NVIDIA at Bilibili World made its first showing of a laptop equipped with the RTX Spark superchip: Blackwell RTX GPU + 20-core Grace CPU directly connected via NVLink-C2C, 1 Petaflop FP4 compute paired with 128GB unified memory, able to run a 120B-parameter LLM locally with context stretched to 1M tokens — a configuration previously only encountered on cloud inference clusters, now stuffed into an ultra-thin laptop. The accompanying software stack is equally critical: the OpenShell runtime breaks Agent permissions into declarative policies, NemoClaw is responsible for keeping sensitive data local; on-site a 35B Qwen multimodal model drives a personal Agent that, after recognizing hand-drawn sketches, can replicate a complete webpage locally in tens of seconds, with zero cloud tokens burned throughout. The "hardware + runtime + privacy layer" trio signals NVIDIA's upgrade of Agent competition from "selling cards" to "selling local Agent compute platforms". Also unveiled at the same time, the DGX Spark desktop supercomputer has nearly identical specs (128GB / 1 Petaflop), but is Linux-based with the full NVIDIA AI stack pre-installed, and two of them can interconnect via ConnectX to handle a 200B model — making it real for independent developers to move training, fine-tuning, and inference from the cloud back to the desk.