Nvidia launches RTX Spark AI desktops and mini PCs reaching 1 PetaFlop of AI performance
Nvidia is attempting to pivot the AI conversation from cloud-based subscriptions to local ownership with the launch of the RTX Spark desktop line. These small-form-factor (SFF) Windows 11 PCs are positioned as high-end AI appliances, designed to move the heavy lifting of agentic AI from distant data centers to the user's own desk.
The foundation of this ecosystem is the RTX Spark N1X Superchip, a piece of hardware custom-built with MediaTek that combines a 20-core Grace CPU with a 6,144-core Blackwell RTX GPU. This configuration is purpose-built for local AI, delivering 1 Petaflop of FP4 compute and supporting up to 128GB of unified LPDDR5X memory. By standardizing on this top-tier silicon for the desktop line, Nvidia enables these machines to run large local AI models, including 120B-parameter models with a 1M-token context window, without relying on cloud dependency.
The Economics of Local AI
This shift toward local compute is driven largely by the escalating costs of cloud API fees. According to Nvidia, a single RTX Spark desktop can generate 61 million output tokens per month for roughly $20 in electricity costs, whereas achieving the same volume through a cloud API like Claude Sonnet 5.5 Medium would cost approximately $1,200. To manage these workloads, Nvidia is introducing PAIR software, which allows the RTX Spark desktop to act as a central coordinating hub, distributing AI subtasks across other idle devices on a home network, including laptops and even MacBooks.
Complementing the hardware is a suite of software integrations from Microsoft, including agentic AI features such as OpenShell, MXC sandboxing, and granular PC capability permissions. Users can also employ NeMo Switchyard for local-cloud routing, which balances workloads between local models and specific cloud tools. For those balancing professional work with leisure, the Nvidia App includes a GPU prioritization feature to allocate resources between background AI agents and active gaming.
Hardware Partners and Market Positioning
While the AI capabilities are the primary draw, the hardware remains versatile enough for traditional workstation tasks and gaming. The Blackwell-based GPU supports the full Nvidia stack, including CUDA, TensorRT, and DLSS 5. Because the systems run on Grace ARM processors, gaming support is arriving via native Windows on Arm updates for titles like Marvel Rivals and World of Warcraft: Forever, while DLSS 5 updates are slated for Hogwarts Legacy and War Thunder.
Several manufacturers have launched systems today, though availability varies by region and brand. The Asus ProArt GR1X is available in the US with 64GB or 128GB of RAM and 1TB or 2TB of storage, with the base 64GB/1TB model priced at $4,499. MSI has released the EdgeMesa N AI+, a 1.3-litre mini PC featuring a vapour chamber and dual turbo fans to maintain a 140W TDP. The EdgeMesa N AI+ is priced at £4,099 for the 64GB/1TB configuration and £6,999 for the 128GB/4TB version. Other models from Dell, Gigabyte, HP, and Lenovo are expected in the coming weeks, while Acer’s SFF RTX Spark is slated for early 2027.
Despite the high specifications, the RTX Spark faces stiff competition from Apple’s Mac Studio, which offers M5 Max and M5 Ultra chips. Some analysis suggests that even at maximum configuration, the RTX Spark may struggle to match the raw CPU and GPU potential of Apple's latest silicon, a trend previously noted in comparisons between the M-series and RTX Spark laptops.
The launch comes at a volatile time for Nvidia's leadership, as CEO Jensen Huang is currently on the outside of the company following a recent fiasco. While the company has focused heavily on the AI training chip market, these consumer-level desktops represent an attempt to establish a permanent footprint in the home and office as an "AI hub." The success of this strategy likely depends on whether users are willing to pay a high upfront hardware premium to avoid the long-term costs of cloud AI tokens.

