NVIDIA RTX Spark debuts at Computex; enters PC market with Windows on Arm chip

At Computex 2026 in Taipei on June 1, NVIDIA CEO Jensen Huang announced the RTX Spark superchip, marking NVIDIA's first major entry into the consumer and creator PC market. The RTX Spark pairs a Blackwell GPU with 6,144 CUDA cores and fifth-gen Tensor Cores (FP4 precision) with a custom 20-core NVIDIA Grace CPU based on Arm, connected via NVLink-C2C at 600 GB/s, all on a single TSMC 3nm SoC. Unified memory enables the CPU and GPU to share up to 128GB of LPDDR5X RAM at 300 GB/s memory bandwidth, eliminating traditional bottlenecks for on-device AI inference. The platform delivers 1 petaflop of FP4 performance and up to 100 FPS at 1440p gaming via DLSS 4.5 upscaling.

NVIDIA and Microsoft are positioning RTX Spark as the foundation for Windows on Arm and local AI agent execution. Dell, HP, Lenovo, ASUS, MSI, and Microsoft Surface will ship RTX Spark laptops and compact desktops starting fall 2026. The announcement sent AMD, Intel, and Qualcomm shares lower, as analysts view it as a credible threat to the x86 duopoly in consumer and edge AI computing. NVIDIA laid out a three-generation roadmap: the initial Grace Blackwell Spark (now shipping), followed by future Vera Rubin and Rosa Feynman variants with improved memory. The company is also collaborating with Adobe on GPU-accelerated versions of Photoshop and Premiere for RTX Spark, and announced DLSS 4.5 Ray Reconstruction for ray-traced gaming and Blender.

The timing is aggressive: NVIDIA faces DRAM and GDDR shortages that are pressuring the entire PC market, which shipped only 11.82 million discrete GPUs in Q1 2026. However, NVIDIA's leverage over software vendors, long-standing GPU ecosystem (CUDA, RTX, TensorRT, OptiX), and focus on on-device agentic AI give it advantages that prior Windows on Arm efforts (Qualcomm, Microsoft) have lacked.

For AI architects and platform strategists, RTX Spark signals a shift in competitive strategy: the edge becomes as important as the data center. If RTX Spark gains traction, local inference for AI agents and creative workloads will reduce cloud dependency. The unified memory model and NVLink interconnect demonstrate NVIDIA's approach to compressing the inference-vs.-training gap—moving GPT-scale AI capability to user devices. Windows on Arm adoption, however, depends on driver maturity, software emulation performance (for x86 apps), and ecosystem lock-in—execution challenges that have sunk prior Arm efforts on Windows.