NVIDIA AI at Home.
On Your Own Hardware.

ClawBox is built on the NVIDIA Jetson Orin Nano Super. It delivers 67 TOPS of AI compute from a 1024-core NVIDIA Ampere architecture GPU with 32 Tensor Cores, paired with a 6-core Arm CPU and 8GB of LPDDR5 memory. This is the same NVIDIA accelerated-computing stack — CUDA, cuDNN, and TensorRT — that powers NVIDIA's data-center AI, shrunk into a low-power device you can keep at home.

AI inference is massively parallel matrix math, exactly what NVIDIA GPUs are built for. The Jetson Orin Nano Super's 1024 CUDA cores and 32 Tensor Cores run language models, speech recognition, and image models far faster and far more efficiently than a CPU. With CUDA and TensorRT, ClawBox accelerates models like Llama 3.1 8B and Whisper to real-time speeds while sipping power — no CPU-only box at this price comes close.

CUDA is NVIDIA's parallel-computing platform that lets AI frameworks (PyTorch, llama.cpp, Ollama) run on the GPU. TensorRT is NVIDIA's inference optimizer that fuses layers, applies INT8/FP16 quantization, and squeezes maximum throughput out of the hardware. Because ClawBox uses the standard NVIDIA software stack, the entire open-source AI ecosystem runs out of the box — you are not locked into a proprietary accelerator.

A cloud GPU (an NVIDIA A100 or H100) is far more powerful, but you rent it by the hour, your data leaves your network, and the bill never stops. The Jetson Orin Nano Super in ClawBox is plenty for running 7–8B models locally, costs €549 once, draws about 15W, and keeps every query on your own hardware. For personal and home AI, local NVIDIA hardware wins on cost and privacy.

Yes. 67 TOPS comfortably runs Llama 3.1 8B at around 15 tokens/second, Mistral 7B, Phi-3, Gemma, Whisper Large v3 for speech-to-text in 90+ languages, and Kokoro TTS for natural voice output — all locally, all GPU-accelerated. The 'Super' revision nearly doubles the AI performance of the original Orin Nano, making it one of the most capable NVIDIA edge-AI platforms for the money.