A 2026 total cost of ownership (TCO) analysis compares local LLMs against cloud APIs across three usage tiers for solo developers, startups, and mid-size engineering teams. The article warns that comparing per-token price alone is a trap, as full TCO must account for electricity, operations labor, hardware depreciation, rate-limit engineering, prompt re-engineering during vendor migrations, and downtime costs. All hardware prices reflect MSRP or estimated street prices as of June 2025, including the RTX 5090, Apple M4 Ultra, and AMD MI325X. API rates come from published rate cards from OpenAI, Anthropic, and Google as of the same date. Power consumption estimates derive from manufacturer TDP specifications with typical system overhead. Cloud API cost projections exclude taxes and network egress costs. The analysis builds complete 12-month and 36-month TCO models across three usage tiers. The light usage tier covers under 1M tokens per day, with solo developers and side projects rarely pushing past 500K tokens per day, and many workloads hovering closer to 100K. Usage is bursty, with a few intense hours of inference followed by long idle stretches. When utilization stays low, fixed infrastructure costs dominate. The article includes head-to-head TCO comparison tables, an interactive cost calculator, and a decision framework. It notes that 2026 break-even points are 40% lower than 2024.