OpenRouter analyzes cost breakeven points for self-hosting open-source LLMs

On November 12, 2026, OpenRouter published an analysis titled "Self-Hosted LLM vs API: When Does Self-Hosting Pay Off," examining the cost break-even point between running open-weight models on your own GPUs and paying for API access. The piece frames the fundamental difference between the two options: GPU rentals are billed hourly whether the hardware is busy or idle, whereas API providers charge only for the tokens you actually consume. That means self-hosting only becomes cheaper when your traffic keeps the GPU busy enough of the time to offset the fixed hourly rent. Working from prices as of October 8, 2026, the analysis arrives at two concrete scenarios. A 27B model running on a single rented RTX 3090 breaks even at roughly 12% utilization — below that threshold the API bill is lower, above it the GPU rental pays for itself. A large mixture-of-experts model running on rented H200s, however, never undercuts the cheapest zero-data-retention (ZDR) API endpoint under the same pricing, regardless of load. The broader takeaway is that smaller models can make self-hosting worthwhile once they hit a certain GPU utilization level, while large MoE models typically cannot overcome the price advantage of API providers. No follow-up reporting has surfaced since the original post, and the underlying GPU rental rates and API pricing tables that feed the calculations were not fully reproduced in the public summary. The November 12 release remains the latest available material on this topic.