I Wanted a Mac Mini for a Year. OpenClaw Ruined My Financial Restraint.

Personal
#homelab #ai-ml #openclaw #self-hosted #privacy #llm

For almost a year, I kept coming back to the same stupid idea: buying a Mac Mini M4 Pro just to run local LLMs.

Not because I needed one. Not because the numbers made sense. Mostly because the idea of a small, quiet box under my desk serving local LLMs to the rest of my homelab was extremely my kind of nonsense.

The version I wanted cost about €2,800, which is a lot of money to spend on what was, for most of that year, basically a homelab-flavoured daydream.

So I didn’t buy it.

Every time I tried to justify it, the argument fell apart. Yes, local models are fun. Yes, running everything on your own hardware is appealing. Yes, a Mac Mini is a very nice little box. But none of that changed the basic problem: I would have been spending serious money to get an experience that was, in most cases, worse than just paying for Claude or ChatGPT.

The Argument That Never Worked

I already had Ollama running in my homelab. I set it up a couple of years ago mostly because I was curious and wanted to see how it worked. Which, in hindsight, was also the problem. The hardware I had at the time was nowhere near fast enough to run bigger models at a speed that felt enjoyable, so the whole thing mostly sat there as an interesting curiosity.

Still, the idea kept coming back. A proper little machine for local LLMs. Ollama serving LLaMA, Mistral, or whatever the open-source flavour of the month was. No API costs, no cloud dependency, everything on my own hardware. Very on-brand, unfortunately.

The problem was simple: open-source models just were not at the frontier. GPT-4o, Claude Sonnet, Gemini - the closed models were still clearly better for writing, brainstorming, research, and reasoning. If I was buying a Mac Mini as a ChatGPT replacement, I would have been paying €2,800 for a worse experience.

And then the cost math made it even less convincing. €2,800 divided by €23 a month for ChatGPT Plus is about ten years of subscription. Ten years before break-even, for an inferior product, and that is before even looking at the electricity bill. As a pitch to my wife, this had absolutely no chance.

So every few months, I would revisit the idea, fail to make the numbers work, and close the tab again.

What Changed: It’s Not About the Chatbot Anymore

Then I set up OpenClaw, a self-hosted AI assistant gateway that connects to messaging apps and homelab services. I wrote about the setup experience in my previous post.

This is where the argument changed. OpenClaw is not really a chatbot in the usual sense. It is an agent that can actually do things: submit insurance claims from Paperless-ngx, control Home Assistant, manage tasks, check calendars, automate workflows. It sits in my Telegram and has access to a slightly concerning amount of my homelab. I did lock a few things down. I am not completely reckless.

That changes the economics in two ways.

The first is cost. Once you have an AI assistant that is always on - checking heartbeats, responding to messages, and running automations throughout the day - API costs add up fast. I hit my Anthropic monthly limit embarrassingly quickly. Topped up. Hit it again. The pay-per-token model that feels fine for occasional chatbot use starts to look a lot less cute when the thing is running all day. Admittedly, using an LLM to turn off the bathroom light is the kind of overkill that should probably come with a warning label.

The second is breadth. Before OpenClaw, turning an AI idea into a real workflow usually meant several evenings of building an app or wiring together n8n automations. Now it is more like 30 minutes to write a skill. When the friction drops that much, the hardware suddenly has actual jobs to do - and lots of them. That changes the calculation completely.

The Key Insight: Agents Don’t Need Frontier Intelligence

This was the bit I had been getting wrong.

Most agentic tasks are not actually that hard. “Check my calendar and send me a summary” does not need GPT-4-level reasoning. “Submit this invoice to my insurance portal” does not need frontier intelligence. “Turn off the heating when I leave” definitely does not.

What those tasks need is reliability, speed, cost efficiency, and - for anything touching personal or business data - privacy. A capable local model like LLaMA 3, Qwen, or Mistral can handle that comfortably. It is not writing a PhD thesis for me. It is executing structured tasks inside a homelab I already built. The quality bar is much lower than “replace my thinking”. It is more like “reliably operate the setup without doing anything stupid”.

The Privacy Bonus

There is a secondary argument too. Researchers at ETH Zurich published a paper at ICLR 2024, “Beyond Memorization”, showing that LLMs can infer personal attributes like age, location, income, and health information from casual text with 85% accuracy, even when nothing is explicitly stated.

I am not going full tinfoil hat here. Cloud providers generally have legitimate data practices. But an AI assistant with access to medical documents, financial records, and daily home routines does reveal a lot by accumulation. Keeping that processing local just feels cleaner to me. It is the same reason I self-host Paperless instead of handing all of that to some random cloud document service.

If you are using AI assistants with real business data - client lists, sales pipelines, product roadmaps - this is worth taking seriously. Your competitive intelligence is in those prompts whether you intended to put it there or not.

But honestly, for me, privacy is more of a bonus than the main reason. The real unlock was the economics of agentic use.

The New Math

With the right framing, the cost calculation looks very different. The €2,800 is a one-time spend. After that, local workloads do not generate per-token costs. There are no rate limits to bump into and no API outages breaking automations. The cloud APIs are still there for the things that actually need frontier intelligence: harder reasoning, difficult coding, nuanced writing.

That hybrid model is the whole point. Route sensitive or routine agentic tasks to the local model. Route genuinely hard problems to Claude or GPT-4o when needed. Best of both worlds, and a much smaller bill for the always-on part.

Suddenly €2,800 is not competing with a chatbot subscription anymore. It is competing with an ongoing API bill for an always-on homelab assistant.

So yes, I finally found a halfway respectable justification for a Mac Mini. Which is obviously terrible news for my financial restraint.

The actual setup - getting Ollama running, connecting it to OpenClaw, and figuring out which models work best for homelab agentic tasks - is probably a follow-up post once the hardware arrives.

Previously: Setting Up OpenClaw on a Homelab.

References: