I tested three AI tools on my GB10. The image generator added servers that do not exist, the vision model misread an IP address, and the model runtime kept 42 GiB of memory reserved while it was idle.
I set up Open WebUI on my GB10 with NVIDIA AI Workbench. A capital letter stopped Compose, renaming the project orphaned a volume, and the change nearly downloaded 19 GB of model weights again.
I connected my local 120B model to five read-only VMware and Veeam APIs. It found that 52 of 63 VMs had no restore point, even though the existing checks reported no unprotected objects.
I moved my Hermes Agent from a laptop to an NVIDIA GB10. This gave me an always-on 120B model, reliable scheduled jobs and remote access through Telegram and Tailscale.
Building a fully on-premises AI assistant that queries VMware vCenter, VCF Operations, and VCF Networks using a local LLM — zero cloud dependency, complete data sovereignty.
A working implementation of Microsoft Copilot Studio as a conversational AI interface for VMware vCenter — turning natural language questions into live infrastructure queries.
