Best Hardware for Running OpenClaw โ Mac Mini vs VPS vs Your Old PC
๐ More on this topic: OpenClaw Setup Guide ยท OpenClaw 100% Local ยท Mac vs PC for Local AI ยท Budget Local AI PC ยท Build a Distributed AI Swarm
Updated 2026-11-06: the Mac mini is no longer a pick. Apple no longer sells the M4 models this page was written around. The $1,000+ tier originally listed a Mac Mini M4 Pro 24GB ($1,399) as Option A, and the quick answer and bottom line recommended the M4 Pro at $1,399-1,900. Apple’s US Mac mini store, checked November 6, 2026, lists an M6 model from $899 and M5 Pro models from $1,699 and $1,899. We haven’t tested them for OpenClaw. The Mac section below stays as background. The 3090 desktop is now ~$1,800-2,125 throughout: the used RTX 3090 band of $1,500-1,625 (eBay US sold, used, excluding shipping, November 6, 2026) plus $300-500 for the desktop around it. Earlier on November 6 this line gave ~$1,700-2,050, from the October 15 band of $1,397-1,552. The quick answer originally said $850+, the $1,000+ tier ~$850-1,200 (with a 2026-09-24 note putting it at $1,500-2,000), the bottom line $1,400-1,900, and the local-models upfront cost in the API-vs-local table $600-1,400 (now $375-2,125).
OpenClaw isn’t a chatbot you open when you need it. It’s an agent that runs all day. It checks your messages, fires heartbeats every 30 minutes, executes scheduled tasks, and waits for instructions on WhatsApp or Telegram. That means whatever hardware you run it on stays powered 24/7.
A gaming desktop with an RTX 3090 draws 90-130 watts at idle. That’s $142-205 per year in electricity before you run a single query. A Mac Mini M4 draws 4 watts. That’s $6 per year. And a free Oracle Cloud VPS costs nothing at all.
The hardware you choose determines three things: how much you spend on electricity, whether you can run local models or need cloud APIs, and how much noise you tolerate in your home. This guide covers all three paths.
Two Modes, Two Sets of Requirements
Before picking hardware, decide how you’re running OpenClaw.
Mode 1: API-Only (Gateway)
The OpenClaw gateway routes your messages to a cloud LLM (Claude, GPT-4, etc.) and returns the responses. The gateway itself is a Node.js process that uses almost no resources:
| Spec | Minimum | Recommended |
|---|---|---|
| RAM | 1-2 GB | 4 GB |
| CPU | 1 core | 2 cores |
| Storage | 10 GB | 20 GB |
| GPU | Not needed | Not needed |
In this mode, a Raspberry Pi, a $4/month VPS, or a used mini PC from 2015 all work fine. The intelligence comes from Anthropic’s or OpenAI’s servers. You’re just paying for the API calls.
Mode 2: Local Models (Gateway + Ollama)
If you want to run models locally through Ollama (no API costs, full privacy), hardware requirements jump:
| Use Case | VRAM / Memory | RAM | Example Hardware |
|---|---|---|---|
| Small models (7B-8B) | 6-8 GB VRAM or 16GB unified | 16 GB | Mac Mini M4 16GB, PC + RTX 3060 |
| Medium models (14B-32B) | 12-24 GB VRAM or 24-48GB unified | 32 GB | Used desktop + RTX 3090 |
| Large models (70B+) | 48GB+ VRAM or 64-128GB unified | 64 GB+ | Mac Studio M4 Max, dual GPU PC |
โ Not sure what fits? Try our Planning Tool.
For the full local setup, see our OpenClaw 100% local guide. This article focuses on which physical hardware to put it on.
Path 1: Mac Mini
The Mac Mini M4 has become the default OpenClaw recommendation in the community. It’s small, silent, sips power, and Apple Silicon’s unified memory lets it run local models that would need a discrete GPU on a PC.
