NVIDIA Preps Windows for DGX Spark as RTX Spark PCs Launch October

NVIDIA spent the week in Berlin laying groundwork for a shift that has been quietly building since the company first shipped its DGX Spark clusters. On September 3rd, the company used the IFA conference to announce that Windows is coming to its local AI hardware—not just as an afterthought, but as a central part of the strategy. Eight manufacturers will ship RTX Spark Windows PCs starting in October: Lenovo, ASUS, Dell, HP, MSI, Acer, Gigabyte, and others, each carrying Blackwell GPUs and Grace CPUs in configurations ranging from thin laptops to compact desktops. But something else happened that week, something NVIDIA didn't announce at all. After applying the latest firmware update to a DGX Spark cluster in the lab, a new entry appeared in the device list: Windows UEFI CA—the Microsoft certificate that tells a machine's firmware it can trust a Windows boot loader. The certificate wasn't there before. It is now.

Understanding why that matters requires a brief detour into how secure boot works. When a computer starts up, its firmware checks a signature database to verify that whatever operating system is about to load is legitimate. Linux distributions boot through a shim signed by Microsoft's third-party UEFI certificate, which the firmware tracks as "UEFI CA." Windows Boot Manager uses a different Microsoft certificate entirely—the Windows production CA, tracked separately as "Windows UEFI CA." DGX Spark, which runs only DGX OS, never needed the Windows certificate before. There was no reason for it to be in the firmware's database. After this week's update, it is there, sitting alongside the BIOS database key, the Key Exchange Key, SBAT, the UEFI revocation list, and the ConnectX-7 network interface certificate. The firmware update log shows no pending updates for any of these components. The groundwork is laid.

Windows on DGX Spark is not imminent. The hardware still needs drivers for the Blackwell GPU, the Grace CPU's platform devices, the ConnectX-7 network interfaces, and the rest of the board. It needs an installer. NVIDIA has said nothing about any of that. What the company has said is that the same silicon going into RTX Spark Windows PCs next month—the GB10-class processors—represents the culmination of work that began at Computex, a collaboration between NVIDIA and Microsoft to build a complete Windows stack for local AI. Adding Microsoft's boot certificate to DGX Spark firmware is the kind of foundational work that makes such a transition possible.

The timing matters because Windows support was the single most requested feature from people considering a Spark. The AMD Ryzen AI Halo, a competing system that arrived earlier this year, supports both Windows 11 and Linux on the same 128-gigabyte machine. DGX Spark runs DGX OS and nothing else. When Windows arrives—if it arrives—that gap closes, and the machine's potential audience expands well beyond AI developers comfortable living in Ubuntu. For now, the RTX Spark Windows PCs are the official story. Lenovo showed the Yoga Pro 9n and Yoga 9n 2-in-1 at IFA. Acer showed a compact desktop concept. The full lineup includes the ASUS ProArt P16, Dell XPS 16, HP OmniBook X 14, Microsoft Surface Laptop Ultra, and MSI Prestige N16 Flip AI+ on the laptop side, with desktops from seven manufacturers. The laptops pack up to 6,144 CUDA cores, a 20-core Grace CPU, and up to 128 gigabytes of unified LPDDR5X memory in a 45- to 80-watt envelope. Desktops are cut to 5,120 CUDA cores, 18 CPU cores, and 64 gigabytes at 140 watts. Both configurations deliver up to one petaflop of FP4 AI performance and include PCIe Gen5, HDMI 2.1b, and three DisplayPort 2.1b outputs.

The software strategy centers on keeping AI agents local. NVIDIA introduced a new Windows Agent framework designed to let agents run safely in the background under operating system control. The company worked with OpenClaw—which it describes as the largest AI project on GitHub with more than 380,000 stars—to simplify Windows app setup on RTX GPUs with 24 gigabytes or more of VRAM. Nous Research's Hermes Agent now has one-click local model setup on Windows, with Linux support to follow. Perplexity's Portable Computer runs today on Linux with a 24-gigabyte-plus RTX GPU, and Windows support is coming soon. The pitch across all three is identical: complete workflows on the local machine, no cloud services, no burned credits.

NVIDIA also released PAIR, the Personal AI Router, a free open-source tool that discovers compatible PCs on a local network and routes inference requests to whichever system has available capacity. It works with Ollama and LM Studio and runs on GeForce RTX 20 Series and newer cards, RTX PRO workstation GPUs back to Turing, DGX Spark, and Apple M4 or later Macs. The beta is available now for Windows, macOS, and Linux in both graphical and terminal interfaces. On the inference side, NVIDIA claims up to 1.9 times higher llama.cpp throughput on a GeForce RTX 5090 from kernel optimizations and improved speculative decoding. vLLM gains reach 1.2 times on the RTX PRO 6000 Blackwell and up to 1.4 times on two-node DGX Spark clusters. The model catalog expanded to include Nemotron 3.5 Lightning, a 30-billion-parameter model sized for RTX PCs and DGX Spark; Z.ai's GLM-5.3-Flash multimodal mixture-of-experts; Qwen3.8-Flash-Next and Qwen3.8-27B; LTX 2.5 and MiniMax-H3 for video generation; Meta's 30-billion-parameter Muse Glimmer for coding and agentic work; and DeepSeek v4 Flash, a 284-billion-parameter mixture-of-experts with 13 billion active parameters that NVIDIA says runs on a two-Spark cluster or a DGX Station. The message is consistent: everything the user needs, running where the user is, without leaving the machine.