You gotta build a "digital twin" of the mess you're actually going to deploy into, especially with stuff like mcp (model context protocol) where ai agents are talking to data sources in real-time.
Tom Fenton reports running Ollama on a Windows 11 laptop with an older eGPU (NVIDIA Quadro P2200) connected via Thunderbolt dramatically outperforms both CPU-only native Windows and VM-based ...