M4 Pricing When This Was Written
| Config | RAM | Storage | Price |
|---|---|---|---|
| M4 base | 16 GB | 256 GB | $599 |
| M4 | 16 GB | 512 GB | $799 |
| M4 | 24 GB | 512 GB | $999 |
| M4 Pro | 24 GB | 512 GB | $1,399 |
| M4 Pro | 48 GB | 512 GB | ~$1,900 |
Update, 2026-10-02: Apple no longer sells the M4 Mac mini. Apple’s US Mac mini store page, checked October 2, 2026, lists an M6 model (12-core CPU, 12-core GPU) from $899 and two M5 Pro models from $1,699 and $1,899. The M4 prices in this table were Apple’s prices when this page was written. New M4 units are now refurbished or third-party stock.
Why People Pick It
Power consumption. 4 watts idle, confirmed by independent testing. Under full LLM inference load, the M4 Pro peaks around 60-65 watts. That’s less than a desktop PC draws at idle. Annual electricity cost at US average rates ($0.18/kWh): about $6.
Silence. The M4 Mini is fanless under light loads. Even under sustained AI inference, it’s barely audible. If OpenClaw runs in your bedroom or living room, this matters.
Unified memory. On a PC, your GPU has 8-24GB of VRAM and models that don’t fit simply don’t run. On a Mac Mini with 48GB, you load a 32B model into unified memory and it runs. No spilling, no crashing. The M4 Pro’s 273 GB/s memory bandwidth is slower than an RTX 3090’s 936 GB/s, so token generation is slower per-byte. But loading a model that fits beats not loading one that doesn’t.
macOS just works. Ollama installs in one command. No CUDA driver issues, no nvidia-smi debugging, no kernel module conflicts. MLX (Apple’s framework) is even faster than Ollama for inference on Apple Silicon.
The Downsides
No NVIDIA GPU. The M4 Pro generates tokens 30-50% slower than an RTX 3090 on models that fit in 24GB VRAM. If the model fits on both, the PC with a 3090 wins on speed.
Expensive to configure. 16GB was tight for local models beyond 7B. The jump to 24GB cost $400 (M4 $999) or $800 (M4 Pro $1,399). Getting 48GB cost $1,900+. The PC counter-argument used to be easy: $850 for a used 3090 and you had 24GB of faster VRAM for less than the Mac. That stopped being true in 2026. By mid-2026 a used 3090 was $1,000-1,400, so the GPU alone cost more than a whole Mac Mini M4 24GB, before you bought anything to plug it into.
Not upgradeable. Memory is soldered. Whatever you buy today is what you have forever. Buy too little and you’re stuck. Buy too much and you overpaid.
What We Recommended, and Why There’s No Mac Pick Now
When this page was written, the M4 base 16GB ($599) was the API-only suggestion and the M4 Pro 24GB ($1,399) the minimum for local models, because the M4 Pro’s memory bandwidth (273 GB/s vs the M4’s 120 GB/s) makes a real difference for inference speed. Those models are gone. The current M6 and M5 Pro Mac minis haven’t been tested here, so the recommendations below don’t include a Mac.
Path 2: VPS / Cloud Server
If you don’t want hardware in your home, a VPS runs the OpenClaw gateway in a datacenter. The tradeoff: no GPU, so you’re locked into API-only mode (unless you rent GPU instances, which cost $0.50-3+/hour and defeat the purpose).
Free and Cheap Options
| Provider | Plan | Specs | Monthly Cost |
|---|---|---|---|
| Oracle Cloud | Always Free ARM | 4 cores, 24GB RAM, 200GB storage | Free |
| Hetzner | CAX11 | 2 vCPU ARM, 4GB RAM, 40GB storage | $4.10 |
| Hetzner | CAX21 | 4 vCPU ARM, 8GB RAM, 80GB storage | $7.00 |
| AWS Lightsail | Cheapest | 1 vCPU, 0.5GB RAM, 20GB SSD | $3.50 |
Oracle Cloud Free Tier
Oracle’s always-free ARM instances are absurdly generous: 4 Ampere A1 cores, 24GB RAM, 200GB block storage. That’s more than enough for OpenClaw’s gateway. The catch: provisioning is difficult. High demand means instances in popular regions are frequently unavailable. You may need to try repeatedly over days or use a script that polls for availability.
