The Local AI Stack: Building a Sovereign Machine Learning Workstation with Hyper-V, WSL2, and GPU Virtualization

The Local AI Stack: Building a Sovereign Machine Learning Workstation with Hyper-V, WSL2, and GPU Virtualization

Renting GPU time from AWS costs more than the GPU. Every prompt you send to a cloud API is a prompt someone else logs. If your ML work runs on hardware you don't own, on a network you don't control, then you don't own your ML work.

This is the 140-page engineering reference for building a sovereign, high-performance local machine learning workstation on Windows — CUDA-accelerated, virtualized, and fully under your control. From silicon to CUDA kernel, one stack, one machine, one operator.

What's inside

The book walks you through the full virtualization stack from bare metal to running CUDA workloads:

Hardware prerequisites: SLAT-capable CPUs, NVIDIA Tensor Core GPUs, IOMMU support

BIOS/UEFI configuration for VT-x/AMD-V and VT-d/AMD-Vi virtualization extensions

Hyper-V feature activation via Windows Features and PowerShell Enable-WindowsOptionalFeature

Virtual Machine Platform and WSL2 optional component installation

Linux kernel update package deployment and WSL2 as global default

Ubuntu 22.04 LTS deployment via Microsoft Store or CLI

.wslconfig tuning for CPU cores, RAM limits, and host OS protection

Custom swap file placement on high-speed NVMe for virtualization paging

NVIDIA Windows GPU Driver (Game Ready and Studio) for GPU Paravirtualization

GPU Paravirtualization (GPU-PV) for shared host/VM acceleration

Discrete Device Assignment (DDA) for exclusive GPU passthrough on high-performance VMs

NVIDIA Container Toolkit installation inside WSL2

Docker Desktop with WSL2 backend, plus Podman and Rancher Desktop as open-source alternatives

NVIDIA runtime configuration in daemon.json for CUDA container access

Verifying GPU acceleration inside WSL2 with nvidia-smi

CUDA Toolkit installation with host driver version alignment

cuDNN library configuration for deep learning workloads

Persistent DrvFs mount points for cross-OS dataset access

I/O optimization: relocating training datasets to native EXT4 VHDX for maximum throughput

And more — 140 pages, densely technical, cover to cover

Who this is for

ML engineers building local training and inference infrastructure. AI researchers who want CUDA workloads on hardware they own, not compute they rent. Infrastructure specialists deploying sovereign AI stacks. Windows-based practitioners who refuse to give up their OS to run serious ML work.

You should be comfortable at the command line, familiar with Linux fundamentals, and know your way around BIOS/UEFI settings. This is a technical reference, not a click-by-click beginner tutorial.

Format

Delivered as PDF and TXT. Read the PDF cover to cover, or feed the TXT straight into your own local RAG, grep it, index it, pipe it into your local LLM. The book about local AI infrastructure, ready to be consumed by locally-run AI.

Part of the Sovereign product line

Sits alongside Sovereign Intelligence (local AI blueprint), The Quantization Black Book (4-bit compression at scale), and The Dark Mesh (sovereign off-grid networking). Different layers of the same philosophy: hardware you own, models you run, networks you control.

A note on responsibility

This material is provided as an engineering reference and architectural roadmap, curated from official Microsoft, NVIDIA, and Linux documentation and current industry best practices. Configurations should be adapted to your specific hardware. You are responsible for complying with the licenses of the tools you deploy, and with the laws of your jurisdiction. Build carefully and build for the right reasons.