Nvidia’s CEO Jensen Huang unveiled the RTX Spark platform during Computex 2026, aiming to “reinvent the PC.” This ambitious initiative combines the GB10 Superchip with a specialized version of Windows on Arm, designed to facilitate the secure execution of local AI agents. Collaborating with industry giants such as Microsoft, Dell, HP, MSI, Lenovo, and Asus, Nvidia envisions a future where PCs serve as active, autonomous companions for both work and creativity.
Exploring the RTX Spark Platform
Recently, we ventured to San Francisco to delve deeper into the RTX Spark platform, returning with Microsoft’s Surface Laptop Ultra in its premium configuration for hands-on testing. In our comparative analysis, we also considered the 14″ MacBook Pro, equipped with Apple’s top-tier M5 Max SoC, featuring a formidable 40-core GPU and an 18-core CPU, along with 128 GB of LPDDR5X RAM. While we currently lack the mobile Ryzen AI Max+ 395 system for testing, we have access to AMD’s compact Ryzen AI Halo, which provides valuable insights through the latest ROCm stack and llama.cpp builds.
Local AI Performance
Assessing local AI performance is an evolving challenge, and our recent review reflects this complexity. Historically, we have focused on evaluating the prompt processing and token generation phases of AI inference separately, recognizing that speed in both areas enhances user experience. However, in the context of an agentic workflow, this method fails to encapsulate the total time required to complete tasks across multiple inference turns.
To address this gap, Artificial Analysis has introduced a new benchmark called AgentPerf-local, which accurately replays inference traces across various agentic AI tasks. This innovative approach emphasizes the importance of completion time: the quicker an AI agent accomplishes a task, the more productive the system becomes. Consequently, instead of merely presenting raw time-to-first-token and tokens-per-second metrics, we now share time-to-completion values akin to traditional CPU and content creation benchmarks.
It is important to note that local AI benchmarking is an ongoing process, and we welcome feedback on this new methodology as we refine our techniques. Currently, the supporting features promised by Microsoft and Nvidia for Windows 11 are still in development.
Despite being a first-generation mobile platform that integrates complex silicon IP from various design teams onto a single chip, the RTX Spark platform has demonstrated impressive maturity. My experience using the RTX Spark-powered Surface Laptop Ultra has been largely seamless, allowing me to accomplish my work much like I would on any other Windows laptop, aside from minor inconveniences. This is a commendable achievement, albeit accompanied by a steep price tag.
The most significant hurdle for the RTX Spark platform lies in its pricing. The launch systems come with eye-watering costs, making cloud alternatives from OpenAI and Anthropic seem more appealing. For instance, the Surface Laptop Ultra we tested retails for ,899 with just 1TB of storage, and lead times are already extending into late December.
For those less concerned about data privacy or cloud usage policies, accessing smarter tokens from cloud services may be a faster and more economical option than relying on a local token factory like the RTX Spark. However, for users prioritizing privacy, offline model availability, or the freedom to experiment, the RTX Spark platform offers a compelling blend of memory capacity, computational power, and software resources, all built on the familiar Windows environment.
The successful launch of such a capable and polished platform, even in its initial iteration, signals a noteworthy advancement for the PC industry. Nevertheless, Nvidia’s ambition to reshape personal computing will likely encounter challenges, particularly due to rising DRAM and NAND prices, partially influenced by the company’s own Vera Rubin rollout.
As we look ahead to the next two to three years, when products based on next-generation RTX Spark Superchips are expected to emerge, we may gain a clearer perspective on Nvidia’s success in achieving its vision. For now, the RTX Spark represents a platform for those who believe in the potential of open AI models to evolve within compact memory constraints, making the investment worthwhile. However, for those who are not yet convinced, it may be prudent to wait until Microsoft and Nvidia fully establish the software infrastructure necessary for this platform to fulfill its promises.