If you get one, it’s the best deal in this entire article. Free is free.
Hetzner
Hetzner’s ARM instances are the reliable paid option. The CAX11 at $4.10/month gives you 2 ARM cores, 4GB RAM, and 40GB storage. That’s plenty for the OpenClaw gateway. Datacenters in Germany and Finland only, so latency from the US is 80-120ms. For an agent that responds over WhatsApp, that’s not noticeable.
Why VPS Works (And Why It Doesn’t)
Works for: API-only OpenClaw where you want the gateway accessible from anywhere, don’t want hardware at home, and are already paying for cloud LLM APIs anyway.
Doesn’t work for: Running local models. VPS instances don’t have GPUs. GPU cloud instances (Lambda, RunPod, etc.) cost $0.50-3+ per hour. Running one 24/7 costs $360-2,160/month. At that point, buy a Mac Mini.
Doesn’t work for: Privacy-focused users. Your messages route through a third-party datacenter. The LLM calls go to another third party. Nothing stays on your hardware.
Path 3: Repurposed PC
This is the path for people who already have hardware sitting around, or who want flexibility on a budget.
The Mini PC Route (API-Only)
Used enterprise mini PCs are ridiculously cheap because businesses lease them, use them for 3-4 years, and dump them by the pallet:
| Machine | Typical eBay Price | CPU | Idle Power |
|---|---|---|---|
| Lenovo ThinkCentre M710Q | $70-100 | i5-7500T | 11-14W |
| Dell Optiplex 3060 Micro | $80-120 | i5-8500T | 11-14W |
| HP EliteDesk 800 G4 Mini | $100-150 | i5-8500T | 12-18W |
These machines are about the size of a paperback book. Silent under light loads. 11-14 watts at idle. Ubuntu installs cleanly. Perfect for running the OpenClaw gateway and routing to cloud APIs.
The ThinkCentre M710Q at $85 is what I use as a light-inference node in my distributed swarm build. It runs embeddings and small model queries on 35 watts. For API-only OpenClaw, it barely breaks a sweat.
The Desktop Route (Local Models)
If you want to run local models through Ollama, you need a GPU. That means a desktop (or at least a tower-style case) with a PCIe slot:
| Setup | Cost | What You Get |
|---|---|---|
| Used Optiplex tower + RTX 3060 12GB | $375-475 | 7B-13B models, 12GB VRAM |
| Used workstation + RTX 3090 | $850-1,200 | 32B models, 24GB VRAM |
| Existing gaming PC + 3090 | $850 (GPU only) | 32B models, 24GB VRAM |
| Update, 2026-09-24: both RTX 3090 rows carry $850-1,200 and $850, the figures they originally had. The card is $1,200-1,400 used as of August 2026, so the workstation option is $1,200-1,800 and the GPU-only option $1,200-1,400. |
Update, 2026-10-02: the RTX 3060 row originally said $250-400. The card is $275-325 as of September 2026, so an Optiplex tower ($100-150) plus the card is $375-475.
The budget PC build guide covers this path in detail. The short version: a used Dell Optiplex tower ($100-150) plus a used RTX 3060 12GB ($275-325) gets you a working local AI machine for about $375-475.
Update, 2026-10-02: this paragraph originally said $170-200 for the 3060 and a machine under $450. The card is $275-325 as of September 2026, so the machine is about $375-475.
The Downside: Power and Noise
A desktop with an RTX 3090 draws 90-130 watts at idle. Under inference load, 400-450 watts. That’s a space heater. Annual idle electricity at US average rates: $142-205. And the GPU fans spin, especially under load. If this machine lives next to your desk, you’ll hear it.
