Where Should Your AI Agent Live? Pi, Mac Mini, VPS or Managed

An always-on agent has to live somewhere. The four candidates are a Raspberry Pi on a shelf, a Mac mini in a cupboard, a VPS, or a managed platform — and the received wisdom that the Pi is the cheap option has quietly stopped being true.

Raspberry Pi raised prices twice inside a year. The company was explicit about why: to “offset the recent unprecedented rise in the cost of LPDDR4 memory” and to “secure memory supplies as we navigate an increasingly constrained market in 2026.” A Pi 5 8GB now lists at $175, and the 16GB at $305.

That changes the arithmetic enough to be worth redoing properly.

The 24-month cost

A Pi is a purchase, a VPS is a subscription, so the only fair comparison is total cost over a period. Two years is a reasonable life for a small server.

Raspberry Pi 5, 8GB, built to run something seriously:

On microSD instead of NVMe it’s about $9.90/month — and you should weigh that saving against microSD cards being the single most common failure point in an always-on Pi.

The VPS it’s actually competing with is one with the same 8GB of RAM, not the $6 entry tier:

So they’re close. The Pi comes out modestly ahead of an 8GB VPS at renewal pricing and slightly behind its introductory rate. What it does not do is undercut a VPS by the margin the folklore assumes — and every dollar of it is paid upfront, before you know whether the project survives the month.

Which means the decision has to be made on something other than price.

One factor this comparison doesn’t price: if your agent browses sites with serious bot protection, moving to a datacenter IP makes detection more likely, not less — we looked at the numbers, and residential connections score 85–99% on protected sites against 20–40% for datacenter address space.

The five things that actually decide it

Uptime. Your home has one power supply, one internet connection and no redundancy. A datacentre has multiple of each. For an agent that catches webhooks or answers messages, a four-hour outage while you’re away is the difference between a tool and a toy. This is the Pi’s biggest structural loss and it can’t be engineered away at home.

Inbound access and addressing. A VPS comes with a static public IP. A home connection usually gets a dynamic one, often behind CGNAT, which means no inbound connections at all without a tunnel. For an agent this matters less than you’d expect — as we’ve covered, Telegram and WhatsApp are outbound connections, so a home box can run one perfectly well. It matters enormously the moment a third-party service needs to POST you a webhook.

RAM headroom. 8GB is the floor once browser automation is involved, and on a Pi that’s what you bought — permanently. On a VPS it’s a checkbox and a restart. If you’re not certain what you’ll need, paying monthly for the option to change your mind is worth something.

Disk speed. NVMe on a VPS is standard. On a Pi it’s an add-on, and on microSD it’s slow enough to matter for anything database-backed.

Blast radius. This one is under-considered. An agent runs shell commands and holds long-lived credentials. On a VPS, a compromise gets an attacker a box with your API keys — bad, and contained. On a Pi sitting on your home LAN, it gets them a foothold on the network your laptop, NAS and everything else lives on. If you self-host at home, segment it onto its own VLAN, and if that sentence sounds like work, that’s a real argument for the VPS.

Where the Mac mini fits

The Mac mini is in this comparison for exactly one reason, and it isn’t price.

The current base model pairs an M4 chip with 16GB of unified memory, configurable to 24GB — and “unified” is the point. On a conventional machine, running a local model means fitting it into GPU VRAM. On Apple silicon, the CPU and GPU share one memory pool, so a 24GB Mac mini can hold models that would need a considerably more expensive discrete GPU.

For running local models at home, that’s a genuinely strong position — quiet, low-power, and with more usable model memory per pound than a consumer GPU. For running an agent that calls a hosted API, it’s an expensive way to do what a $6 VPS does, and you don’t need the hardware at all.

Apple doesn’t publish the price on its specifications page, so check the current figure — but it’s several times the Pi and many times a VPS. It earns its place through unified memory or not at all.

The verdict by profile

The tinkerer — a Pi. Not because it’s cheapest, but because owning the hardware is part of what you want, and an agent that occasionally goes down is an acceptable cost of learning. Put it on NVMe, not microSD.

The privacy maximalist — a Mac mini if you’re running local models, a Pi if you aren’t. The whole point is that nothing leaves the house, which a VPS can’t offer at any price. Accept the uptime and segment the network.

