The hardware story is the unified memory. RTX Spark pairs a Blackwell GPU with a 20-core Grace CPU and up to 128GB that the CPU and GPU address as one pool — so a 120B model loads without being copied across a narrow PCIe bus, the same trick that lets Apple's M-series hold weights a faster discrete card can't. NVIDIA quotes a petaflop of AI performance and a 1-million-token context for agents running entirely on the machine. For anyone building agentic apps, OpenShell is the part worth reading. Microsoft and NVIDIA describe it as a runtime that defines what an agent is allowed to do, decides when a request can be served by a local model, and masks personal data before anything leaves for the cloud. It builds on the late-May Windows preview and the Surface dev box shown earlier this month; this is the wider OEM launch, with laptops and small desktops due this fall from ASUS, Dell, HP, Lenovo, Surface and MSI, with Acer and GIGABYTE to follow.