Compare that to a Mac Mini at $6/year or a ThinkCentre at $20/year. Two years of the desktop’s idle power alone ($284-410) costs more than buying the ThinkCentre or the Pi outright.
Update, 2026-10-02: this sentence originally said the desktop’s electricity bill can exceed the cost of the hardware itself within two years. A used RTX 3090 is $1,200-1,400 as of August 2026. With it, the whole build is $1,500-2,000, and $142-205 a year of idle power takes 7-14 years to match that.
Raspberry Pi 5 (The $65 Gateway)
The Pi 5 works as an OpenClaw gateway. Quad-core ARM Cortex-A76 at 2.4GHz, available with 2-8GB RAM. The 4GB model ($85) is the sweet spot for OpenClaw’s Node.js process.
| Pi 5 Model | RAM | Price |
|---|---|---|
| Pi 5 2GB | 2 GB | $65 |
| Pi 5 4GB | 4 GB | $85 |
| Pi 5 8GB | 8 GB | $125 |
Power draw: 3-5 watts. Annual electricity: $5-8. Use the official 5V/5A USB-C power supply for 24/7 reliability, and boot from an NVMe SSD (via the Pi’s PCIe slot) instead of a microSD card. SD cards degrade under constant writes.
The Pi won’t run local models. It’s gateway-only. But at $85 total cost and $7/year electricity, it’s the cheapest physical hardware option that works.
Power Consumption: The Hidden Cost
OpenClaw runs 24/7. Electricity costs add up. Here’s what each option actually draws and costs over a year at the US average of $0.18/kWh:
| Hardware | Idle Power | Annual Electricity | 3-Year Electricity |
|---|---|---|---|
| Mac Mini M4 | 4W | $6 | $19 |
| Raspberry Pi 5 | 3-5W | $5-8 | $15-24 |
| ThinkCentre M710Q | 11-14W | $17-22 | $52-66 |
| Dell Optiplex Micro | 10-15W | $16-24 | $47-71 |
| Desktop + RTX 3090 | 90-130W | $142-205 | $426-616 |
| VPS (Oracle free) | 0W (your end) | $0 | $0 |
| VPS (Hetzner CAX11) | 0W (your end) | $49 (hosting) | $148 (hosting) |
The Mac Mini and Pi 5 are in a different category. At 4 watts, the Mac Mini costs less to run for a year than a single dinner out.
The desktop with a 3090 is the outlier. If you run it as an always-on OpenClaw server, electricity alone costs more than a Mac Mini M4 base within two years. The 3090 makes sense for active inference workloads (running models when you need them), not as a 24/7 idle server.
Recommendations by Budget
$0: Use What You Have
Option A: Oracle Cloud free tier. 4 ARM cores, 24GB RAM, free forever. Run the OpenClaw gateway, connect to Claude or GPT-4 via API. The hardest part is provisioning the instance (high demand, limited availability). Once you get one, it’s the best deal here.
Option B: Old laptop or desktop. Any machine from the last decade with 4GB+ RAM and a wired Ethernet connection works as an API-only gateway. Install Ubuntu Server, run OpenClaw, leave it on. The electricity cost depends on the machine, but even a 40W laptop is only $63/year.
$50-100: Used Mini PC
A ThinkCentre M710Q ($85 on eBay) or Dell Optiplex Micro ($80-120) running Ubuntu Server. API-only. 11-14 watts idle. Small, quiet, reliable. This is the practical choice for most API-only users who want physical hardware.
Add a monitor-less headless setup (SSH in from your main machine) and it disappears into a shelf or behind a router.
$500: The Optiplex Build
Update, 2026-10-09: this section originally recommended a $599 Mac mini M4. Apple no longer sells it; as of October 2026 this tier is the Optiplex build only.
