Notebook
Blog
Personal notes on security, local AI, sovereignty and building AgentVUE.
-
40× RTX 5070 Ti: the rig behind AgentVUE
The current rig: 40 consumer GPUs, 5.5k+ tokens/s prompt processing and 800 tokens/s for 8 agents in parallel.
-
The final monster: 32 GPUs and 1 TB of VRAM
32× AMD MI50, one terabyte of VRAM to run Kimi K2.6 locally. And why I said: never again.
-
Scaling to 16 GPUs: DeepSeek v3.2 and Qwen3.5 at home
512 GB of VRAM, ~€5.6k: DeepSeek v3.2 then Qwen3.5 397B running at home, with the real numbers.
-
My first local AI rig: 8 GPUs, 256 GB of VRAM, $880
How 8 second-hand AMD GPUs became one of the most cost-effective local inference rigs in the world.
-
What is the origin story of AgentVUE?
The personal story of how AgentVUE began: origins, systemic risk and a philosophy of vigilance.