If you want a local AI assistant running in your home — something that reads your messages, automates browser tasks, and answers questions without sending your data to someone else's cloud — the two boards everyone considers first are the Raspberry Pi 5 and the NVIDIA Jetson Orin Nano. They look similar on the shelf: small, quiet, low-power single-board computers. For running actual AI models, they are very different machines.
Here's the honest comparison, without marketing spin.
TL;DR
- Raspberry Pi 5 is the better board for general home-server duty: cheaper, huge community, runs Home Assistant, Pi-hole, media servers and small services beautifully. It can technically run small language models — slowly.
- Jetson Orin Nano (Super, 8GB) is the better board for local AI: its GPU runs 4–8B parameter models at conversational speed, which is the difference between an assistant you actually use and a demo you show once.
Price
The Pi 5 wins on sticker price: roughly €90–130 for the 8GB/16GB board, before you add a decent power supply, NVMe HAT, SSD and case — realistically €150–200 for a complete, reliable setup.
A Jetson Orin Nano Super developer kit typically lands around €270–330, plus NVMe storage and a case — realistically €350–420 complete. So the Jetson costs roughly twice as much once both are actually built out.
If the budget is the whole story, the Pi wins. But for AI workloads, price-per-token tells a different story.
Running local LLMs: the part that actually matters
This is where the two boards stop being comparable.
The Pi 5 runs language models on its four ARM CPU cores. With a quantized 7–8B model (think Llama 3.1 8B or Qwen 2.5 7B at Q4), commonly reported speeds are in the 2–4 tokens/second range, and prompt processing on longer contexts is painfully slow. Small 1–3B models are more usable, but their answers are noticeably weaker for assistant-style work. There's no CUDA, so the broader GPU-accelerated AI ecosystem simply isn't available.
The Jetson Orin Nano Super runs the same models on a 1024-core Ampere GPU (67 TOPS) with CUDA and TensorRT support. Quantized 4–8B models typically run at 10–20 tokens/second — fast enough that chatting with a local model feels like chatting, not like waiting for a fax. The 8GB of unified memory comfortably fits a quantized 8B model plus a vision or speech model alongside it.
That 4–5× speed gap is the practical difference between "my assistant answered" and "I gave up and opened ChatGPT."
Power draw
Both are excellent here, and both embarrass any desktop:
- Pi 5: ~3–5W idle, ~8–12W under load
- Jetson Orin Nano: ~7W idle, 15–25W under sustained AI load
Even at full tilt, the Jetson draws less than a lightbulb. Running 24/7 at average load, either board costs a few euros per month in electricity.
Setup effort
The Pi is famously easy for classic server stuff — flash an SD card, boot, done. The pain starts when you try to make AI work well on it: you'll spend evenings compiling llama.cpp flags and accepting compromises.
The Jetson is the opposite: more initial friction (JetPack SDK, NVIDIA's flashing process), but once it's up, the AI tooling actually works — CUDA builds of llama.cpp and Ollama, TensorRT, hardware-accelerated speech and vision.
Budget 10–20 hours to take either board from unboxing to a polished, always-on AI assistant with messaging integrations, browser automation and backups — most of it software setup, not hardware.
So which one?
- You want a home server that occasionally touches AI → Raspberry Pi 5.
- You want a local AI assistant that's actually pleasant to use every day → Jetson Orin Nano.
The shortcut
If the Jetson is the right answer but 10–20 hours of JetPack, drivers and integration glue is not how you want to spend your weekends, that's exactly why we build ClawBox: a Jetson Orin Nano Super (8GB, 67 TOPS) with a 512GB NVMe SSD and the open-source OpenClaw assistant pre-installed and pre-configured — €549, plug in, scan a QR code, and your assistant is running on hardware you own. Local-first, with the option to bring your own cloud AI key (OpenAI, Anthropic) when you want a bigger brain.
Either way you go: running your own AI at home stopped being a science project. It's now just a choice of how much speed you want — and how much setup you enjoy.