Building a Homelab: Compact Guide with Concrete Recommendations
What this article covers
- How I would build a homelab at home: layer by layer.
- Affordable but optimized: no budget traps, no overkill spending.
- Concrete product recommendations with reasoning.
- What to watch out for to keep your setup future-proof.
Note: Product links are affiliate links (Affiliate disclosure): you pay nothing extra.
Introduction
A homelab doesn’t need a rack or server hardware. The most sensible foundation in 2026: a mini PC with plenty of RAM (for AI models and VMs), a managed switch with 10G uplinks (future-proofing), and a UPS (data survival). Everything else is optional.
This guide is intentionally compact: one recommendation per layer, one rationale, one budget.
Layer 1: Networking, the managed switch
Recommendation:
- Budget: MikroTik CSS610-8G-2S+IN (~170 €): 8× GbE + 2× 10G SFP+
- Alternative: Grandstream GWN7721: 8× GbE + 2× 10G SFP+, often cheaper
Why managed instead of unmanaged:
- VLANs: Separate IoT devices, guests, and servers, the most important security feature in a homelab.
- LACP/Link aggregation: Bundle multiple ports for higher throughput to your NAS.
- Monitoring: View port statistics, spot errors, invaluable when troubleshooting.
- 10G SFP+: The real reason for these switches: uplink to NAS or cluster without bottlenecks.
Future-proofing considerations:
- 10G SFP+ uplinks are mandatory, otherwise you’ll hit a bottleneck in two years.
- SFP+ instead of RJ45-10G: cheaper optics and DAC cables, lower power draw.
- PoE only if needed (access points, cameras), otherwise save the money.
Layer 2: Compute, the mini PC
My clear recommendation: MINIS FORUM MS-S1 MAX: AMD Ryzen AI Max+ 395, 128 GB RAM, 2 TB SSD (~1,900-2,200 €)
Why this specific model:
- 128 GB Unified Memory: The Ryzen AI Max+ 395 (Strix Halo) shares RAM between CPU and iGPU, letting you run large 120B-MoE models like gpt-oss-120b or Llama-4-Scout locally, models that don’t fit on standard PCs.
- Cluster-capable: Connect two MS-S1 Max units via 10G SFP+ for distributed inference or failover, scales without a new server.
- Quiet and efficient: ~30-80 W under load, versus 300 W+ for tower workstations. Can run in a living room.
- VM-ready: Install Proxmox, then create VMs for n8n, Ollama, chat platforms, all on one machine.
Budget alternative: A used mini PC (Intel NUC, Lenovo ThinkCentre) with 32-64 GB for ~300-600 €, see AI mini PC and Used server hardware. But for AI work, the VRAM is limited: the MS-S1 Max is the sweet spot for this use case.
Layer 3: Storage: NAS or DAS
Recommendation: 4-bay NAS (Synology/QNAP) or a custom mini-ITX build, see NAS basics. For getting started, though:
- 2× NVMe in the mini PC (2 TB built-in + expansion) for fast local storage.
- External USB or network drive for backups, more important than RAID.
Core principle: Backups beat RAID. Plan your backup strategy first, then think about RAID. See Backups.
Layer 4: Power, the UPS
Recommendation: Compact UPS 600-1000 VA (~80-150 €), e.g. APC Back-UPS or CyberPower: Search on Amazon.
Why: Power loss during a write corrupts your filesystem. The UPS buffers for 5-10 minutes, enough time for a clean shutdown via NUT integration (Network UPS Tools) with Proxmox or Linux.
Layer 5: The framework, optional
A 10” mini rack or a solid shelf. Cable ties, short patch cables, labels. Buy only when needed, not first.
Total budget
| Layer | Product | ~Price |
|---|---|---|
| Switch | MikroTik CSS610-8G-2S+IN | 170 € |
| Compute | MINIS FORUM MS-S1 Max (128 GB) | ~2,000 € |
| Storage | 2nd NVMe 2 TB | ~120 € |
| UPS | APC Back-UPS 700 VA | ~100 € |
| Total | ~2,400 € |
Budget version (used mini PC + cheaper switch): possible from ~600 €, see AI PC for beginners.
What runs on this
- Ollama + gpt-oss-120b / Qwen3-MoE locally: 128 GB Unified Memory handles large MoE models.
- Proxmox with VMs: n8n, Nextcloud, Vaultwarden, Chatwoot, all tools from the integration articles.
- Bot stack: Telegram bots, Mattermost, etc., all on the mini PC.
- Clustering later: second MS-S1 Max via 10G SFP+ for distributed inference.
Further reading
- IRC-Coding.de: Programming tutorials.
- Software stacks for the homelab: Which software to run: AI, small business, backups.
- AI hardware: Detailed hardware buying guides.
- MoE models: Why 128 GB RAM matters for these models.
- Docker: Deploy services.
- Proxmox: Virtualization on the mini PC.
- RackPeek: Document your homelab.
Key takeaways:
- Homelab foundation: managed switch with 10G uplink + mini PC with plenty of RAM + UPS.
- Managed switch for VLANs (security), LACP, and monitoring.
- MS-S1 Max: 128 GB Unified Memory = large MoE models locally, cluster-ready.
- Future-proof: 10G SFP+, not just GbE; RAM over everything else for AI.
- Backups before RAID; UPS before fancy racks.
FAQ
Why a managed switch?
Why the MS-S1 Max instead of a custom build?
Do I really need 10G?
Do I need a rack?
Power consumption?
How does clustering work with two MS-S1 Max units?
Sources and further reading
- MikroTik CSS610: Switch datasheet.
- MINIS FORUM MS-S1 Max: Affiliate link (schutzgeist-21).
- Amazon shop: All recommendations.
- IRC-Coding.de: Programming tutorials.