Introduction

The scope of this research is to perform a comprehensive audit of conversational plans and systemic documentation related to upgrading and repurposing a local PC into a dedicated AI server. The methodology leverages Deep Research protocols to triangulate technical feasibility, architectural risks, and systemic limitations described in the provided internal files (future-server transcripts). The primary assumption is that the hardware (Ryzen 5 3600, 64GB RAM, RTX 5060 Ti, GTX 1060) remains fixed, and the objective is maximizing AI throughput, security, and remote accessibility without incurring unnecessary cloud costs.

Main Analysis

Finding 1: Fedora 42 Headless is Superior to Fedora 43 Upgrade

Upgrading the current infrastructure to Fedora 43 presents catastrophic risks to the hybrid GPU topology. The transition to a Wayland-exclusive environment completely breaks the GBM buffer sharing between the proprietary NVIDIA driver (RTX 5060 Ti) and the open-source Mesa driver (GTX 1060) [1]. Furthermore, the glibc 2.41 exception specification mismatch in Fedora 43 inherently breaks CUDA 13 compilation, severely disrupting the local AI toolchain [1]. Transitioning the existing Fedora 42 installation to a headless target (multi-user.target) retains stability, preserves existing Btrfs/Snapper configurations, and bypasses the graphical display issues entirely [3].

Finding 2: Alpine Linux vs. Fedora for Blackwell Compute

While Alpine Linux offers a lighter, systemd-free alternative, it requires manual compilation and deployment of the proprietary NVIDIA open-kernel modules (570/575+) and CUDA libraries, which are not natively supported in Alpine's repos [3]. For the RTX 5060 Ti (GB206), which exhibits volatile PCIe negotiation issues on unsupported kernels, maintaining the stable RPMFusion driver ecosystem on Fedora 42 provides a substantially lower administrative burden [1][3]. The systemd "age verification" concern is functionally irrelevant in a headless container host [3].

Finding 3: AI Workload Viability within 16GB VRAM Constraints

The 16GB VRAM provided by the RTX 5060 Ti is optimal for quantized local models (e.g., Qwen 2.5 Coder 14B Q4, Ministral 8B) operating at 40-70 tokens per second [2]. Deploying OpenClaw with a hybrid routing strategy—handling daily tasks (coding, file ops) via local models while routing complex reasoning tasks to OpenRouter (DeepSeek-R1, Claude 3.5)—is a mathematically sound approach to reduce API expenditure by 70-80% [2].

Finding 4: Security and Containerized Isolation

Exposing the host system to local AI agents poses a high security risk. The proposed mitigation strategy—running OpenClaw via rootless Podman with strictly mounted Btrfs subvolumes (/sandbox) and explicitly denied system capabilities (rm, chmod)—provides adequate isolation [2]. Utilizing a dedicated partition for the sandbox ensures that any rogue agent activity remains contained and cannot corrupt the host OS or /home directory [2]. Remote management via Cockpit and Tailscale further reduces the attack surface by eliminating the need for exposed external ports [3].

Synthesis & Insights

The transcripts highlight a convergence toward an "infrastructure-as-code" homelab approach, where the workstation ceases to be a desktop and operates as a unified API endpoint for laptop clients. By preserving Fedora 42 and leveraging its native Podman + Cockpit ecosystem alongside Tailscale, the architecture achieves cloud-like resilience (Btrfs snapshots) and GPU access without the overhead or latency of a true hypervisor (like Proxmox) or a completely unsupported OS (Alpine). The hybrid AI model routing acts as a cost-effective bridge between localized data privacy and cloud-based reasoning capabilities.

Limitations & Caveats

• The Ryzen 5 3600 CPU may become a bottleneck during heavy container orchestration or data preprocessing, despite the 64GB RAM upgrade [3].
• The RTX 5060 Ti Blackwell architecture currently exhibits known VRAM-pressure kernel panics on newer Linux kernels; strict adherence to the stable 580.65.06 driver on kernel 6.12/6.17 is mandated [1].

Recommendations

1. Execute Headless Conversion: Immediately disable the GUI (systemctl set-default multi-user.target) on Fedora 42 to reclaim resources for AI workloads.
2. Defer Fedora 43 Upgrade: Remain on Fedora 42 until the Wayland hybrid-driver buffer sharing limitations are upstream patched and glibc 2.41 CUDA incompatibility is formally addressed by NVIDIA.
3. Implement Hybrid Model Routing: Configure OpenClaw with RouterSmith or ClawRouter to default to local Qwen 14B (Ollama backend) and conditionally escalate to OpenRouter.
4. Enforce Sandbox Architecture: Ensure all AI containers run via Podman with restricted capabilities and are mapped only to a discrete Btrfs subvolume (/sandbox).