NVIDIA and Microsoft drop RTX Spark for AI agents

For forty years, using a personal computer meant the same tedious routine: launch an app, click around, type, and repeat. NVIDIA and Microsoft want to change that dynamic entirely, turning Windows PCs from passive tools into active digital coworkers.

At GTC Taipei, the two tech giants revealed RTX Spark, a new superchip engineered specifically to run local AI agents directly on Windows. The hardware packs an NVIDIA Blackwell RTX GPU with 6,144 CUDA cores alongside a 20-core NVIDIA Grace CPU, tied together by high-speed NVLink-C2C. NVIDIA even partnered with MediaTek to co-design the custom CPU, aiming for all-day battery life on thin-and-light laptops.

Desktop-class compute inside a laptop

The most striking technical leap is the memory footprint: RTX Spark delivers up to 128GB of unified memory and 1 petaflop of local AI performance.

That massive unified memory pool fundamentally shifts what can run on a desk. You can run 120-billion-parameter LLMs locally with a 1-million-token context window, edit 12K 4:2:2 video streams, or render 90GB 3D scenes without leasing expensive cloud clusters. For gamers, it drives AAA ray-traced titles at 1440p over 100fps using DLSS 4.5 Ray Reconstruction.

Software makers are rushing to adapt. Adobe announced it is rebuilding Photoshop and Premiere from scratch for RTX Spark to double AI and graphics speeds. Hardware makers including ASUS, Dell, HP, Lenovo, MSI, and Microsoft Surface will ship notebooks and compact desktops this fall, with Acer and GIGABYTE to follow.

Sandboxing the digital assistant

Letting an autonomous AI agent roam freely inside your primary computer sounds terrifying from a privacy standpoint. Microsoft and NVIDIA are addressing that head-on with new native system plumbing.

Windows is introducing new security primitives covering identity, isolation, and policy enforcement. On top of that, NVIDIA built OpenShell, a runtime that lets users define exact permissions for what an agent can and cannot touch. It can route queries locally to keep data private, or automatically redact sensitive personal info before sending requests to frontier cloud models.

Popular open-source agent projects like OpenClaw and Nous Research’s Hermes Agent are already integrating the stack into dedicated Windows apps. Microsoft also showed off the Surface RTX Spark Dev Box, an aluminum-clad desktop mini PC tailored for AI builders, preloaded with developer-mode Windows 11 Pro, WSL 2 with CUDA passthrough, and native tooling.

Why it matters

For AI developers and heavy creators, RTX Spark is an absolute dream machine. Running agentic loops and local model fine-tuning used to mean either burning cloud credits or rigging multiple noisy desktop GPUs. Having 128GB of unified memory on a local desk eliminates cloud subscription anxiety while keeping proprietary code completely offline. If your workflow hinges on local inference or privacy-critical data, this hardware is worth budgeting for immediately.

Strategically, this is the most formidable counterattack the Windows ecosystem has ever mounted against Apple Silicon. For the past few years, developers flocked to high-spec MacBooks primarily because Apple unified system RAM with GPU memory, making them the only practical portable rigs for local LLMs. By combining Grace, Blackwell, and MediaTek's low-power expertise, NVIDIA and Microsoft have eliminated that exclusive Apple advantage.

That said, pulling off an Arm-based Windows shift is never frictionless. While Microsoft and NVIDIA have nailed the raw specs, legacy application compatibility and developer tooling across diverse Windows environments will face scrutiny. These machines won't come cheap either, making them a specialized powerhouse for professionals rather than a casual upgrade for mainstream users.