DGX Spark Prepares for Windows as NVIDIA Confirms October RTX Spark Launch

Windows Support Transforms the DGX Spark

A recent firmware update applied to a DGX Spark cluster revealed a new "Windows UEFI CA" entry, according to StorageReview. This is the specific Microsoft certificate required to trust a Windows boot loader. Previously, the DGX Spark ran DGX OS exclusively, which meant users had to operate entirely within an Ubuntu environment.

Native Windows support has been the single most requested feature for the Spark platform. Removing the strict Linux requirement allows South African businesses to integrate these powerful AI nodes directly into their existing Windows domains. You no longer need to hire a dedicated Linux administrator just to manage your local AI infrastructure.

While drivers for the Blackwell GPU and Grace CPU are still pending, adding the boot certificate is the critical foundational step. For local enterprises budgeting in rand, this upcoming compatibility fundamentally changes the return on investment. It turns a niche developer tool into a standard, easily managed corporate asset.

RTX Spark PCs Arrive in October

NVIDIA has confirmed that RTX Spark PCs from vendors including Lenovo, Acer, Dell, and HP will land in October. The laptop variants feature up to a 6,144-core Blackwell RTX GPU and a 20-core Grace CPU. They pack up to 128GB of unified LPDDR5X memory within a highly efficient 45W to 80W power envelope.

Desktop models will offer 5,120 CUDA cores, 18 CPU cores, and 64GB of memory operating at 140W. Both form factors deliver up to one petaflop of FP4 AI performance. They also include PCIe Gen5 connectivity, HDMI 2.1b, and three DisplayPort 2.1b outputs for comprehensive multi-monitor workstation setups.

This low power draw is a massive advantage for South African operations, allowing heavy local AI workloads to run comfortably on standard office UPS systems during load-shedding. Global October availability means local buyers should expect initial lead times stretching into late November as global supply chains move.

Importing early units individually carries significant freight, customs, and warranty burdens for a single business. At Server Hub, we absorb those logistics and risks at scale. We manage the complex import process so you simply receive fully supported, production-ready hardware directly to your office without the administrative headache.

Local AI Agents and Network Routing

NVIDIA is heavily pushing local AI to help businesses eliminate ongoing cloud credit burn. A new Windows Agent framework allows AI agents to run safely in the background under operating system control. Tools like OpenClaw and Nous Research’s Hermes Agent now feature simplified, one-click Windows setups for machines with 24GB or more of VRAM.

To maximise hardware utilisation, NVIDIA introduced PAIR, a free Personal AI Router. This open-source tool discovers compatible PCs on your local network and routes independent inference requests to whichever system has spare capacity. It supports hardware ranging from older RTX 20 Series GPUs up to the new DGX Spark clusters.

For South African businesses, PAIR offers a brilliant way to pool local compute resources. Instead of buying dedicated high-end AI hardware for every single employee, you can route office-wide requests to a central DGX Spark or a cluster of RTX workstations. This approach maximises your rand spend on enterprise compute.

Accelerated Inference and New Local Models

Software optimisations are driving significant performance gains on existing and upcoming hardware. NVIDIA reports up to 1.9x higher llama.cpp throughput on the GeForce RTX 5090 due to kernel optimisations and improved speculative decoding. These raw inference improvements flow directly through to popular local tools like LM Studio and Ollama.

Enterprise clusters are also seeing a boost. The vLLM framework gains a 1.2x speedup on the RTX PRO 6000 Blackwell and up to a 1.4x improvement on two-node DGX Spark clusters. This allows businesses to process larger batches of internal data faster without upgrading their physical hardware.

A fresh batch of local models is now available to leverage this speed. This includes Nemotron 3.5 Lightning, a 30-billion-parameter model sized specifically for RTX PCs and the DGX Spark. Meta’s 30-billion-parameter Muse Glimmer is also available for coding tasks, while DeepSeek v4 Flash can run on a two-Spark cluster.

These software and model updates mean local hardware is becoming increasingly capable of handling complex enterprise workflows. Waiting for next year's hardware is entirely unnecessary when October-bound systems can run these advanced models locally, securely, and completely offline today.