabolic discount through the lens of Derrida’s pharmakon—a substance intrinsically both remedy and poison. It argues that the discount’s emancipatory potential is inseparable from a metabolic shadow of exploited labor and ecological cost, and that its design under platform capitalism structurally favors cognitive capture. We examine how Reinforcement Learning from Human Feedback (RLHF) enacts epistemic closure, how the recursive contamination of training data threatens irreversible epistemicide, and how the dynamics of platform capitalism push the pharmakon toward its poison-face. Distinguishing between scaffold deployment and infrastructure capture, the paper proposes the principle of “the body votes last” as a navigational practice and identifies endogenous feedback loops through which organized resistance could contest the default trajectory toward capture.