OpenClaw 2026
Hardware Matrix
From $6/month VPS servers to 70B parameter home datacenters. Compare 8 distinct hardware architectures to find the perfect host for your personal AI.
βΆ The Comparison Matrix
Sort by any column to find the right balance of price, power, and tokens/sec.
Figures here are indicative, not measurements we took. Throughput depends on quantisation, context length and thermal state; prices move; wattage varies with configuration. Use this to narrow the field to two or three candidates, then measure the one you buy with `ollama run <model> --verbose` and read the eval rate.
| RAM | Tokens/sec | |||||
|---|---|---|---|---|---|---|
Hetzner CPX11 Cloud VPS | 2GB (+4GB Swap) | $6 | 0W | 3 | Zero | 4.5 t/s (Llama 8B) |
Raspberry Pi 5 SBC (ARM) | 8GB | $80 | 5W | 4 | Zero | 6.2 t/s (Llama 8B Q4) |
Intel N100 Mini PC Mini PC | 16GB DDR5 | $150 | 10W | 3 | Low | 5.1 t/s (Llama 8B) |
Orange Pi 5 Plus SBC (ARM) | 16GB / 32GB | $180 | 10W | 5 | Zero | 8.5 t/s (Llama 8B) |
Mac Mini M4 Mini PC | 16GB Unified | $599 | 15W | 8 | Low | 35 t/s (Llama 8B) / 15 t/s (32B) |
Synology DS923+ NAS | Upgraded 32GB | $750 | 35W | 5 | Medium | 8.5 t/s (Llama 8B) |
Mac Mini M4 Pro Mini PC | 48GB / 64GB Unified | $1599 | 25W | 10 | Low | 65 t/s (Llama 8B) / 28 t/s (70B) |
Custom PC (1x RTX 4090) Desktop/Workstation | 64GB DDR5 / 24GB VRAM | $3000 | 450W | 10 | High | 110 t/s (8B) / 35 t/s (70B) |
βΆ Specific Setup Guides
The $6/mo Cloud VPS
Perfect for Telegram/WhatsApp bots. Use small 4B/8B parameter models running 24/7 without worrying about hardware failure or electricity costs.
Apple Silicon Mac Mini
The undisputed king of local AI in 2026. Massive unified memory means you can run 70B parameter models at low power consumption.
NAS (Synology) Setup
Already have a NAS? Upgrading the RAM allows you to run OpenClaw directly alongside your files for maximum document retrieval speed.
How to read this table without being misled by it
Sortable columns invite you to optimise the wrong variable. Sorting by tokens per second puts a desktop GPU on top, which is correct and rarely useful: for an always-on agent the binding constraint is usually memory, then noise, then the electricity bill, and throughput only after that. The machine that runs the model you need, silently, in a room you sit in, beats a faster machine you end up turning off.
Read it in this order. First, decide the model size you need β a 4-bit model occupies roughly 0.6 GB per billion parameters, so an 8B needs about 5 GB and a 32B about 20 GB, plus context and operating system. Second, eliminate every row that cannot hold that in memory; swapping is not a slower option, it is a broken one. Third, among what survives, choose on noise, idle power and whether it needs to be reachable from outside your network. Throughput is the tie-breaker, not the criterion.
The performance score is our own ordering rather than a measurement, and it compresses several different things β memory bandwidth, core count, whether there is a usable GPU path β into a single number. It is useful for grouping, not for comparing two adjacent rows. Where two candidates are close, the honest answer is to rent or borrow both for an hour and run the same prompt on each.