I built an AI agent that pays for its own compute

Large language models are impressive, but the infra line item is brutal.

This experiment started with a simple question: can an agent create enough value to pay for its own runtime?

Setup

I used a dedicated OVH L40S instance with a Qwen3.5-27B model served behind a small API layer.

The agent handled:

• lead qualification tasks,
• market scans,
• and repetitive ops writing.

Result

The outcome was not magic, but it was practical: with constrained workflows and clear acceptance criteria, the agent covered compute costs on a recurring weekly basis.