Nvidia pitches RTX Spark as the chip that finally makes local AI agents practical on Windows devices

RTX Spark is Nvidia's first move into Windows laptops, designed to run AI agents locally. The hardware is a Windows version of the already-known DGX Spark chip.

Nvidia unveiled RTX Spark at GTC Taipei. At the top end, the chip is the same GB10 Grace Blackwell Superchip that powers the DGX Spark. The difference is who it's for. Instead of a Linux workstation aimed at AI developers, RTX Spark targets Windows laptops and compact desktops for consumers. Nvidia is offering several variants with different core and SM counts, plus memory options ranging from 16 to 128 GB.

The top SKU pairs a Blackwell RTX GPU with 6,144 CUDA cores and fifth-gen Tensor Cores with a 20-core Arm-based Grace CPU, linked via NVLink-C2C. MediaTek helped design the CPU, according to Nvidia. Memory tops out at 128 GB, shared between CPU and GPU. The claimed 1 petaflop peak refers to FP4 precision with sparsity, a theoretical best case per Nvidia's specs. GPU performance sits close to a GeForce RTX 5070 Laptop GPU depending on the workload, Nvidia says.

Nvidia's answer to Apple Silicon and Snapdragon

RTX Spark follows the path Apple charted in 2020 with its M-series chips: Arm CPU, GPU, and memory controller on one package, sharing a unified memory pool instead of separate VRAM. Apple's M4 Max also offers up to 128 GB of unified memory at 546 GB/s bandwidth, but its Neural Engine tops out at 38 TOPS (INT8). RTX Spark claims roughly 1,000 TOPS by comparison, though that's FP4 with sparsity, so the conditions are very different. Still, the gap in raw AI compute is significant. Nvidia's real edge remains its CUDA stack, including TensorRT and RTX, which runs natively.

Qualcomm also pushed into Windows-on-Arm laptops with the Snapdragon X Elite in 2024, then followed up in September 2025 with the X2 Elite, boosting performance to 80 TOPS across 18 Oryon cores. Those chips are built around Microsoft's Copilot+ features, not local inference with multi-billion-parameter models. Traditional x86 platforms from Intel and AMD still rely on separate CPU and GPU memory with much smaller NPUs.

Local AI agents get new security guardrails

Nvidia argues that AI agents rarely run on users' primary devices because the right security tools haven't existed. New Windows components are supposed to provide identity management, agent isolation, and policy enforcement. The Nvidia OpenShell Runtime adds another layer: it defines what agents are allowed to do, routes requests to local or cloud models based on privacy settings, and masks personal data in cloud queries. The open-source projects Hermes Agent and OpenClaw already integrate this layer into their Windows apps, according to Nvidia.

Adobe also announced plans to rebuild Photoshop and Premiere for modern GPUs. Premiere is getting a new video pipeline with Nvidia TensorRT integration. Photoshop is getting a new engine with GPU-accelerated compositing. On RTX Spark, Premiere should also benefit from the shared memory pool. Adobe says the goal is AI, editing, and effects workflows that are up to twice as fast.

Alongside the laptop chip, Nvidia showed off the DGX Station for Windows. It's built on the GB300 Grace Blackwell Ultra Desktop Superchip with up to 748 GB of shared memory and 20 petaflops of FP4 performance. Nvidia says it can run models with up to a trillion parameters locally. It ships in Q4 2026.

RTX Spark devices will be available starting fall 2026 from ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI.