Every piece of hardware on this page is something I’ve tested in my own homelab — running Proxmox, Docker, local LLMs, Immich, Frigate, and UniFi. No spec-sheet reviews. These are the picks that survived months of 24/7 uptime.
🚀 Minimum Viable Homelab (Under $700)
You don’t need a rack to start. Three items get you a fully functional homelab that runs Docker, Proxmox, Pi-hole, Home Assistant, media servers, and even lightweight local AI — all under $700.
💻 Mini PCs
Mini PCs are the backbone of a modern homelab. Silent, low-power (15-35W idle), and small enough to velcro behind a monitor. All three below run Proxmox, Docker, and Ubuntu Server without issues.
💾 NAS & Storage
Your homelab needs reliable storage. NAS for bulk data (photos, backups, media), NVMe for VMs and containers. Don’t cheap out on drives — a failed consumer drive at 2 AM is not fun.
🎮 GPUs for Local AI
If you’re running local LLMs, Stable Diffusion, or want fast Immich/Frigate ML processing, GPU VRAM is the bottleneck. More VRAM = bigger models. Here’s the real-world hierarchy — not the marketing one.
🍓 Single-Board Computers
Not everything needs a full mini PC. Single-board computers handle lightweight, always-on services where silence, size, and power draw matter. Pi-hole, Home Assistant, DNS, VPN endpoints — these are perfect for a Pi 5.
⚡ AI Accelerators
Dedicated AI accelerators handle specific ML tasks (object detection, classification) without tying up your CPU or GPU. Essential for Frigate NVR and Immich photo processing. They sit alongside your main hardware.
-
🪸
Google Coral USB Accelerator ~$60 — Best for Frigate/ImmichCheck price →
Plug-and-play Edge TPU. Handles 100+ fps object detection. The standard for Frigate NVR and Immich face/object ML. -
🔷
Hailo-8L M.2 AI Accelerator ~$75 — Coral AlternativeCheck price →
Drops into an M.2 A+E or M.2 B+M slot. 13 TOPS. Growing Frigate support. Good if you have a free M.2 slot and no USB ports. -
🚀
NVIDIA Jetson Orin Nano 8GB ~$500 — Serious Edge AICheck price →
40 TOPS, runs small LLMs and vision models natively. Overkill for Frigate, but perfect for custom edge AI pipelines or multiple camera streams with YOLO.
📡 Networking
Your homelab is only as good as your network. VLANs, proper APs, and a managed gateway make the difference between “it works” and “it works reliably for everyone in the house.” UniFi is the homelab standard for a reason.
🔋 Power & Memory
Two things that are easy to forget but critical: a UPS to survive power blips (especially if you’re running ZFS or databases), and enough RAM to actually use your hardware.
- ❌ Don’t skip the UPS. One dirty power spike will corrupt your ZFS pool or Proxmox config. A $180 UPS protects $1,000+ in hardware and days of setup time.
- ❌ Don’t run Proxmox on 8GB RAM. The host OS needs 2-3GB, each VM needs 2-4GB. 16GB is tight, 32GB is comfortable, 64GB is future-proof.
☁️ Cloud Overflow
Sometimes your homelab isn’t enough — you need a public-facing server, GPU burst compute, or just don’t want to maintain hardware for a specific workload. That’s where cloud makes sense alongside (not instead of) your homelab.
🔗 Related Tools & Guides
🎯 LLM Hardware Checker
Enter your GPU or Mac and see exactly which LLMs you can run locally, with quantization and speed estimates.
🧮 VRAM Calculator
Calculate exactly how much VRAM a model needs at different quantization levels. Match models to your GPU.
🤖 Local LLMs Guide
Complete guide to running AI models locally — Ollama, llama.cpp, model recommendations, and performance benchmarks.
📸 Self-Hosted Photos Guide
Immich vs PhotoPrism vs Synology Photos — with NAS sizing, Coral TPU setup, and migration paths from iCloud/Google.
Frequently Asked Questions
What is the best mini PC for a homelab in 2026?
The Beelink SER5 with AMD Ryzen 5 is the best budget pick at ~$300 — it handles Proxmox, Docker, and 10+ containers at 15W idle. For heavier workloads, the Minisforum UM790 Pro (Ryzen 9, DDR5) steps up. Intel NUC 13 Pro is best for dedicated Proxmox hosts with vPro.
How much should I spend on a homelab to start?
Under $700 gets you running. A Beelink SER5 ($300) + Samsung 990 EVO 1TB ($90) + APC UPS ($180) = ~$570. Add NAS and drives later when you actually need shared storage. Start small, expand based on what you use.
Should I use a GPU or CPU for local AI?
GPU, always. A $300 RTX 3060 12GB does 30-40 tok/s on 7-8B models; a Ryzen 9 CPU does 5-10 tok/s on the same model. GPU VRAM determines max model size: 12GB for 7-8B, 16GB for 14B, 24GB for 27-34B. CPU-only is too slow for regular use.
Is a NAS better than direct-attached storage?
NAS is better when multiple devices need shared access (photos, backups, media). DAS is simpler for single-server setups. Most homelabbers start with NVMe in their mini PC, then add a Synology NAS when they need shared storage or RAID redundancy.
Is UniFi networking worth it?
Yes, if you want VLANs, IoT isolation, and centralized management. A single U6+ AP ($100) is a massive upgrade over consumer routers. The Dream Machine SE ($500) adds IDS/IPS and camera support. If you just need better Wi-Fi without VLANs, any standalone Wi-Fi 6 AP works.
Do I need a UPS for my homelab?
Yes. A $180 APC Back-UPS Pro 1500 protects against power spikes and gives 15-20 minutes for clean shutdown. Without one, a power blip can corrupt ZFS pools and Proxmox configs. Use NUT via USB for automatic shutdown scripts. Cheapest insurance for your whole stack.