128GB of LPDDR5x unified memory (273GB/s bandwidth) is shared coherently between CPU and GPU — the whole point of this design, letting a single compact unit hold and run large AI models locally that would otherwise need a full multi-GPU server, for prototyping, fine-tuning, and inferencing models up to roughly 200 billion parameters standalone. It ships with 4TB of self-encrypting NVMe storage and a ConnectX-7 200Gbps SmartNIC that lets two DGX Spark units link together to handle models up to around 405 billion parameters. Despite that horsepower, the whole unit measures just 150 x 150 x 50.5mm and weighs 1.2kg — genuinely desk-sized — and runs NVIDIA's DGX OS with the full NVIDIA AI software stack preinstalled, aimed at developers who want local, private AI development without cloud GPU rental costs.