First Edition · July 2026

From Box to Cluster:
Building a Personal AI Supercomputer

with NVIDIA DGX Spark Bundle — from bare hardware to a production inference cluster running large language models on Kubernetes

8 Chapters
256 GB Combined GPU Memory
GB10 Blackwell GPUs
7 Stack Components
Front Matter

Preface

Why We Wrote This Book

In early 2026, NVIDIA released the DGX Spark — a personal AI supercomputer the size of a Mac mini, packing a Grace Blackwell GB10 GPU with 128GB of unified memory into a single quiet desktop unit. For the first time, running frontier-scale AI workloads from a home lab or a startup office became genuinely practical.

We bought two of them.

What followed was weeks of hard-won lessons: first-boot quirks, ARM64 compatibility landmines, networking configurations that broke SSH, container images that silently ran on the wrong architecture, and tensor parallelism tuning that only worked after we understood the interplay between NCCL, Ray, and the k3s networking layer. None of this was fully documented in a single place.

This book is the documentation we wish we had when we started.

What This Book Covers

  • Hardware setup — physical connections, first-boot wizard, static IP, SSH, Docker
  • System updates — OS, CUDA 13.0, and driver updates via the DGX Dashboard
  • Kubernetes — k3s, GPU Operator, Helm, namespace architecture
  • Distributed inference — KubeRay, cross-node Ray cluster, ARM64 gotchas
  • vLLM — tensor parallelism across both nodes, HuggingFace secrets
  • AIBrix — AI gateway for routing, multi-tenancy, agent lifecycle
  • Monitoring — Prometheus, Grafana, GPU metrics via DCGM Exporter

Who This Book Is For

This book assumes you are comfortable with a Linux terminal, have basic Kubernetes familiarity, and have purchased or are evaluating the NVIDIA DGX Spark Bundle. You do not need to be a DevOps engineer or CUDA expert — we explain every decision and command, including why certain approaches were tried and abandoned.

Note on ARM64
The DGX Spark runs on ARM64 (Grace CPU). Many popular container images — including the standard Ray image — are x86 only and will silently fail on ARM64. We call these out explicitly throughout the book and always provide the correct ARM64-native alternative.

Mohinish Shaikh & Sanwi Sarode, July 2026

Front Matter

Table of Contents

Part I — Node Setup

ChapterTitleKey Topics
1Introduction — What You Are BuildingArchitecture, software stack, namespace layout
2Hardware Setup and First BootPhysical setup, static IPs, SSH, Docker
3CUDA and System UpdatesDGX Dashboard, CUDA 13.0, driver updates
4Kubernetes Cluster with k3sk3s, worker join, Helm, GPU Operator

Part II — Model Serving

ChapterTitleKey Topics
5KubeRay for Distributed AIKubeRay operator, ARM64 image, RayCluster
6vLLM Inference EngineTensor parallelism, API verification, models
7AIBrix AI GatewayRouting, multi-tenancy, ModelAdapter
8Cluster Overview and MonitoringGrafana, GPU metrics, roadmap

Back Matter

SectionContents
Command CheatsheetQuick-reference kubectl, vLLM, port-forward, iptables commands
TroubleshootingSSH loss, GPU visibility, ARM64 errors, model download failures
FAQCommon questions about DGX Spark, models, and cluster setup
About the AuthorsMohinish Shaikh and Sanwi Sarode