Budget PC + RTX 3060 12GB (~$375-475). A used Optiplex tower plus a used 3060 gives you 12GB VRAM for local models up to 13B. Louder and power-hungry compared to the Mac Mini, but you get real GPU inference. See the budget build guide for the full parts list.
Update, 2026-10-02: this option originally said ~$400. The RTX 3060 is $275-325 as of September 2026, so the Optiplex-plus-3060 build is about $375-475, still under this section’s $500.
$1,000+: Local Models, Always-On
Used desktop + RTX 3090 (~$1,800-2,125). 24GB VRAM. Runs 32B models at Q4 with room for context. Downside: 90-130W idle, fan noise, needs a case and PSU that fit a 350W three-slot GPU. Best if you already own a compatible desktop and just need the GPU.
The catch: a 3090 desktop is loud under load and draws 90-130W at idle. If it lives in a closet or garage, that doesn’t matter much. If it lives next to your desk, keep the gateway on a mini PC or Pi and use the hybrid setup below.
When to Run Local Models vs Just Use APIs
This is the decision that determines your hardware requirements more than anything else.
| Factor | Run Local | Use APIs |
|---|---|---|
| Monthly budget | $0 (already own hardware) | $15-150/month (API costs) |
| Privacy | Data never leaves your machine | Messages route through cloud providers |
| Model quality | Good (32B local) to great (70B on Mac) | Best available (Claude Opus, GPT-4) |
| Inference speed | 15-45 tok/s (depends on hardware) | 50-100+ tok/s |
| Reliability | Depends on your hardware and power | 99.9%+ uptime |
| Upfront cost | $375-2,125 | $0-85 (gateway hardware) |
| Complexity | Install Ollama, manage models, tune configs | Point at API, done |
The Hybrid Approach
Most practical OpenClaw setups use both. The token optimization guide covers this in detail, but the short version:
- Heartbeats (every 30 minutes, just keep-alive pings): route to a free local model via Ollama
- Simple tasks (file management, reminders, basic queries): route to a cheap cloud model (Haiku at $1/$5 per million tokens)
- Complex tasks (multi-step reasoning, code generation, research): route to Claude Sonnet or Opus
This lets you run the gateway on cheap hardware (mini PC, Pi, VPS) with Ollama handling only heartbeats. Complex work goes to the cloud. You get the cost savings of local heartbeats without needing a 3090 desktop to run the smart model.
See the model routing guide for the exact openclaw.json configuration.
Bottom Line
The hardware decision comes down to one question: are you running local models or not?
API-only (no local models): Get the cheapest thing that stays on. Oracle free VPS, a $65 Raspberry Pi 5, or an $85 used ThinkCentre. The gateway needs almost nothing. Spend your budget on API credits instead of hardware.
Local models (privacy, no recurring costs): Get a used desktop with an RTX 3090, and price it honestly: the card is $1,500-1,625 used (November 2026) and the desktop around it is another $300-500, so you’re at $1,800-2,125. You pay for it in noise and 90-130W at idle. On a smaller budget, the Optiplex plus RTX 3060 12GB build (~$375-475) handles 7B-13B models.
Not sure yet: Start API-only on cheap hardware. If you decide you want local models later, you can always add them. The gateway hardware doesn’t change. You just add Ollama and a capable machine to run it on.
Related Guides
- OpenClaw Setup Guide โ installation from scratch
- Running OpenClaw 100% Local โ zero API costs with Ollama
- OpenClaw Token Optimization โ hybrid routing for 97% cost reduction
- OpenClaw Model Routing โ route tasks to the right model
- Mac vs PC for Local AI โ deep dive on the platform question
- Build a Local AI PC for Under $500 โ the desktop route
- Building a Distributed AI Swarm โ multi-node setup
- How Much VRAM Do You Need? โ model sizing by hardware
A weekly email with every new guide and measured benchmark.
Subscribe โ free, no spam