Business use, or anything anyone depends on — a VPS. Static IP, redundant power, RAM you can change, and a compromise that doesn’t reach your home network. At roughly the same 24-month cost as a Pi, this is not a close decision once someone else relies on the agent.

The local-LLM user — a Mac mini for unified memory at home, or rent a GPU by the hour if the workload is bursty. Buying a GPU box for intermittent inference is the expensive version of a good idea.

Undecided — a VPS, on a monthly plan. It’s the only option where changing your mind costs nothing, and you can move to hardware later having learned what you actually need.

Which VPS

Hostinger is the easiest landing spot, with 8GB tiers and one-click agent templates — the closest thing to a Pi’s convenience without the hardware. Hetzner is the cheapest per gigabyte when its CX tier is in stock — worth checking before you commit, because that tier was showing as unavailable when we last looked, and the CPX line you’d fall back to is several times dearer. In stock, it is the right answer if you’re comfortable at a terminal, which you probably are if a Pi was ever in the running. Cloudzy bills monthly with no term, which suits the undecided case above. Railway removes the server entirely, though as we found in Railway vs a VPS, its metered pricing works out dearer than a flat VPS for anything resident. UltaHost is worth pricing at the budget end.

DigitalOcean is a common choice here too, but its tracked link isn’t working on our end, so we’ve named it without linking.

Sizing: 4GB for a text-only agent, 8GB with browser automation. Our best VPS for AI agents compares providers, and best VPS for OpenClaw covers the brackets.

How we checked this

Raspberry Pi 5 prices are from Raspberry Pi’s own product page and its published announcement of memory-driven price rises, read in August 2026 — including the quoted explanation about LPDDR4 costs. Note that these have moved twice in a year: an initial rise announced in December 2025, then further increases, which is why we’ve dated them explicitly rather than presenting them as settled. Mac mini specifications — M4, 16GB unified memory configurable to 24GB — are from Apple’s specifications page, which does not display a price.

The 24-month Pi total is our arithmetic on those prices plus typical accessory costs, with electricity estimated at an average 6W draw — about 53 kWh a year, roughly $9–13 depending on your rate. That power figure is an estimate, not a measurement, and a Pi under sustained load draws more. The accessory prices are approximate market rates rather than quotes.

What we did not do: we did not run an agent on each of these four platforms and measure uptime, and we haven’t benchmarked local model throughput on Apple silicon. The unified-memory argument for the Mac mini is architectural rather than something we’ve timed — it’s why the machine can hold a larger model, not a claim about how fast it runs one.

The hosting links above are affiliate links. The Raspberry Pi and Mac mini recommendations earn us nothing, which is worth knowing given that two of the five profiles above point at hardware rather than hosting.

FAQ

Is a Raspberry Pi cheaper than a VPS for an AI agent?

Marginally, and by much less than it used to be. A Pi 5 8GB build works out around $11/month over 24 months against roughly $15 for an 8GB VPS at renewal pricing — but the Pi is paid upfront and loses on uptime, static IP and disk speed.

Why did Raspberry Pi prices go up?

Raspberry Pi attributed it to “the recent unprecedented rise in the cost of LPDDR4 memory” and the need to secure supply in a constrained 2026 market. The 8GB Pi 5 now lists at $175 and the 16GB at $305.

Can I run an AI agent at home?

Yes, and more easily than you’d think — Telegram and WhatsApp connections are outbound, so no inbound ports or static IP are required. The problems are uptime, and that a compromised agent sits on the same network as everything else you own.

Is a Mac mini good for self-hosting AI?

For local models specifically, yes — unified memory means a 24GB Mac mini can hold models that would otherwise need an expensive discrete GPU. For an agent calling a hosted API it’s an expensive way to do what a cheap VPS does.

What’s the security risk of hosting at home?

Blast radius. An agent runs shell commands and holds credentials; on a VPS a compromise is contained to that box, while on your home LAN it’s a foothold next to your laptop and NAS. Put a home agent on its own VLAN.

How much RAM does my agent need?

4GB for text-only against a hosted model, 8GB once browser automation is involved, since a single headless Chromium instance wants 1–2GB. On a Pi that choice is permanent; on a VPS it’s a restart.

What if I’m not sure yet?

Take a monthly VPS. It’s the only option where being wrong costs nothing, and you’ll know within a month what you actually need before committing to hardware.