Why Your Next Home Server Should Be a Mac Mini, Not a Raspberry Pi

I own three Raspberry Pis. They’ve run my DNS, my sensors, and various half-finished weekend projects for a decade, and I’ll defend the little boards in any argument — up to a specific point. That point arrived when I asked what a home server should actually do in 2026, and the answer included the words “run AI workloads.” The Pi’s case collapses there, and the machine that replaces it is, surprisingly, the cheapest Mac Apple sells.

My home server today is an M2 Mac Mini bought used. Here’s the full argument, including the honest section about when the Pi is still the right answer.

The capability gap is not 2x, it’s 10–30x

The Raspberry Pi 5 (16 GB, ~$120 with PSU, case, SD card, and an NVMe HAT pushing the real build to ~$180) is a wonderful general-purpose Linux board. But ask it to do AI inference and the arithmetic gets grim. Running Llama 3.2 3B quantized on a Pi 5, I measured ~2 tok/s generation. The same model on a base M2 Mac Mini: ~45 tok/s. Whisper transcription of a 10-minute recording: about 18 minutes on the Pi 5 with the small model, versus 50 seconds on the Mini with large-v3-turbo — a better model, 20x faster.

The reason is structural, not fixable with overclocking: the Pi has no usable GPU acceleration for these workloads and slow memory bandwidth (~17 GB/s). Apple Silicon’s unified memory runs at 100 GB/s on the base M2 and the GPU does the heavy lifting via Metal. For LLMs, memory bandwidth practically is performance. Depending on workload, the Mini lands 10–30x ahead — which is the difference between “AI server” and “AI demo that you stop using.”

What an AI home server actually does all day

This isn’t hypothetical; this is the workload list running on my Mini right now:

• Ollama serving the household at 0.0.0.0:11434. My MacBook, my wife’s laptop, and my shortcuts all hit the Mini for Qwen 2.5 14B. One always-on model server beats four laptops each burning battery on local inference. (Set OLLAMA_HOST=0.0.0.0 via launchctl setenv to expose it on the LAN.)
• Whisper transcription jobs. A watched folder (Hazel, or a launchd job running a script) — any audio file dropped in comes back as text minutes later. Voice memos in Czech included.
• Photo indexing. An overnight job runs CLIP-style embeddings over our photo archive so I can search “kids at the lake” locally.
• Home Assistant in a Docker container (OrbStack runs it with near-zero overhead), now wired to Ollama for the local voice assistant pipeline.
• Tailscale node, so all of the above is securely reachable from anywhere without a single open port.
• Time Machine target for both laptops, over SMB to an external SSD.
• Media server (Plex or Jellyfin), where the Mini’s hardware video encode embarrasses the Pi’s transcoding story completely.

A Pi 5 runs items four through seven happily. It cannot meaningfully run the first three — and the first three are the reason to have a home server in 2026.

The power bill: measured, not guessed

The Pi’s traditional trump card is power draw. I put a metering plug on both machines for a week. Czech electricity runs me about 7 CZK/kWh (~€0.28), which makes the math vivid:

The Mini idles at 6 watts. Read that again if you’re picturing a tower server — it’s a rounding error, €10 a year more than the Pi, achieved while being completely silent (the fan is inaudible at idle and barely audible under sustained inference). The “Pi saves power” argument is true and worth less than one pizza annually. Meanwhile every inference job finishes 10–30x sooner, so the Mini often uses less total energy per task.

macOS reliability deserves a word too: my Mini’s uptime is bounded only by security updates, there’s no SD card to corrupt (the classic Pi failure mode — I’ve lost two cards over the years), and the hardware is a sealed, fanless-quiet unit with Apple-grade build quality.

Headless setup: making macOS behave like a server

macOS isn’t a server OS out of the box, but ten minutes of settings fix that:

1. Screen Sharing + Remote Login. System Settings → General → Sharing: enable Screen Sharing (full GUI when needed) and Remote Login (SSH for everything else).
2. Never sleep. System Settings → Energy: enable “Prevent automatic sleeping,” enable “Start up automatically after a power failure.” Or from the terminal: sudo pmset -a sleep 0 displaysleep 1.
3. Auto-login — think before enabling. A server must come back up after a power cut without a keyboard. Auto-login (Settings → Users & Groups) makes that happen, but it means FileVault must be off or the disk password entered manually anyway, and anyone with physical access gets a session. My compromise: auto-login on, FileVault off for the system disk, sensitive data on an encrypted external volume mounted via a password prompt I handle over Screen Sharing after rare reboots.
4. Login items. Ollama, OrbStack, and Tailscale all install as login items, so a reboot restores the full stack with zero intervention.
5. No monitor needed. Modern Minis run headless without dummy HDMI plugs; Screen Sharing performance is fine.

# Health-check the stack from your laptop:
ssh mini "uptime && curl -s localhost:11434/api/tags | head -c 200"

The price math, including the used market

New base M4 Mini (16 GB/256 GB): $599 / ~15,000 CZK — and frankly the best server value Apple has ever shipped. But the used market is where this argument gets unfair: M1 Minis (16 GB) go for $250–350, M2 Minis for $350–450 on Czech bazaar sites and eBay alike. A used M1 Mini still delivers 35 tok/s on an 8B model — that’s a fully capable AI home server for roughly 2x the price of a fully built-out Pi 5 kit, with 15–25x the AI throughput. Add an external 2 TB SSD ($130) for Time Machine and media and you’re done. There is no x86 mini-PC at that price that matches the inference-per-watt, and nothing at any price that does it this quietly.

Where the Pi is still the right answer

Honesty section, because I’m keeping all three of my Pis:

• GPIO projects. The Mini has no pins. Sensors, relays, LED strips, garage-door hacks — Pi territory, permanently.
• Dirt-cheap single tasks. A Pi Zero 2 W running Pi-hole costs $15 and sips 1 W. Buying a Mac to block ads is silly; my DNS still lives on a Pi.
• Clusters for learning. Four Pis teach you Kubernetes, Ansible, and network booting for the price of one Mini. As an education platform it’s unbeatable.
• Hostile environments. The shed, the camper, the project that might get rained on — risk $60, not $600.

The decision rule I’d give a friend: if the job involves pins, costs under $50 to solve, or might die in a shed, buy a Pi. If the sentence “and then it runs AI on it” appears anywhere in your plan, buy a used M-series Mini and thank yourself every single day it answers in two seconds instead of two minutes.

The Pi didn’t get worse. The definition of a home server changed underneath it. Mine now answers questions, transcribes meetings, and finds photos — and it does it from a silent aluminum square that costs €22 a year to feed.