Self-Hosting AI Models After Claude's Usage Limits

The author, writing about switching from Claude Max's weekly-limit change, concludes that self-hosting large coding models is not economically worth it today, and that paying per token is the saner choice. They recommend Claude Code as the main terminal tool, opencode with Qwen 3 Coder hosted on Alibaba as an alternative for simpler tasks, and Gemini for debugging. They report that an 8x H200 rented rig on DataCrunch costs $15/h, roughly $11,000 a month if run 24/7, while Alibaba was slightly faster than their optimized rig in their tests. The post also includes a setup guide for running Qwen3-Coder-480B on Vast.ai.