Institutional computing resources are unevenly distributed, and commercial cloud costs create a structural barrier for independent researchers, small manufacturers, and individual practitioners who nonetheless require capable compute for AI workloads, knowledge-graph systems, and persistent services. We present Lala City, a 10-node heterogeneous compute cluster built on a legacy-hardware-first principle: every compute node is either a previously-owned device repurposed into a server role or a used machine acquired on the secondhand market, while three marginal additions follow a cheapest-viable-add-on rule (a Raspberry Pi 5 for the home-server role, a Google Coral USB Accelerator for edge ML inference, and a secondhand ARM64 SBC). Measured recurring subscription fees to cloud providers, SaaS vendors, and network services total approximately INR 300/month (USD 3.60/month); the marginal capital outlay to build this cluster from the author's existing hardware pool is approximately USD 325, against a historical acquisition cost across all prior owners and uses in excess of USD 13,000. Nodes are stitched together by a self-hosted Headscale (v0.28) control plane, published selectively via Cloudflare Tunnels, and exposed as a federation of seven Model Context Protocol (MCP) filesystem servers. The architecture is deliberately minimal: no Kubernetes control plane, no commercial observability stack, no paid managed services, no new enterprise hardware. We describe the topology, hardware-acquisition strategy, authentication model, service mesh, and practical gotchas encountered over approximately 14 months of continuous operation, and argue that this design is a viable sovereignty-and-sustainability-first pattern for personal and small-organization compute. We explicitly do not account for electricity, hardware amortization, or operator labor, none of which have been instrumented. We release a complete reference configuration at github.com/Vibhav-Aggarwal/lala-net-public.
Vibhav Aggarwal (2026) studied this question.