Microsoft Overhauls Windows for On-Device AI Agents, Unveils Surface Laptop Ultra With Nvidia Chip

The Surface Laptop Ultra starts at $2,599.99 and carries an Nvidia RTX Spark Superchip, which combines a Blackwell RTX GPU, a Grace CPU and unified memory on one platform, according to the Chinese technology outlet iFanr. Microsoft says the machine offers up to 128GB of unified memory and as much as 1 petaflop of AI performance.

Microsoft presented its own benchmarks against Apple's hardware, saying the Surface Laptop Ultra generates its first token 2.1 times faster than a 16-inch MacBook Pro with an M5 Pro chip, produces AI images 4.3 times faster and generates AI video 6.2 times faster. Those figures come from Microsoft and have not been independently verified.

The notebook is under 18mm thick and weighs less than 4.5 pounds, with a 15-inch PixelSense Ultra touchscreen. Microsoft says its cooling system holds up to 2.5 times the heat capacity of current Surface Laptop models. It includes three USB-C ports, HDMI, USB-A, an SD card reader and a headphone jack, supports up to three external 4K displays and has a user-replaceable SSD. Preorders opened October 7 and shipments begin October 16.

Microsoft has dropped the Copilot+ PC branding from the new hardware, a company spokesperson told foreign media, saying naming has been simplified around the Surface brand while the company continues to offer high-end hardware and hybrid AI experiences.

A second device, the Surface RTX Spark Dev Box, is a small desktop development machine priced at $5,999 that ships in November. It comes with Visual Studio Code, Git, GitHub CLI, GitHub Copilot, WSL, Python and Node preinstalled. iFanr reported its main specifications as an Nvidia RTX Spark N1X with a 6,144-core GPU and 20-core CPU, 128GB of unified memory and a 2TB SSD.

Microsoft also extended RTX Spark into a Builder PC line. Preorders are open for the ASUS ProArt P16 and P14, Dell XPS 16 Creator, HP OmniBook Ultra 16, Lenovo Yoga 9n 2-in-1, MSI Prestige N16 Flip AI+ and the Surface Laptop Ultra, all shipping from October 16. For compact desktop machines, Microsoft prepared a native Windows gateway for OpenClaw and put the MXC sandbox into the out-of-box setup. At the high end, it announced DGX Station for Windows, built on an Nvidia GB300 Grace Blackwell Ultra Desktop Superchip with up to 748GB of unified memory and 20 petaflops of FP4 AI compute, which Microsoft says can run models with more than one trillion parameters locally.

On the software side, Copilot moves from a chat entry point into the system through three experiences: Home, which reads the files and recent activity a user is working with; Code, which builds native Windows applications from prompts and runs code in a more isolated environment; and Autopilot, which is intended to run continuously and act on tasks. The company said these capabilities depend on local context, local operations and local models, with heavier tasks routed to the cloud.

Windows Search will execute thousands of actions directly from the taskbar, including turning on dark mode or do-not-disturb, dimming the screen, arranging windows or sending a message, without opening Settings or other applications. Search will also connect to the new Copilot so users can get a quick answer in the taskbar before moving to the full application.

For privacy and control, Microsoft introduced Microsoft Execution Containers, or MXC, which restrict which files and networks an agent can reach and apply isolation at different levels, including process isolation, session isolation, virtual machines, WSL and Windows 365 for Agents. Codex, GitHub Copilot, OpenClaw, Replit, LM Studio, Nvidia OpenShell and Unsloth already support MXC, according to iFanr, while Claude Code, Box, Egnyte, Manus, Perplexity and Raycast are preparing to add support. Meta's Muse for Windows will integrate MXC as a native application.

Microsoft framed the launch around what it calls hybrid intelligence, routing tasks to local hardware when possible and to the cloud when needed. The local model lineup includes Microsoft's own MAI Code 1.1 Flash, with 137 billion total parameters and 6.8 billion active parameters, reduced in size by nearly 80 percent after 3-bit processing and supporting a 256K context window. Nvidia's upcoming Nemotron model, at more than 70 billion parameters, occupies just over 20GB of memory after 2-bit quantization, and DeepSeek V4 Flash carries 284 billion parameters. GitHub's HydraFusion routing has been extended from the cloud to Windows to decide per task whether local or cloud models are used, and Windows ML now supports llama.cpp, allowing models to be deployed across GPUs, NPUs and CPUs.

The launch follows two years in which Windows AI features drew criticism for amounting to little more than a Copilot entry point, iFanr reported, noting that open-source plugins for removing Windows AI features had collected more than 10,000 stars. Microsoft's previous AI PC push, launched in 2024 under the Copilot+ PC name, emphasized NPU compute rather than sustained local model and agent workloads.