Hardware is the bottleneck. AI infrastructure demand has driven up prices and constrained supply for the same gear homelabbers depend on, GPUs, NVMe drives, and dense compute nodes. Techno Tim and Adam Stacoviak spend this episode mapping the real cost of that scarcity against a software ecosystem that has never been more capable. Tim calls 2026 the Year of Self-Hosted Software, and the stack backing that claim is specific: Proxmox VE for virtualization, TrueNAS with ZFS special vdevs on discontinued Intel Optane drives, an RTX 3090 pulling double duty for Plex transcoding and local LLM inference via Ollama, and Paperless-NGX extended with Paperless-GPT for AI-assisted document tagging.
The more interesting thread is how Claude and Model Context Protocol are being used to automate infrastructure that was previously hand-configured. Adam built PXM, a custom Proxmox automation CLI, and DNSHole, a Pi-hole replacement written in Rust. Claude is being pointed at a UDM Pro via the Ubiquiti API to handle network tasks that would otherwise require navigating the UniFi UI. The episode also covers IBM's Docling for document parsing in RAG pipelines, PaddleOCR for multi-language OCR, and a Medallion Architecture approach applied to home document ETL. These are not toy projects.
The full conversation is worth your time not for the conclusions but for the architecture decisions underneath them: why ZFS special vdevs matter for metadata-heavy workloads, how PCIe bifurcation expands NVMe capacity without adding slots, and where Tailscale fits against a self-hosted Authelia and Bitwarden stack. If you run any serious homelab infrastructure, the specific tool choices and the reasoning behind them will give you something to act on.
[READ ORIGINAL →]