NVIDIA Unveils RTX Spark Superchip

NVIDIA used its GTC Taipei keynote at COMPUTEX (June 1–2, 2026) to unveil RTX Spark, a new Arm-based “superchip” the company says will power a fresh class of Windows laptops and compact desktops designed for local AI agents.

NVIDIA describes RTX Spark as a single package that pairs a 20‑core Arm host CPU with a Blackwell‑architecture GPU, and supports up to 128 GB of unified LPDDR5X memory. The company claims the chip can deliver roughly one petaflop of on‑device AI throughput.

Under the hood the GPU element is reported to include up to 6,144 CUDA cores and fifth‑generation tensor engines, while the platform is built to expose the full CUDA and RTX software stack to Windows on Arm. NVIDIA and reviewers say the chip targets heavy local AI workloads rather than just graphics.

NVIDIA framed RTX Spark as a hardware foundation for what it and Microsoft call “agentic” Windows — machines that run persistent personal AI agents which can act across apps, manage tasks, and run large models locally rather than relying solely on the cloud. Microsoft has worked with NVIDIA on the stack and on Windows‑level integration.

Major PC makers already confirmed RTX Spark systems in development, with Microsoft, Dell, HP, Lenovo, ASUS and MSI named as early partners. NVIDIA and Microsoft say the first RTX Spark laptops and small form‑factor desktops should start shipping in the fall of 2026.

Alongside consumer devices, Microsoft previewed a compact developer system — the Surface RTX Spark Dev Box — and NVIDIA pointed back to its earlier DGX Spark developer hardware as a technical ancestor. Those systems are intended to let developers train and test models locally before scaling to the cloud.

Software makers are already lining up to tune apps for the platform. NVIDIA says the RTX Spark machines will run the full CUDA toolchain and that major creative and game developers are rearchitecting workloads to use unified memory and on‑device inference. The company also told press it is working with anti‑cheat and DRM vendors to smooth game compatibility on Windows on Arm.

NVIDIA and some reporting suggest the extra memory and on‑device performance let RTX Spark run models in the 70–120 billion‑parameter range locally, a scale that could support more capable agents without constant cloud calls. The company’s one‑petaflop figure is a vendor claim that will need independent testing in real‑world workloads.

Windows on Arm has historically faced app compatibility and emulation challenges, and analysts at the show pressed NVIDIA and Microsoft on that risk. Microsoft is pitching new developer tools and AI‑assisted porting to address gaps, but third‑party app and driver support will be central to the platform’s mainstream success.

The move places NVIDIA directly into the traditional PC CPU/GPU fight, inviting comparisons with Intel, AMD and Apple’s silicon. Observers at COMPUTEX noted that RTX Spark pivots NVIDIA from a datacenter and discrete‑GPU focus toward selling tightly integrated Arm‑based systems for consumer and professional PCs.

NVIDIA gave few public details about pricing or battery impact for thin‑and‑light laptops, and OEMs say final configurations will vary; many laptops will ship with smaller memory configurations than the 128 GB maximum. That leaves cost, thermals and real battery life as open questions for reviewers and buyers this fall.

If the platform delivers on its claims, RTX Spark could reshape how Windows handles background AI, local models and multimedia workflows — but success depends on software support, realistic power budgets, and whether users value persistent local agents over cloud services. NVIDIA, Microsoft and partners will spend the rest of 2026 trying to prove that case.