Skip to content

Instantly share code, notes, and snippets.

@yalexx
Last active June 18, 2026 12:55
Show Gist options
  • Select an option

  • Save yalexx/8358392506daab8a854b7a665ac81dd8 to your computer and use it in GitHub Desktop.

Select an option

Save yalexx/8358392506daab8a854b7a665ac81dd8 to your computer and use it in GitHub Desktop.
OpenClaw Hardware Requirements 2026: Minimum, Recommended, and Production tiers — what you need to run OpenClaw locally

OpenClaw Hardware Requirements 2026: Minimum, Recommended, and Production tiers — what you need to run OpenClaw locally

OpenClaw Hardware Requirements (2026): Minimum, Recommended, and Production

Last updated: May 2026.

OpenClaw hardware requirements are simpler than most people expect — but the difference between "will it run" and "will it run well 24/7" is the reason we wrote this. This is a focused buyer's guide: what the OpenClaw runtime actually needs on the hardware side, where most DIY builds fail, and what the official OpenClaw hardware — ClawBox — ships with out of the box.

The Three Tiers

1. Minimum (it will run, barely)

The lowest spec where OpenClaw boots, accepts a chat, and responds without crashing.

  • CPU: 4-core ARM or x86, 2.0 GHz+
  • RAM: 4 GB unified or system memory
  • AI acceleration: Optional (CPU-only inference possible with small models)
  • Storage: 64 GB, any medium (microSD tolerable for learning)
  • Network: 100 Mbps Ethernet or Wi-Fi
  • Power envelope: <10 W
  • Example: Raspberry Pi 5 (4 GB), Intel N100 mini PC

Realistic experience: 3B-parameter quantized model, 4–8 s first-token latency, single user, no concurrent browser automation. Fine for "I want to see what OpenClaw is."

2. Recommended (always-on assistant, single user)

Where OpenClaw starts to feel like the product it is — responsive, multi-modal, reliable.

  • CPU: 6-core ARM or x86, 2.0 GHz+
  • RAM: 8 GB unified memory (unified matters — PCIe traffic kills latency)
  • AI acceleration: 20+ TOPS dedicated (Hailo-8L, NPU, or integrated Jetson GPU)
  • Storage: 256 GB NVMe (not microSD)
  • Network: 1 Gbps Ethernet
  • Power envelope: 15–25 W
  • Example: Jetson Orin Nano 8 GB (DIY), ClawBox

Realistic experience: 7B-parameter quantized model (Llama 3 8B, Qwen 2.5 7B), 0.5–1 s first-token latency, concurrent browser automation, voice transcription in the background. This is the sweet spot for home or single-operator office.

3. Production (always-on, multi-user, business workloads)

Where OpenClaw can carry a whole team or household without degrading.

  • CPU: 8+ core, 3.0 GHz+
  • RAM: 16 GB+ unified
  • AI acceleration: 60+ TOPS dedicated (Jetson Orin Nano 8 GB Super, Mac Mini M3 Neural Engine)
  • Storage: 512 GB NVMe minimum
  • Network: 1 Gbps Ethernet (wired is worth it here)
  • Power envelope: 15–40 W
  • Example: ClawBox (official OpenClaw hardware appliance, €549), Mac Mini M3, DIY Jetson Orin Nano 8 GB Super

Realistic experience: 7B-parameter models with long context, 3–5 concurrent users, simultaneous browser automation + transcription + chat, 24/7 uptime on 15 W.

What Most DIY Builds Get Wrong

After shipping hundreds of OpenClaw appliances, we see the same four failure modes in DIY hardware builds:

1. Running on microSD

MicroSD wears out. It also bottlenecks inference when the model weights stream off disk. Use NVMe. Every ClawBox ships with a 512 GB NVMe — and the single most common upgrade a DIY Jetson owner performs is "add NVMe."

2. Undersized power supply

Jetson Orin Nano needs a solid 5 V / 4 A or USB-C PD supply. A phone charger works until the GPU spikes during inference, then the whole system resets. ClawBox ships with a matched PSU. DIY builds need one that can deliver 25 W sustained.

