Sovereign Agency without the Capex Supercycle, Volume II: An Anatomical Review and Augmentation of Capital-Light Pathways to Sovereign AI Ecosystems for the Global South

Volume I of this series argued that sovereign AI is a question of agency across the computational value chain, constrained above all by power, cooling and capital (Mamun, 2026a). This second volume asks the question that follows for most of the world: what can a state do now, before it can afford a national AI cluster? It reviews the "Volume II" section of the same AI-generated report, which proposes capital-light shortcuts: small language models, parameter-efficient fine-tuning with LoRA and DoRA, low-communication distributed training (DiLoCo, NoLoCo, DeMo), decentralised serving (PlanetServe, KVCompass), India's digital public infrastructure, and the Atlas Global South AI Policy Framework. Following the KSCCN practice of anatomical review, each claim was checked against the primary paper or institutional source. Most of the techniques are real and correctly described in outline. Several details were corrected. NoLoCo's 4 per cent figure refers to final perplexity, and its speed claim is up to four times faster wall-clock convergence. KVCompass serves one provider's geo-distributed regions, not volunteer edges. The Atlas framework's "Sovereign Compute Allocation" is the 60 per cent share of a proposed 1 per cent levy, not a tax of its own. One claimed method could not be found and was removed, as were untraceable capital figures. The augmentation adds computed informatics on memory, trainable parameters and communication time, a visual derivation of DoRA, schematic panels for digital public infrastructure and the proposed fiscal loop, and a five-rung capital-light ladder, with implications for Bangladesh. Volume II; continues Volume I, doi:10.5281/zenodo.23669609.