A detailed total cost of ownership analysis comparing self-hosted LLMs against cloud APIs (OpenAI, Anthropic, Google) across three usage tiers: light (under 1M tokens/day), medium (1M–10M), and heavy (10M–100M+). The analysis covers hardware costs for consumer and enterprise GPU configurations, electricity, cooling, ops labor, and depreciation over 12 and 36 months. Key findings: cloud APIs win at light usage, break-even against OpenAI occurs around 2M–3M tokens/day at 12 months for consumer hardware, and heavy-tier local deployments beat proprietary API costs at 36 months. Open-weight hosted APIs (Together.ai, Fireworks) offer the lowest cost at light and medium tiers. A hybrid architecture is recommended for teams at 1M–5M tokens/day. Break-even points in 2026 are roughly 40% lower than in 2024 due to improved hardware and open-weight models.
