The NVIDIA Jetson Orin Nano is the most capable edge AI platform you can buy for under $300. With 67 TOPS of AI compute, 1024 CUDA cores, and just 15W power draw, it's purpose-built for always-on AI workloads.
But turning a bare Jetson module into a functional AI assistant takes work. This guide covers everything — from choosing your hardware to having a fully working AI assistant on Telegram, WhatsApp, or Discord.
Why the Jetson Orin Nano for an AI Assistant?
Before diving into the build, let's talk about why the Jetson Orin Nano is the best platform for this:
67 TOPS of real AI compute. Not theoretical benchmarks — actual inference performance with NVIDIA's CUDA and TensorRT stack. For comparison, a Raspberry Pi 5 offers roughly 2 TOPS (CPU-only), and even the Mac Mini M4's Neural Engine tops out at 38 TOPS.
15 watts. Running 24/7 costs about €3/month in electricity. A gaming PC doing the same job would cost €50-80/month. For an always-on assistant, power efficiency isn't a nice-to-have — it's essential.
Full NVIDIA CUDA support. Every major AI framework (PyTorch, TensorFlow, ONNX Runtime, TensorRT) runs natively. No workarounds, no CPU fallbacks, no "it works but slowly."
8GB unified memory. Shared between CPU and GPU, so you can load 7-8B parameter models comfortably. That's Llama 3.1 8B, Mistral 7B, CodeLlama 7B — all running at conversational speeds.
What You'll Need
Hardware
| Component | Recommended | Budget Option |
|---|---|---|
| Compute Module | Jetson Orin Nano Super 8GB (~$250) | Jetson Orin Nano 4GB (~$200) |
| Storage | 512GB NVMe SSD (~$40) | 256GB NVMe SSD (~$25) |
| Case | Any passive cooling case ($20-50) | 3D printed enclosure |
| Power Supply | 20W USB-C PD ($15) | Official NVIDIA PSU |
| Networking | Built-in WiFi 6 + Ethernet | Same |
Total DIY cost: $325-355
Software
- Ubuntu 22.04 (via JetPack 6.0)
- Node.js 20 LTS
- OpenClaw (open-source AI assistant platform)
- Chromium + Playwright (for browser automation)
- Whisper (speech-to-text, optional)
- Kokoro TTS (text-to-speech, optional)
Step 1: Flash JetPack OS
Download NVIDIA JetPack 6.0 from the NVIDIA developer site. Flash it to your NVMe SSD using the SDK Manager or balenaEtcher.
# After booting, update everything
sudo apt update && sudo apt upgrade -y
JetPack 6.0 includes CUDA 12.2, cuDNN 8.9, and TensorRT 8.6 — everything you need for AI inference right out of the box.
Step 2: Install Node.js
OpenClaw requires Node.js 18+. Install the LTS version:
curl -fsSL https://deb.nodesource.com/setup_20.x | sudo -E bash -
sudo apt install -y nodejs
node --version # Should show v20.x
Step 3: Install OpenClaw
OpenClaw is the open-source AI assistant platform that ties everything together — messaging, browser automation, scheduling, memory, and AI model management.
# Install OpenClaw
npm install -g openclaw
# Run the setup wizard
openclaw setup
The setup wizard walks you through:
- Choosing your AI provider (Anthropic Claude, OpenAI GPT, Google Gemini, or local models)
- Connecting messaging platforms (Telegram, WhatsApp, Discord)
- Configuring browser automation
- Setting up voice pipeline (optional)
Step 4: Connect Telegram
The fastest way to interact with your AI assistant:
- Open Telegram and search for
@BotFather - Send
/newbotand follow the prompts - Copy the bot token
- Paste it into the OpenClaw setup wizard
That's it — send a message to your bot and your Jetson-powered AI responds.
Step 5: Set Up Browser Automation
This is where the Jetson really shines. Browser automation lets your AI assistant:
- Search the web and summarize results
- Fill out forms automatically
- Monitor websites for changes
- Take screenshots on command
- Extract data from any webpage
# Install Chromium and Playwright
npx playwright install chromium
OpenClaw handles the integration — just ask your AI to "search for flights to Rome" or "check if this product is in stock" and it opens a browser, navigates, and reports back.
