AI Knowledge HubMac Mini AI Server vs One Rental Server — In the AI Era, Where Should My Server Live?
Mac Mini AI Server vs One Rental Server — In the AI Era, Where Should My Server Live?
In the last post I wrote that "renting a server beats setting up a personal server on a Mac mini." That statement needs one caveat attached to it. If you want to run AI directly, the story changes. This post puts the Mac mini on the table as a variable and actually runs the comparison.
1. The premise — these are two different sports
First, let's be honest. A 4,500-won-a-month rental VPS and a Mac mini aren't playing the same game. The rental VPS is a small web server with nothing but a CPU. Saying you're going to run an LLM on it is like saying you'll move house with a compact car. The Mac mini, meanwhile, is a machine built for AI inference. Thanks to Apple Silicon's unified memory architecture, it runs models you couldn't even load onto a regular PC with small GPU memory. So this comparison isn't "which is more powerful" — it's "which one fits what I'm trying to do."
2. Mac mini price list (official store)
| Config | Memory · SSD | Price |
|---|---|---|
| M4 base | 16GB · 256GB | 890,000 won |
| M4 | 16GB · 512GB | 1,190,000 won |
| M4 | 24GB · 512GB | 1,490,000 won |
| M4 Pro | 24GB · 512GB | from 2,090,000 won |
| M4 Pro high spec | 48GB · 1TB | around 2,690,000 won |
The 890,000-won starting price catches your eye, but if AI is in the picture, the 24GB / 512GB (about 1.49 million won) is the realistic starting line. Memory is the most important choice — Apple basically doesn't let you upgrade after purchase, so you have to settle it when you buy.
3. How local LLMs actually feel by memory size (4-bit quantization)
| Unified memory | Models it runs | How it feels |
|---|---|---|
| 16GB | 7B–9B | Simple chat, translation, summarizing. Tight once the OS and apps are running too |
| 24GB | 13B–14B | Practical coding assist and document analysis |
| 32GB | 26B–27B | Can run quality models like Qwen 32B-class |
| 48GB+ | 32B–70B | Large models comfortably, concurrent requests |
A 7B-class model like the supergemma I use runs on 16GB. But if you want to use it seriously, 24GB is the floor, and going roomy means M4 Pro or above.
4. Mac mini setup — easier than you'd think
The standard combo for a Mac mini AI server:
| Role | Tool | Notes |
|---|---|---|
| LLM inference | Ollama | Install the official app, then ollama run. Port 11434 |
| Chat UI | Open WebUI | One Docker line for a ChatGPT-style screen. Port 3000 |
| Public access | Cloudflare Tunnel | HTTPS exposure with no port forwarding. Free |
Cloudflare Tunnel is the key. The biggest walls for a home server are the router and the dynamic IP. Some ISPs block ports 80 and 443, and if you have to do DDNS plus port forwarding on top, beginners burn out before they even start. A tunnel works by having the Mac mini throw an encrypted channel outward first, so you never touch the router. HTTPS comes free with it. And it doesn't drop when your home IP changes.
5. Measured electricity — 2,000 to 3,000 won a month
Per Apple's own environmental reporting, the Mac mini idles at 3W and pulls around 20W under load. Even running it at 20W all day works out to about 26,000 won a year, or 2,200 won a month. Over three years that's a bit over 80,000 won in electricity — a couple of months of rental server fees. You can stop worrying about the power bill.
6. Three-year total cost
| Item | Mac mini M4 24GB/512GB | Rental VPS (entry) |
|---|---|---|
| Upfront | 1,490,000 won | 0 won (no setup fee) |
| Monthly | ~2,200 won electricity | from 4,500 won |
| 3-year total | ~1,570,000 won | ~162,000 won |
| Residual after 3 yrs | Substantial recovery via resale | 0 won |
| LLM inference | 13B-class possible | Not possible (CPU-only mini VPS) |
| Public service | Tunnel needed, upload-line dependent | Fixed public IP, instant exposure |
By the numbers the rental server is 10x cheaper. But a rental VPS can't run an LLM. It isn't a comparable target. If AI inference is the point, the Mac mini's real rival isn't a 4,500-won VPS — it's GPU cloud (RunPod and the like, billed by the hour). That side breaks into the ten-thousand-won-per-hour range once you actually run it. If you're hitting a local model for a few hours a day, buying a Mac mini outright is the cheaper play.
7. The home server's weaknesses — not hiding them
| Weakness | What it means |
|---|---|
| Power outage | When the house loses power, the server dies with it |
| Upload line | Home internet is download-heavy, so access from outside is slow |
| Public address | You have to solve public IP, DDNS, or tunnel yourself |
| Scalability | If traffic suddenly explodes, there's nothing you can do |
| Dedicated management | When it breaks it's all on you. I check it at night too |
| Fine-tuning | No NVIDIA CUDA, so not suited for training (inference only) |
8. Verdict by use case
| Use case | Choice |
|---|---|
| Running a homepage or blog | Rental server. Fixed IP and snapshots, nothing to debate |
| Local AI inference (daily use) | Mac mini 24GB+. Cheaper than GPU cloud |
| Local AI (occasional dabble) | Mac mini 16GB base, or rental server plus free API |
| Linux for study and experiments | Rental server. Break it and a snapshot brings it back |
| Files, media, home automation | Mac mini. Things used inside the house belong on a house machine |
| Commercial service | Rental server. SLA and uptime are bought with money |
9. My conclusion — I use both
I'm not taking back the praise I gave rental servers last post. The one I use is https://my.yourit.co.kr/. It's the official page where signup, management, and billing all happen. My plan has 2TB of traffic, and having used it, I think it's genuinely great. For a public homepage the rental server is the answer. It's up in five minutes, you can pay by bank transfer without a card, and if you break it, you break it.
But if you're someone who touches AI every day, a Mac mini belongs under your desk. The experience of hitting a local model without worrying about API fees is something only people who've done it understand.
To put it plainly: what you show the outside world is the rental server, what you run on the inside is the Mac mini. A 4,500-won server and a Mac mini that costs 2,000 won of electricity a month. Combined that's under 10,000 won a month. The answer for a personal server in the AI era isn't "which one" — it's "both, each in its own place."