Choosing hardware for a home server

Nearly every regretted home server purchase is the same mistake: someone picked a CPU first. Cores are the resource a home server runs out of last. The order that actually predicts whether a machine will still be adequate in three years is RAM, then a modern Intel iGPU if and only if you transcode video, then disks, then cores.

Write the stack down before you open a parts list#

List the services by name, not by category. "Media" is not a workload. "Jellyfin serving three concurrent streams to a Roku, Immich with 90,000 photos, Home Assistant plus Zigbee2MQTT, Paperless-ngx, Vaultwarden and Caddy in front" is a workload, and it sizes itself.

The arithmetic is unglamorous: add the idle footprints, add the peak of whichever one spikes hardest, then double the total so the page cache has somewhere to live. Stack planner does this for you against the profiled figures. The reason it works is that most self-hosted software is small. The median idle footprint across the index is 150 MB. What blows the budget is a short list of specific tools, and you know which ones you picked.

Build one: the 40 dollar-a-year box#

An Intel N150 mini PC, 16 GB of DDR5 SODIMM, one 500 GB NVMe drive. New, that is roughly 180 to 250 USD depending on whether you buy it barebones and supply your own memory. Refurbished 8th-generation Dell and Lenovo micro desktops sit in the same price band with more expandability and more idle draw.

The N150 is a 4-core part with a 6 W processor base power and 24 execution units of Intel graphics with Quick Sync. Note the ceiling: Intel documents a 16 GB maximum memory size for it, which is the reason this tier is a light-stack tier and not a general answer.

Measured at the wall with a disk attached, expect 7 to 12 W idle. Call it 10 W. At the US average residential price of 18.34 cents per kWh (EIA, June 2026) that is 10 W times 8760 hours, or 87.6 kWh, or about 16 USD a year. At a European 0.30 EUR per kWh it is closer to 26 EUR, and once you hang a spinning 3.5 inch disk off it you are at 40 a year in either currency. That is the whole running cost. What a home server costs to run does the full total-cost arithmetic including disks and backup storage.

This build runs a reverse proxy, a DNS blocker, a password manager, a wiki, a feed reader, a bookmark manager, an uptime monitor and a media server doing direct play, all at once, without effort. It will not run Immich comfortably at 16 GB alongside all of that, and it will not hold your photo library on a 500 GB NVMe.

Build two: the mid tier with an iGPU#

This is the build most people should end up with, and the one that survives a stack growing sideways.

• Intel Core i3-12100 or i5-12400 (or a 13th or 14th generation equivalent), on a B660 or B760 microATX board. Roughly 100 to 180 USD for the CPU, 90 to 130 for the board.
• 32 GB of DDR4 or DDR5, two sticks. 60 to 110 USD.
• 1 TB NVMe for the OS, container data and databases. 60 to 90 USD.
• Two 8 to 14 TB CMR drives for bulk media. 130 to 250 USD each.
• A 450 W efficient PSU and a case with real drive mounts. 120 to 180 USD.

Total around 650 to 900 USD, of which the disks are a third. Idle draw with two spinning disks lands between 20 and 35 W, so 30 to 55 USD a year at US prices.

The iGPU is the point. Jellyfin's hardware guidance names an N100 or an i5-11400 or newer as the recommended Intel integrated graphics, and explicitly removed 7th to 10th generation parts from that list because Intel deprecated the toolkit for them. Those older chips still work; they are just no longer the thing to buy. The same /dev/dri device serves Frigate for OpenVINO detection and hardware decode, and Immich for video transcoding, so one iGPU covers three workloads. Details of which app needs which flags are in GPUs, transcoding and local AI.

32 GB, not 16, is what makes this tier different. Immich's docs ask for 8 GB recommended on its own. Add Frigate with semantic search, which documents an 8 GB minimum and recommends 16, and 16 GB total stops being a plan.

Build three: storage first#

When the requirement is "hold 40 TB and do not lose it", the compute is an afterthought and the case is the hard part.

