NVIDIA RTX Spark: an Arm superchip with a Blackwell GPU aims to redefine the Windows PC for agents
On May 31, 2026, at GTC Taipei, NVIDIA and Microsoft announced RTX Spark, a superchip for Windows PCs that puts an NVIDIA GPU and CPU in the same package. Jensen Huang summed up the ambition by saying “the PC is being reinvented”. Before anything else, an important caveat: this is an announcement. Laptops and desktops with RTX Spark are expected this fall, so none of what follows has been tested by third parties yet.
What RTX Spark is
According to NVIDIA’s announcement, the chip combines a Blackwell RTX GPU with 6,144 CUDA cores and fifth-generation Tensor Cores with FP4 precision, connected over NVLink-C2C to a 20-core Arm NVIDIA Grace CPU. MediaTek collaborated on the design of that CPU. NVIDIA quotes up to 1 petaflop of AI performance and up to 128GB of unified memory.
Unified memory is the number that catches my attention most. On traditional PCs, the size of the model you can run locally is limited by the card’s VRAM. With the CPU and GPU sharing a large pool, the math changes. NVIDIA claims, for example, that it can run 120-billion-parameter LLMs with 1 million tokens of context. It’s a vendor claim, but it shows where the product is aimed.
The laptops will range from 14 to 16 inches, as slim as 14 mm, according to NVIDIA. Partners for the fall launch are ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI, with Acer and GIGABYTE following later. Compact desktops are also planned.
Windows ready for agents
The software side is where the announcement gets more interesting for developers. NVIDIA is bringing OpenShell to Windows, built on new system security and containment primitives. According to NVIDIA, OpenShell lets you define what agents can and cannot do, routes queries to local models based on privacy policies, and masks personal information in queries sent to cloud models. Hermes Agent and OpenClaw are adopting this stack.
Microsoft’s post adds platform details:
- Windows ML gained GPU acceleration, and AI developers will be able to use TensorRT natively on Windows.
- Prism, the emulation layer that runs 32- and 64-bit x86 apps on Windows on Arm, was tuned for the RTX Spark microarchitecture.
- MATLAB now officially supports Windows on Arm via Prism.
- Windows got changes for unified memory, including a higher limit on how much system memory the GPU can access.
- Planned for RTX Spark are CUDA-accelerated PyTorch, llama.cpp, TensorRT, Hugging Face frameworks, Unsloth and Kohya.
Microsoft is also extending Windows to NVIDIA DGX Station, with the GB300 Grace Blackwell Ultra superchip, later this year, with access to the Linux AI ecosystem through WSL.
What changes for developers
For anyone working on local AI, the combination of large unified memory with native CUDA and TensorRT on Windows is the main draw. Today, people who want to run big models locally usually choose between a Mac with unified memory and a PC with a discrete GPU. RTX Spark tries to offer both.
The thing to watch is the Arm architecture. Native Arm tools already exist in good numbers, and Prism helps with the rest, but the list of CUDA tools is described as “planned”. Before switching machines, it’s worth checking whether your specific workflow (drivers, native Python extensions, build tools) will work on day one.
My take
It’s the most relevant hardware announcement of the year for anyone building with AI on Windows. The proposal makes sense: agents running locally, with policies that decide what can leave for the cloud, is exactly what privacy-conscious companies want to hear.
But for now it’s all promises: NVIDIA’s performance numbers, planned tools and machines that aren’t on sale yet. I’d wait for the first independent tests, especially of battery life, emulation and tool compatibility, before treating RTX Spark as a replacement for a current workstation.