Hypothetically, if someone wanted to host their own private, continuously-on AI server on Apple hardware—accessible anywhere in the world from an iPhone—what is the most cost-effective way to do it without throwing money away on specs that don't scale?

The Goal:
A personal 70B LLM (Llama 3.3 70B @ Q4) with a long context window and persistent memory (Mem0/Letta), running 24/7 at home, piped directly to an iPhone over an encrypted mesh network (Tailscale).

How to avoid wasting money on this setup:

1 Avoid Over-specifying Brand-New Retail Storage/RAM from Apple:

A maxed-out M5 Max Mac Studio with 128GB RAM and 2TB SSD hits $5,899 retail. However, buying a Certified Refurbished M2/M3 Ultra with 192GB Unified Memory for ~$4,500–$5,200 gives you 64GB more usable VRAM for less money.

2 Don't Waste Money on High Power Bills:

Building a 4× RTX 3090 PC workstation costs ~$5,000 upfront, but drawing 1,400 Watts continuously can add $600–$900 per year to an electric bill. A Mac Studio sitting idle uses ~15 Watts (~$2/month) and tops out at ~160W under full query load, paying for itself in energy savings over 2–3 years.

3 Rely on Smart Context Management over "Raw Context Size":

Instead of trying to force a massive, slow 256k context window into VRAM on every query (which causes massive pre-fill delay), use a local vector database (ChromaDB) + Mem0. Storing memories externally and injecting only the most relevant ~16k–32k tokens keeps response times under 2 seconds on mobile.

4 Zero Cloud API Subscriptions:

At cost, a $4,500 Mac Studio running for 3 years works out to ~$125/month. For a power user hitting API limits on Claude/ChatGPT Pro across multiple devices, a private, uncensored, 100% confidential personal AI server reaches breakeven quickly.

Has anyone deployed a similar 24/7 Mac-based private server for mobile access? What software stack (Ollama vs. llama.cpp + Open-WebUI) gave you the best stability for unattended uptime?