Cloud AI is still facing questions about how its revenue is measured, while desktop AI continues to gain momentum. At an event in San Francisco, Microsoft and Nvidia shifted the focus from “being able to chat” to “being able to run agents locally.” Nvidia’s RTX Spark connects its Blackwell GPU and Grace CPU, and is said to deliver around one petaflop of local AI compute. Microsoft, meanwhile, is making Execution Containers a general-purpose capability in Windows 11, giving agents an isolated runtime environment. Tools such as coding assistants have already integrated with it, and more vendors are expected to follow.
In addition to the previously announced Surface Laptop Ultra, the two companies also previewed the DGX Station for Windows, bringing data-center-class GB300 computing power to the desktop, with unified memory reaching hundreds of gigabytes. It is designed for heavy training workloads and private, on-premises deployment. Microsoft also updated Copilot: with user permission, the assistant can access local files and recent activity to handle tasks such as organizing and troubleshooting. The direction is clear: a hybrid approach in which lightweight tasks run locally and heavy workloads run in the cloud—reducing latency while partly avoiding cloud costs and privacy concerns.
This does not contradict the controversy over OpenAI’s revenue: the monetization efficiency of large cloud-based models is being repriced, but the advantages of on-device inference, privacy, and lower latency could instead accelerate hardware replacement. For NVIDIA, this is a second growth curve beyond data centers; for the PC supply chain and memory makers, it could also mean higher prices and volumes.
My view: On-device AI is a genuine trend, but it sits in a relatively high price bracket, so in the short term it looks more like a fit for professional developers and enterprise pilots. From an investment perspective, I remain bullish on GPUs, memory, and high-quality PC brands. At the same time, don’t overlook which software players can turn AI agents into a truly daily-use product.
Beyond getting the hardware into users’ hands, the software ecosystem will be the deciding factor. Whoever can integrate local models, permission management, and enterprise IT workflows will capture the software profits from the next replacement cycle. Early developer feedback and the model sizes that can actually run matter more than product launch events.
For enterprise customers, on-premises deployment can also meet compliance and data residency requirements—selling points that purely cloud-based solutions are hard-pressed to match.
The above content is for reference only and does not constitute investment advice.