Nvidia is finally walking into the Windows PC market on its own silicon. At GTC Taipei, timed alongside Computex 2026, the company showed off RTX Spark, a system-on-chip built with Microsoft and MediaTek that pairs a 20-core Arm-based Grace CPU with a Blackwell RTX GPU packing 6,144 CUDA cores. The two halves talk over Nvidia's NVLink-C2C interconnect, sharing up to 128GB of unified LPDDR5X memory. Microsoft laid out the collaboration in a Windows blog.
The pitch is local AI. Nvidia says the chip can run 120-billion-parameter language models on-device with context windows up to 1 million tokens, no cloud bill required, according to its press release. Those figures are the company's own. The silicon is essentially the same GB10 superchip already shipping inside the Linux-based DGX Spark workstation, repurposed here for Windows 11 on Arm.
Gaming gets folded in too. The GPU lands around the mobile GeForce RTX 5070 in capability, with DLSS 4.5 support and a claimed 1440p at 100-plus fps. The whole thing tops out at an 80W power ceiling.
Laptops and compact desktops arrive this fall from ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI. Nvidia hasn't named consumer pricing yet, though it's flagged the premium end of the market; for reference, the enterprise DGX Spark now runs $4,699 after a recent DRAM-driven hike. Further up the stack, Nvidia is prepping DGX Station for Windows, built on the GB300 Grace Blackwell Ultra chip with up to 748GB of coherent memory and a 72-core Grace CPU.
Bottom Line
RTX Spark laptops ship this fall, claiming on-device support for 120-billion-parameter models in up to 128GB of unified memory.
Quick Facts
- 20-core Arm Grace CPU, Blackwell GPU with 6,144 CUDA cores
- Up to 128GB unified LPDDR5X memory
- 80W power ceiling
- Runs 120B-parameter models locally (company-reported)
- Ships fall 2026; consumer pricing not announced