• A microATX or ATX board with six or more SATA ports, or four plus an LSI 9207-8i HBA flashed to IT mode from the used market at 30 to 60 USD.
• A case with real bays: a Fractal Design Node 804 or Define R5 from the used market, or a Jonsbo N3/N5 if you want it small. 100 to 200 USD.
• Six 16 to 20 TB CMR drives. This is 1200 to 1800 USD and it is most of the build.
• 32 to 64 GB if you run ZFS. Proxmox's own ZFS guidance says start at 8 GB and add about 1 GB per TB of pool, and installs since 8.1 cap the ARC at 10 percent of RAM with a 16 GiB ceiling.
• Any modern 6-core CPU. Genuinely any.

Total 1800 to 2600 USD, and it idles at 45 to 70 W with six drives spinning. That is 70 to 110 USD a year in electricity, which is a real line item and the reason people spin disks down. Whether you should is covered in ZFS, btrfs, mdadm or one disk, along with why ZFS on a snapshot schedule tends to keep them awake anyway.

ECC, honestly#

ECC RAM detects and corrects single-bit memory errors. The people who genuinely need it are running a large pool of irreplaceable data, or a database that many people depend on, or a machine that stays up for years between reboots. If you are buying a used server board or a Xeon-based workstation, ECC is already there and free, so take it.

What is not true is that ZFS requires ECC, or that ZFS without ECC is more dangerous than any other filesystem without ECC. TrueNAS's own hardware documentation says ECC is recommended but not required, and says it plainly. The OpenZFS tuning documentation does not raise the subject at all.

The uncomfortable version: ECC is often sold as an upgrade to people whose actual exposure is that they have one copy of their data. Memory errors are a low-probability corruption path. Having no restorable backup is a certainty-of-loss path. If your budget forces a choice between an ECC platform and a second copy of the data somewhere else, buy the second copy. See Backups that actually restore.

Idle watts, not peak watts#

A home server spends 95 percent of its life doing nothing. The number that determines what it costs and how hot the cupboard gets is idle draw, not TDP and not benchmark power. Two machines with identical spec sheets can idle 25 W apart depending on the power supply's efficiency at 10 percent load, how many case fans you fitted, whether the board reaches package C-states with your PCIe cards installed, and whether an HBA is fitted at all (a flashed LSI card adds 8 to 12 W and never sleeps).

Over five years, 8 W versus 60 W at 18.34 cents per kWh is the difference between 64 USD and 482 USD. That gap is a second machine.

Buy a 15 USD plug-in power meter before you buy anything else. It is the only way to know what you actually own.

arm64 is mostly fine, and the exceptions are the ones you care about#

95.2 percent of the profiled tools publish official arm64 builds, so the container ecosystem is no longer the reason to avoid ARM. The reasons that remain are specific:

• There is no Quick Sync. ARM video engines exist but the software support is a patchwork.
• Jellyfin dropped 32-bit armhf builds entirely at 10.11, so a Pi running a 32-bit OS is stuck on 10.10.x.
• Immich runs on arm64 but only accelerates machine learning through ARM NN on Mali or RKNN on Rockchip; CUDA, OpenVINO and ROCm are x86-64 only.
• Emby publishes arm64 as a separate Docker Hub repository rather than a multi-arch tag, so a copied compose file pulls the wrong image.
• Proxmox VE got official arm64 builds in August 2026, but they target NVIDIA Grace class servers, not single-board computers.

The trap: buying for the stack you imagine#

The most expensive home server is the one specced for the Kubernetes cluster, the local language model and the six-camera NVR that the buyer has not yet installed and, statistically, will not. Two years later the machine is a generation behind, has been idling at 60 W the whole time, and is running eleven containers that would fit on a 200 USD mini PC.

The counter-move is to buy for the stack you will run in the next six months, and to buy a platform that can take a second RAM stick and a second disk. RAM and disks are the two things you will actually add. A CPU upgrade almost never happens.

If you genuinely might add Ollama later, note that the upgrade is a GPU with enough VRAM, which is a case and PSU decision you can make now for free by picking a case that fits a full-height card and a PSU with the connectors. That costs nothing today and saves a rebuild.

What to do next#

Run your candidate list through Stack planner to get a RAM figure you can defend, then decide between form factors with Mini PC vs NAS vs used enterprise. If you have not settled the operating system question yet, Proxmox vs TrueNAS and Your first self-hosted server cover the decisions that are hard to reverse once data is on the disks.