Local vs Cloud Break-Even

Every solo builder and small lab eventually runs the same math: keep paying the cloud per token forever, or buy a GPU once and pay only the wall plug. The cloud is frictionless and elastic, but the bill never stops. Local hardware has a real upfront sting, and then the marginal cost collapses to electricity plus a slot of your desk. This calculator answers the only question that actually matters: how many months until the local box pays for itself, given your real usage and your real provider mix. As of June 2026, the going rate for frontier models is roughly Claude 3.5 Sonnet at $3 per million input tokens and $15 per million output, GPT-4o at $2.50 / $10, and Gemini 1.5 Pro in the $1.25–$5 / $5–$15 range depending on context length. Open-weight models on the local side — Llama 3.3 70B, Qwen 2.5 72B, DeepSeek-V3, Mistral Large 2 — now match or beat GPT-4 on most non-frontier tasks, and they run on a single RTX 4090 (quantized) or a pair of 3090s. A used 4090 currently runs about $1,600–$2,000; a new 5090 lands around $2,500–$3,000. The calculation here is deliberately narrow. It compares cash out the door — your current cloud bill versus the electricity needed to keep a single GPU under load. It does not credit your time, the quality gap on hard reasoning tasks, the cost of a UPS, or the convenience tax of having someone else handle scaling. Those matter, but they don't fit in a one-screen calculator. What this calculator is good at is killing the worst version of the argument — the one where "the cloud is so much cheaper" or "owning hardware is so much cheaper" gets asserted with no numbers. Plug in your last month's API invoice, pick a realistic GPU price and wattage, set electricity at your actual utility rate (the U.S. residential average is about 16¢/kWh in mid-2026, but it varies 9¢–35¢ by state), and read the months-to-break-even number. Anything under 12 months usually means local is worth it. Over 24 months and the cloud is probably still the right call.