Generative AI reaches users as a remote information service, but inference is produced in data centers tied to local electricity systems and policy-constrained accelerators. This paper asks how advanced-chip controls reshape the electricity burden and production geography of AI inference. In a two-location framework with tradable token services, local power-market feedback, and asymmetric chip frictions, an import restriction raises unit electricity use when the domestic substitute is cheap relative to its energy disadvantage. Domestic chip subsidies strengthen this channel, while power-system responsiveness determines where expanding demand is served. Public evidence on serving locations, model performance, chip systems, data-center load, electricity systems, application programming interface prices, and platform use grounds the numerical illustration. The results show that AI governance and digital infrastructure policy need to treat chips, data-center siting, and power-system expansion as connected policy margins.