NVIDIA and Microsoft made hardware and software announcements on Oct. 7 at the Windows AI and Surface event in San Francisco, aiming to expand local AI workloads on Windows PCs. Preorders for NVIDIA RTX Spark-based laptops opened the same day; the company said sales are planned to begin Oct. 16, while compact desktop versions are planned for November. These are announced dates; actual delivery and stock availability are separate stages.

NVIDIA says RTX Spark laptops will use a Blackwell RTX GPU, a Grace CPU with up to 20 cores, a 600 GB/s link and up to 128GB of unified memory. The company lists 1 petaflop of FP4 AI performance. This is a manufacturer-stated peak compute capability; real application speed can vary with the model, precision, software, power limits and thermal conditions. The announcement provides no independent comparative benchmark.

NVIDIA says RTX Spark can run the 125-billion-parameter Qwen 3.8 Flash Next model, including its 51-billion n-gram component, on device. The company says this can be done without sending data to the cloud and without per-token charges. Model and usage conditions matter: local execution does not eliminate the need for online services, updated models or cloud capacity for larger workloads. The claim of intelligence comparable to cloud models is the company’s characterization, not an independent measurement.

The manufacturers named are Acer, ASUS, Dell, HP, Lenovo, Microsoft, MSI and Gigabyte. RTX Spark laptop and compact desktop designs are intended for always-on local agents and developer workflows. NVIDIA also previewed DGX Station for Windows, a deskside AI supercomputer. For the GB300 Grace Blackwell Ultra-based system, the company lists 748GB of coherent memory and up to 20 petaflops of FP4 compute; it gave no price or firm sale date.

Microsoft announced general availability of Microsoft Execution Containers at the same event. The company describes this operating-system-level infrastructure as allowing agents to run persistently in the background under Windows control, with security and observability support. NVIDIA’s local AI devices and Microsoft’s execution controls together aim to create an on-PC agent ecosystem; security outcomes still depend on software configuration and permissions. The announcement gave no enterprise adoption rate or measured productivity result.

The product schedule illustrates AI compute moving beyond data centers toward developer and professional PCs. Local systems may offer data control and more predictable usage costs, while requiring hardware purchases, electricity, maintenance and upgrades. Opening preorders is not proof of completed sales or broad distribution. NVIDIA and Microsoft describe technical positioning and planned availability; actual performance, deliveries and cost advantages will require independent user measurements.