Nvidia's RTX Spark puts an AI agent inside the PC

Nvidia announced the RTX Spark, a CPU and GPU package the company calls a "superchip," with a stated 1 petaflop of throughput aimed at running an AI agent locally rather than shipping every request to the cloud. ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI confirmed Spark-based PCs for later this year, with Acer and Gigabyte to follow. Microsoft is branding its version the Surface Laptop Ultra. Jensen Huang pitched the move as a play for a $200 billion CPU market that has belonged to Intel and AMD.

The hardware side is the obvious headline, but the more interesting commitment is on the software side. Nvidia and Microsoft co-developed a sandbox layer for agent execution, which is the same problem Anthropic and OpenAI are spending serious engineering effort on for their own agent products. Spark ships with enough CPU, GPU, and RAM headroom to run a local model that drives the agent, so a desktop user can hand off a task and let it finish without the network round-trip or the privacy cost of streaming screens to a cloud. Over 100 software vendors signed on at launch, including Adobe, Blender, Riot Games, and Xbox.

Huang said the goal is "billions of agents" running across PCs, and the launch is consistent with that ambition. The market test is whether a typical Windows buyer will pay the Spark premium for a capability they have so far only seen in chat windows.

Why it matters

If you ship software that runs on Windows, the assumption that the user's agent lives in a browser tab is about to break. Plan for a local sandbox that can read your app's state, request actions, and report back, because the OEMs are now shipping the hardware that makes that the default.