NVIDIA launches RTX Spark desktops for local AI agents
NVIDIA's RTX Spark compact desktops let businesses run AI agents and models locally, cutting cloud costs and keeping data on the company network.
On November 18, 2026, NVIDIA announced (opens in a new tab) the general availability of its RTX Spark compact desktop PCs. Built by hardware partners including ASUS, Dell, and MSI, these systems run on Windows on Arm architecture and are designed specifically for local artificial intelligence workloads. For a small or mid-sized business, this release offers a different way to deploy and finance artificial intelligence. Instead of paying cloud providers for every prompt, generation, or automated action, companies can purchase a dedicated machine to run agents, process data, and fine-tune models locally. By moving these workloads from the cloud to a physical box in the office, businesses gain a predictable cost structure and complete control over their proprietary data.
The hardware moving models out of the cloud
The technical specifications of the RTX Spark desktops show a clear intent to handle heavy, continuous computing tasks. Each unit includes a 20-core NVIDIA Grace central processing unit and an NVIDIA Blackwell RTX graphics processing unit equipped with 6,144 CUDA cores. The systems deliver up to 1 petaflop of artificial intelligence performance. The most significant feature for business applications is the inclusion of up to 128GB of unified memory.
In artificial intelligence workloads, memory capacity dictates the size of the model a computer can run. With up to 128GB available, these compact desktops can run large language models with 120 billion parameters entirely on the device. They can also render 3D scenes larger than 90GB and process 4:2:2 video formats in real time. A 10-gigabit Ethernet connection allows the desktop to pull massive datasets, model files, and creative assets from a company network quickly.
Running a 120-billion-parameter model previously required renting expensive cloud infrastructure. By packaging this capacity into a compact desktop, NVIDIA offers businesses a fixed-cost alternative. These machines are built to run personal artificial intelligence agents constantly, handling scheduled tasks while employees are away or working on other devices.
A maturing software ecosystem
Hardware requires software to be useful. The RTX Spark launch indicates that major software vendors are treating Windows on Arm architecture as a primary platform for professional work.
Adobe has released native Windows on Arm versions of Illustrator and InDesign. Illustrator operates up to 30% faster than its emulated counterpart. Substance 3D Painter is also available in public beta with native Arm support, requiring no compatibility layer. Premiere and Photoshop include specific hardware acceleration for features like speech enhancement, scene edit detection, and object removal. For video editors, the update improves support for professional formats including Canon RAW, Red RAW, and ProRes RAW. This enables video editors to complete everyday H.264 deliverables more than 50% faster.
The 3D modeling software Blender has integrated DLSS 4.5 Ray Reconstruction into its version 5.3 viewport. This integration provides artists with near-final-quality ray-traced previews in real time, eliminating the need to wait for a noisy image to resolve.
For developers, the systems support Linux development tools through the Windows Subsystem for Linux. NVIDIA also introduced Switchyard, a tool within its Hermes software that automatically routes tasks. Switchyard sends everyday requests to the local desktop while directing complex reasoning tasks to larger cloud models. This hybrid approach helps companies manage their cloud spending without sacrificing capability when a task exceeds the local machine's limits.
What it means for your business
The arrival of capable, unified-memory desktops changes the financial math of artificial intelligence for a company of 50 or 100 people. Currently, most businesses access artificial intelligence through application programming interfaces, paying a fee for every token processed. This variable cost structure works well for occasional use but becomes expensive when a company deploys automated agents that run continuously.
Purchasing an RTX Spark desktop shifts this expense from a recurring operational cost to a one-time capital expenditure. A business can set up an agent to process internal documents, sort data, or generate assets overnight without accumulating per-token cloud charges. The machine runs constantly, handling the bulk of routine automated work.
Data privacy provides a second financial and operational advantage. Processing files locally means proprietary code, unreleased product designs, and sensitive client information never leave the company network. Businesses in regulated industries, or those with strict client confidentiality agreements, can use advanced artificial intelligence tools without routing their data through third-party cloud servers.
For creative agencies and development shops, the benefit is immediate time savings. A dedicated rendering and generation box allows employees to offload heavy processing tasks. A designer can use natural-language instructions in ComfyUI to turn sketches into rendered images on the RTX Spark desktop, and immediately return to other work on their primary laptop.
What to watch
The current availability includes systems from ASUS, Dell, and MSI. The hardware ecosystem is expanding rapidly. Acer, GIGABYTE, and HP plan to release their own mini PCs soon, alongside a Microsoft Surface RTX Spark Dev Box. The pricing and availability of these upcoming models will determine how accessible this technology becomes for the smallest businesses.
Enterprise infrastructure is also adapting to this local trend. Microsoft announced at its Ignite conference that Windows Server 2028 support is coming to the NVIDIA DGX Station for Windows platform, which is powered by the GB300 Grace Blackwell Ultra Desktop Superchip. This signals that server-grade local artificial intelligence is moving closer to standard enterprise environments, giving companies more hardware choices at different price points.
Monitor the development of local models and developer tools optimized for this hardware. Developers can now bring any GGUF model from Hugging Face and run it within Windows Machine Learning using the llama.cpp backend. As open-source models become more efficient, the capabilities of a local system will grow. The balance between what must be sent to the cloud and what can be done locally will continue to shift toward the office desk.