Self-hosting an LLM pays off only above about 12% GPU load
AIOpenRouter says a self-hosted model on a rented RTX 3090 breaks even with the API at about 12% load for Qwen3.8 27B, but a DeepSeek V4.1 Flash setup on four rented H200s never beats the cheapest zero-data-retention endpoint. The break-even divides GPU rent per hour by the API cost of the tokens that GPU serves in the same hour.
Source details
OpenRouter Blog · openrouter.aiPublished · added here
Original title: Self-Hosted LLM vs API: When Does Self-Hosting Pay Off