Step 6: Voice Pipeline (Optional)
For a truly hands-free experience, set up on-device voice processing:
# Whisper for speech-to-text (runs on GPU)
pip3 install faster-whisper
# Kokoro for text-to-speech
pip3 install kokoro-tts
Both run locally on the Jetson's GPU tensor cores — your voice never leaves the device. Support for 90+ languages, and the latency is low enough for natural conversation.
Step 7: Configure for 24/7 Operation
Your AI assistant should run always-on:
# Create systemd service
sudo systemctl enable openclaw
sudo systemctl start openclaw
# Verify it's running
openclaw status
Set up automatic recovery so it restarts after power outages or crashes. The Jetson Orin Nano's 15W draw means you can run it on a small UPS for hours of battery backup.
What Can Your Jetson AI Assistant Actually Do?
Once everything is set up, here's what's possible:
Daily assistant tasks:
- "Summarize my unread emails"
- "What's on my calendar today?"
- "Set a reminder for 3 PM"
- "Find the cheapest flight from Berlin to London next weekend"
Browser automation:
- "Monitor this eBay listing and tell me if the price drops"
- "Fill out this job application form with my details"
- "Take a screenshot of my website's homepage"
Smart home integration:
- "Turn off the living room lights"
- "Set the thermostat to 20°C"
- Works with Home Assistant, MQTT, and REST APIs
Development tasks:
- "Review this pull request"
- "Run the test suite and tell me what failed"
- "Help me debug this Python script"
Performance Benchmarks
Real-world performance on the Jetson Orin Nano Super 8GB:
| Task | Speed | Notes |
|---|---|---|
| Llama 3.1 8B inference | ~15 tok/s | Conversational speed |
| Whisper speech-to-text | Real-time | No perceivable delay |
| Kokoro TTS | Real-time | Natural-sounding output |
| Browser page load | 2-5 seconds | Full Chromium rendering |
| Web search + summary | 10-15 seconds | End-to-end |
For tasks requiring frontier model intelligence (complex reasoning, large context windows), OpenClaw supports BYOK (Bring Your Own Key) — route to Claude, GPT, or Gemini APIs while keeping your data, memory, and automation local.
The DIY vs Pre-Built Decision
Building from scratch takes 3-4 hours and costs $325-355. You learn a lot and have full control.
If you'd rather skip the setup, ClawBox is a pre-built Jetson Orin Nano AI assistant that comes with everything pre-configured — OpenClaw, browser automation, voice pipeline, carbon fiber case, 512GB NVMe — ready in 5 minutes. It's €549, which includes the case, power supply, larger SSD, and the time you'd spend on setup.
Either way, the Jetson Orin Nano is the best platform for a personal AI assistant in 2026. Nothing else gives you 67 TOPS at 15W with full CUDA support.
Troubleshooting Common Issues
"Model loading is slow" — Make sure you're using the NVMe SSD, not a microSD card. NVMe is 10-50x faster for model loading.
"Browser automation crashes" — Increase swap space: sudo fallocate -l 4G /swapfile && sudo mkswap /swapfile && sudo swapon /swapfile
"WiFi keeps disconnecting" — Use Ethernet for always-on operation. WiFi power management can cause drops. Disable it: sudo iw dev wlan0 set power_save off
"Out of memory with larger models" — Stick to 7-8B parameter models. For larger models, use BYOK to route to cloud APIs.
What's Next?
Once your Jetson AI assistant is running:
- Install skills from ClawHub — 580+ community-built skills for everything from weather to stock trading
- Set up proactive monitoring — have your AI check your email, calendar, and websites automatically
- Connect multiple platforms — use one AI across Telegram, WhatsApp, and Discord simultaneously
- Build custom skills — Python-based skill system lets you automate anything
The Jetson Orin Nano isn't just a development board — it's the foundation of a genuinely useful, always-on, private AI assistant. At 15W and $250, there's nothing else like it.