3. Cooling that almost works

Passive cooling is fine at idle. Under sustained inference, an un-sinked Jetson Orin Nano thermal-throttles to half speed within 90 seconds. Either buy a proper heatsink+fan assembly or accept the throttling. ClawBox uses a tuned passive heatsink sized for the 15 W envelope so it stays silent without throttling.

4. Wi-Fi for everything

OpenClaw speaks MCP, runs webhooks, and does browser automation. Wi-Fi latency jitter turns cleanly into user-visible lag. Use wired Ethernet for anything production.

OpenClaw Hardware Requirements vs. Other Popular Setups

Setup Meets minimum? Meets recommended? Meets production? Notes
Raspberry Pi 5 (4 GB, microSD) Learning only
Pi 5 (8 GB) + Hailo-8L HAT + NVMe ✅ (tight) 13 TOPS ceiling
Intel N100 mini PC No AI accelerator
Intel NUC 13 + discrete GPU ✅ (power-hungry) 45 W+ idle
Mac Mini M3 (16 GB) €1,000+, ARM/macOS
DIY Jetson Orin Nano 8 GB ✅ (with NVMe + PSU) 8–12 h setup
ClawBox (Jetson Orin Nano 8 GB Super) 5-min setup, €549

The Official OpenClaw Hardware: ClawBox Spec Sheet

If you want something that meets the production tier with zero integration effort:

  • Silicon: NVIDIA Jetson Orin Nano 8 GB Super (67 TOPS, 1024-core Ampere GPU)
  • RAM: 8 GB LPDDR5 unified
  • Storage: 512 GB NVMe SSD pre-installed
  • Power: 15 W peak, 5–8 W idle
  • Network: 1 Gbps Ethernet + Wi-Fi 6
  • OS: Ubuntu 22.04 (aarch64) with NVIDIA JetPack, CUDA, TensorRT pre-configured
  • Software: OpenClaw pre-installed and tuned to this silicon
  • Setup time: 5 minutes (QR code → chat app → done)
  • Warranty: 2 years
  • Ships to: 108 countries (free EU shipping via DHL Express)
  • Price: €549 one-time

Order ClawBox — the official OpenClaw hardware →

Checklist: "Will my hardware run OpenClaw well?"

Print this and tick boxes before you commit to a build.

  • ≥ 8 GB unified RAM (not system + discrete-GPU split)
  • ≥ 20 TOPS dedicated AI acceleration
  • NVMe SSD for OS and model weights
  • Power supply rated for sustained peak (not just idle)
  • Cooling tested under 15-minute inference load without throttle
  • Wired Ethernet available
  • Power envelope ≤ 25 W if running 24/7 (keeps electricity cost sane)
  • CUDA 12+ (for Jetson-based) or Metal (for Mac) or OpenVINO (Intel)
  • A backup plan for when microSD dies (if you're using one)

If you can't tick most of these, either drop down to the "minimum" tier or skip the DIY route — the official OpenClaw hardware ships with every line already ticked.

FAQ

Can I run OpenClaw on a Raspberry Pi?

Yes — a Pi 5 with 8 GB RAM meets the minimum tier. Add a Hailo-8L AI HAT and NVMe M.2 HAT and you reach the recommended tier, but with a 13-TOPS ceiling. For production OpenClaw usage, pick Jetson Orin Nano 8 GB or ClawBox.

Does OpenClaw need a dedicated GPU?

No, but you want dedicated AI acceleration. Integrated NPUs (Jetson, Apple Neural Engine) beat discrete desktop GPUs on power efficiency for edge inference. An RTX 3060 will run OpenClaw fine but pulls 170 W — 11× the ClawBox envelope.

What about an old desktop PC?

Works if it has 16 GB+ RAM and a reasonably modern CPU. Expect 40–80 W idle, which adds up on 24/7 deployments. The TCO math almost always favors purpose-built edge hardware over a repurposed tower.

Will these requirements change in 2027?

Probably yes — as OpenClaw ships larger local models, the 8 GB unified memory bar will feel tight. Buy for the recommended or production tier if you want 3+ years of headroom.


Guide maintained by the team behind ClawBox — the official OpenClaw hardware. Updated roughly every 60 days as silicon and OpenClaw versions evolve.


Read the full guide on ClawBox: OpenClaw hardware requirements — full guide →

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment