Jensen Huang's line yesterday at Computex deserves a closer read: "For forty years, you launched apps. Click. Type. With RTX Spark and Microsoft Windows, you ask—and the PC does the work."
True—for those whose PC can do that work.
A few months ago, I wrote about the emerging battle over the runtime where agents live. Apple had already established a strong position there—M-series chips, unified memory architecture, and a Unix foundation that makes local code execution native and frictionless: a terminal, a package manager, a local runtime, and you're running a model. RTX Spark is NVIDIA's answer on the Windows side: 128GB unified memory, 1 petaflop of AI compute, 120-billion-parameter models running locally. The runtime battle just got a new, powerful player.
But there's a gap worth naming. RTX Spark requires machines that can execute code, run local inference engines, operate with enough rights to spin up an agent runtime. That description does not match the managed Windows laptop most enterprise employees use—provisioned through device management, locked by policy, on a multi-year refresh cycle. It's the rational outcome of decades of security thinking built around one assumption: software equals approved applications, centrally deployed.
But agents are not applications.
Microsoft is clearly aware of this—new device APIs sketch a path toward managed local inference. But sketching a path and having it ready for a fleet of tens of thousands of devices are very different things.
NVIDIA actually ships two distinct products. RTX Spark for developers and power users. A shared local compute node for enterprise teams, managed by IT. That distinction is the real story.
What Computex perhaps signals is the simultaneous erosion of two established models: the homogeneous corporate fleet—same machine, same image, same rights for everyone—and the all-cloud enterprise AI stack, where every query defaults to a hyperscaler's hosted model.
Local inference at this scale makes a third path architecturally serious, not just philosophically attractive.
What all of this points toward is distributed AI: central models for frontier tasks, edge inference for sovereign and sensitive ones, interoperability protocols making the routing seamless and governable. MCP is one early signal—and its most significant update since launch is telling: stateless transport for distributed deployments, async task management for enterprise workflows, distributed tracing for governance. A dedicated enterprise compute node is another signal. RTX Spark, whatever its current enterprise limitations, is a third.
For large organizations, the harder challenge is composability—building a stack that can route intelligently across all three layers without being captive to any of them.
