Jensen Huang and Satya Nadella used a joint appearance at a Microsoft event in San Francisco on Wednesday to map out a strategy for running AI agents natively on Windows PCs, anchored by NVIDIA's RTX Spark chip and its one-petaflop FP4 AI performance in a laptop form factor S¹. The pitch breaks with the assumption that serious AI inference belongs in cloud data centres, and bets that every Windows desktop becomes a host for persistent agents.
My read: This is the tightest NVIDIA-Microsoft alignment on local agent infrastructure I've seen, and the general availability of Microsoft Execution Containers is the part worth watching most closely. Hardware specs are one thing, but OS-level execution containers for persistent background agents address the missing piece that has kept local agents from being useful on Windows. I don't buy Huang's claim that MXC will "revolutionise how agents are built" yet, because we have no third-party benchmarks and no shipped agent workloads to test against. October 16 is the date that matters: that is when the first laptops arrive and a petaflop of FP4 can be measured against real agent tasks.
MXC gives agents a controlled space inside Windows
The concrete software announcement is Microsoft Execution Containers, or MXC, now generally available S¹. Pavan Davuluri, Microsoft's EVP of Windows and Devices, called it operating-system-level infrastructure designed to let agents operate safely and continuously in the background while remaining under OS supervision S¹. Nadella was blunt: "We needed to make the desktop the most secure place for agents to execute" S¹.
Huang compared MXC's potential to what Windows and DirectX did for applications S¹. That is a promotional claim from a CEO on a corporate stage, and nobody outside NVIDIA and Microsoft has tested MXC with production agent workloads. No independent benchmarks have been published, and no third-party agent code has been run through MXC publicly.
RTX Spark pairs Blackwell and Grace for laptop-scale inference
RTX Spark pairs an NVIDIA Blackwell RTX GPU with as many as 6,144 cores and an up to 20-core NVIDIA Grace CPU, linked at 600 GB/s S¹.
The chip is rated at one petaflop of FP4 AI performance. FP4 is a low-precision number format that sacrifices some accuracy for faster inference — the process of actually running a trained model S¹. NVIDIA's petaflop figure is the company's own; no independent testing has confirmed sustained FP4 throughput on production silicon.
Microsoft's Surface Laptop Ultra, introduced by Davuluri, is built around RTX Spark and offers as much as 128GB of unified memory — meaning the CPU and GPU draw from the same RAM pool, alongside up to a petaflop of AI compute S¹. Those are maximum configurations, not baseline specs.
RTX Spark laptop preorders opened the day of the event, with shipping on October 16 S¹. Compact desktops follow in November S¹. Systems are coming from Acer, ASUS, Dell, HP, Lenovo, Microsoft, MSI and Gigabyte S¹.
NVIDIA first showed RTX Spark at COMPUTEX in Taipei earlier this year. The October event shifts the focus from gaming to agentic computing. NVIDIA has also been expanding cloud-side capacity, but this event pushes compute in the opposite direction, onto the device. The day before, OpenAI was reported to be training agents on Ironclad contracts — a reminder that the most capable agent workloads still run on server-grade hardware.
For a machine-learning engineer who currently routes agent inference through a cloud API, RTX Spark offers a local alternative: running those workloads on a laptop inside MXC, with no per-query cloud cost and no network latency. Whether the petaflop figure holds up under sustained agent loads remains unverified. The first laptops ship October 16, which is when MXC can be tested on real silicon rather than on a stage.
Sources: S1 — NVIDIA, Microsoft Kick Off a New Beginning for Windows PCs With RTX Sp · S2 — NVIDIA, Microsoft Kick Off a New Beginning for Windows PCs With RTX Sp · P3 — NVIDIA and Microsoft Reinvent Windows PCs for the Age of Personal AI | · P4 — NVlabs/OmniVinci · P5 — Faster Local AI Agents on RTX PCs and DGX Spark | NVIDIA Blog · P6 — Satya Nadella
Written from 6 sourced items, 5 of them primary.
