NVIDIA and Microsoft Pair RTX Spark With Secure Agent Containers to Put Always-On AI Agents on Windows PCs

On Wednesday, October 7, Microsoft held a Windows AI and Surface event at Dogpatch Studios in San Francisco. NVIDIA founder and CEO Jensen Huang and Microsoft CEO Satya Nadella took part in a fireside chat hosted by Sriram Krishnan, a former senior White House policy advisor on AI. The message was clear: the two companies are co-engineering hardware and software so that AI agents can run on Windows PCs. Huang traced NVIDIA's history back to Windows and said that without Windows there would be no GeForce. He then linked that history to the agentic era. When an agent lives on your own computer, he argued, it becomes your personal assistant. Nadella credited Huang with a consistent vision held over decades, which he described as a secular trend that explains why this moment has arrived.

The most important technical element of the event is the security foundation Microsoft described. Pavan Davuluri, Microsoft's executive vice president of Windows and Devices, announced general availability of Microsoft Execution Containers, or MXC. It is operating-system-level infrastructure that lets agents run safely and persistently in the background, under OS control. Davuluri said Microsoft is building these primitives directly into Windows so that agents can be secured, observed and governed, working together with Microsoft Security and Agent 365. Nadella added that the desktop must become the most secure place for agents to execute. Huang compared the shift to earlier platform moments: just as Windows and DirectX changed how applications were built, he said, MXC will change how agents are built and deployed. The reasoning is sound. An agent differs from a classic application because it calls tools on its own, reads and writes files, and stays resident for long periods. For such software, permission boundaries and audit trails are a precondition, not an optional extra.

On the hardware side, the lead product is RTX Spark. NVIDIA's blog describes a Blackwell RTX GPU with up to 6,144 cores and a Grace CPU with up to 20 cores, connected at 600 GB/s. The package offers one petaflop of FP4 AI performance and up to 128GB of unified memory. Davuluri said Microsoft built Surface Laptop Ultra around RTX Spark, and that with that memory and compute a user can run models that simply do not fit on a traditional machine. NVIDIA's example is Qwen 3.8 Flash Next, described as a 125B model with 51B n-gram that matches the intelligence of many cloud models, unmetered and without sending data to the cloud. Readers should treat these comparisons with care. They come from the vendor, and real results will depend on quantization, context length and the task at hand. Independent benchmarks should settle the question. Systems are planned from Acer, ASUS, Dell, HP, Lenovo, Microsoft, MSI and Gigabyte, in forms from slim laptops to compact desktops built for always-on agents. Laptop preorders opened on the day of the event, with availability on October 16. Compact desktops go on sale in November.

For developers, the unified software stack may be the most valuable part. RTX Spark runs the full NVIDIA CUDA platform, so developers can move models, tools and workflows without rewriting them, using the same NVIDIA AI stack from RTX Spark to DGX Station. Creators get fifth-generation Tensor Cores with NVFP4 support, hardware-accelerated AV1, and 4:2:2 video encode and decode. Gamers are promised AAA titles at 1440p above 100 frames per second with DLSS 5, Reflex and G-SYNC. The compact desktop uses the same RTX Spark superchip in a small chassis designed for 24/7 operation, a dedicated local AI system that keeps agents running continuously. The event also previewed NVIDIA DGX Station for Windows, called the first deskside AI supercomputer built for the Windows enterprise desktop. It brings GB300 Grace Blackwell-class AI infrastructure into the Windows ecosystem. Davuluri said it unlocks the power to run frontier-class models locally, capabilities that once required renting a cluster. The blog notes that DGX Station previously ran on Linux, so Windows enterprise environments can now use this class of machine directly.

Taken together, the significance of this announcement lies less in any single number and more in the fact that a complete chain was shown at once: OS-level isolation and governance, local compute with large unified memory, and one CUDA stack from laptop to deskside supercomputer. For enterprise IT, keeping data on the device and paying per machine rather than per call changes both the cost structure and the compliance discussion. For developers, a path of build once and deploy across tiers lowers migration cost. The risks are equally clear. Agents that stay resident on local machines move the attack surface from cloud APIs to every endpoint. The isolation strength, default policies and auditability of MXC will need to be tested by the security community in real deployments. Pricing, battery life and thermal behavior of the new hardware were not disclosed in the source and await hands-on testing. Finally, whether local models truly close the gap with frontier cloud models, as the vendors suggest, is a question for independent benchmarks. The next points to watch are the first reviews after laptops ship on October 16, the pace of MXC adoption in enterprises, and whether developers actually move agent workloads onto local machines.