TL;DR
On October 7, 2026, Microsoft and NVIDIA unveiled the RTX Spark computing platform: a single ARM-based SoC integrating CPU, GPU and up to 128GB unified memory, delivering up to 1 PFLOPS of FP4 AI performance for running 100B+ parameter models locally, at the core of the next-generation Windows AI PC.
Background
At NVIDIA GTC and Microsoft Build in June 2026, the two companies announced plans to 'reinvent the PC'; the Windows & Surface special event on October 7, 2026 (local time) was the second half of that collaboration, with NVIDIA CEO Jensen Huang attending (per ifanr citing The Verge). Microsoft and NVIDIA began working together in early 2024, and RTX Spark is the platform NVIDIA built for Microsoft's personal-AI vision (per The Verge, via ifanr).
The Core Architecture of RTX Spark
RTX Spark is an ARM-based SoC that integrates CPU, GPU and up to 128GB unified memory in one package, not a standalone GPU. Similar in approach to Apple Silicon, RTX Spark lets the CPU and GPU share one memory pool, avoiding data transfers between VRAM and system memory. NVIDIA also partnered with MediaTek to bring a customized Grace CPU to mainstream Windows devices for the first time (per The Verge, via ifanr).
On-Device AI Capability
RTX Spark delivers up to 1 PFLOPS of FP4 AI performance and can run 100B+ parameter models locally. Official materials state the platform can run models such as Qwen 3.8 Flash Next with 125 billion parameters on-device, with native support for NVIDIA's CUDA, DLSS and ray tracing stack (per official materials, via ifanr). In the age of AI agents, RTX Spark is positioned as a personal compute node that can independently carry agent workloads (per ifanr).
The Flagship Device: Surface Laptop Ultra
Surface Laptop Ultra is the first flagship device built on RTX Spark, with a CNC aluminum unibody. Hardware includes a 15-inch 3:2 Mini-LED high-refresh display (262 PPI), 3x USB4, HDMI, USB-A, SD slot and 3.5mm headphone jack; the chassis is under 18mm thin and weighs about 2kg. Microsoft also introduced Magnetic Connect, a magnetic USB-C charging design the company calls a world first (per ifanr).
A New Division of Labor: Windows and Copilot
Microsoft frames the next-generation Copilot+ PC as a dual-OS story: Windows manages the base layer, while Copilot Agent becomes a task operating system. The goal is to bind coding, Office work, video and photo editing, and personal file management to local Copilot Agents. Unlike the earlier Copilot+ PC definition built on low-power NPU small models, RTX Spark follows a high-compute local-model route, using near-zero marginal-cost local compute for high-frequency AI calls (per ifanr).
Common Misconceptions
- Misconception 1: RTX Spark is a graphics card. In fact it is an integrated SoC combining CPU, GPU and unified memory (per official materials).
- Misconception 2: it is the same as Qualcomm Copilot+ PCs. The Qualcomm route centers on low-power NPU small models, while RTX Spark centers on high-compute local GPU models (per ifanr).
Summary
RTX Spark is the core product of the Microsoft-NVIDIA collaboration in 2026: a unified-memory ARM SoC carrying 1 PFLOPS of FP4 performance, bringing a customized Grace CPU into mainstream Windows devices for the first time. Its direct reference point is the Apple Silicon-based MacBook Pro; the difference is RTX Spark's native support for CUDA, DLSS and G-SYNC (per ifanr).
Sources
- ifanr (October 8, 2026): Jensen Huang unveils the most powerful Surface yet, built for unlimited intelligence (citing The Verge)
- Tech Bytes (October 8, 2026): Microsoft launched Nvidia-chip AI PCs at its Surface Laptop Ultra event