Nvidia just unveiled RTX Spark, a new superchip built to turn Windows PCs into machines that run personal AI agents on-device. According to Hacker News, the company announced it at GTC Taipei alongside a deep partnership with Microsoft, pitching it as the first class of computer designed from the ground up for the era of personal agents. The framing from Nvidia is blunt: the PC moves “from tool to teammate.”
What stands out here is the scope. This isn’t a faster gaming card. It’s a full reimagining of what a Windows machine does, packaged into both slim laptops and compact desktops.
What’s inside RTX Spark
The chip pulls together 30 years of Nvidia hardware into a single package. The headline specs Hacker News reports:
- 1 petaflop of AI compute with up to 128GB of unified memory
- A Blackwell RTX GPU with 6,144 CUDA cores and fifth-gen Tensor Cores using FP4 precision
- A 20-core Nvidia Grace CPU, co-designed with MediaTek for Arm-based power efficiency
- NVLink-C2C linking the GPU and CPU directly on the chip
- The full stack: CUDA, RTX, DLSS, TensorRT, OptiX, Reflex and G-SYNC
Built for agents that run locally
The real bet is on-device AI agents. Nvidia and Microsoft argue that adoption has stalled because people can’t run agents securely and privately on their main computer. Their fix is a security layer: new Windows security primitives plus Nvidia’s OpenShell runtime.
OpenShell lets users define what an agent can and can’t do, route queries to local models based on privacy rules, and mask personal information before anything goes to the cloud. Open-source agent projects OpenClaw and Hermes Agent are already adopting it for their new Windows apps. Those agents can execute tasks across Windows apps, reason through cross-app workflows, generate images and video, write plug-ins, and search local files by meaning.
Nous Research CEO Dillon Rolnick put the pitch plainly: “You realize you’re buying a full-fledged assistant, not a typical laptop.”
Creators and gamers get the same silicon
RTX Spark isn’t agents-only. On the same hardware, users can render 90GB+ 3D scenes, edit 12K 4:2:2 video with the Blackwell decoder, generate 4K AI video, run 120-billion-parameter models with up to 1 million tokens of context, and play AAA games at 1440p above 100 fps with ray tracing and DLSS.
Adobe is rebuilding Photoshop and Premiere from the ground up for the chip, promising 2x faster AI and graphics performance. That kind of ground-up rewrite from a major software vendor signals this is more than a spec bump.
Availability and who’s shipping it
RTX Spark machines arrive this fall. The first wave covers slim laptops with all-day battery and compact desktops from ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI, with Acer and GIGABYTE to follow. Nvidia hasn’t named pricing in the announcement, and exact configurations per vendor remain open.
Why it matters: This is significant because it reframes the privacy problem that’s been holding back personal AI. Sending your files, your screen, and your context to a cloud model is a hard sell for a lot of people. Running a capable agent locally, with hardware-level containment and the option to mask data before it ever leaves the device, is a genuinely different proposition.
There’s also a strategic read. Nvidia and Microsoft are trying to define the personal AI computer before Apple, Qualcomm, or a pure cloud play locks in the category. Putting agents in the Windows taskbar, on silicon that also crushes creative and gaming workloads, is a wide net.
The caveats are worth keeping in mind. No pricing yet, fall is still months out, and “1 petaflop” is an FP4 figure, not a number you can compare cleanly to older benchmarks. The agent security model also only matters as much as developers actually adopt it, though OpenClaw and Hermes signing on early is a real start.
The bigger question is whether “ask and the PC does the work” holds up outside a keynote demo. Full details are at the original source.