Signals›Chips & compute
Chips & computeShipping

Nvidia's first Windows chip is built to run AI on your laptop, not in the cloud.

On sale since October 16 from six PC makers, RTX Spark pairs an Arm CPU and a Blackwell GPU with up to 128 GB of shared memory — enough, Nvidia says, to run a 120-billion-parameter model on your lap.

ZN ZilverNet EditorialPublished · 5 min read · 3 sources On sale Oct 16, 2026
A laptop shown open at several angles, with its internals visible through a translucent shell
Image credit: Nvidia
The short versionShipping
What it is
One "superchip" combining an Nvidia Blackwell RTX GPU (6,144 CUDA cores) and a 20-core Arm-based Grace CPU, designed with MediaTek, sharing up to 128 GB of memory.
How far along
ResearchPrototypePilotShippingEverywhere
Shipping. Laptops from ASUS, Dell, HP, Lenovo, Microsoft and MSI since Oct 16. Desktops were promised for the fall but have no date yet.
Price
From $2,599, per Nvidia's press release — though Nvidia's own blog says $2,399. Pricing varies by region.
Why it matters
Memory the GPU can use directly is what lets large models run locally. Nvidia says up to 120B parameters.
Still unproven
"1 petaflop" of AI performance, game frame rates and "all-day battery life" are Nvidia's claims.

What's inside

Nvidia announced RTX Spark on May 31 at GTC Taipei and put it on sale on October 16. It is a single chip with two parts: a Blackwell RTX graphics processor with 6,144 CUDA cores and Tensor Cores that support the low-precision FP4 format used for AI, joined by a high-speed link to a 20-core Arm-based Grace CPU. MediaTek, Nvidia says, "collaborated with NVIDIA on the custom CPU design."

Nvidia rates it at "1 petaflop" of FP4 AI performance. That is its own figure, measured at the precision most favorable to big numbers.

It also marks Nvidia's entry into Windows PCs with a CPU of its own. To make that practical for developers, Nvidia says it built "the first PyTorch-CUDA build for Windows on Arm" and that 97% of the top 15,000 Python packages are supported.

Why the memory is the headline

For running AI models, the limit is usually memory, not speed: a model has to fit before it can run. RTX Spark gives the GPU access to up to 128 GB of unified memory, and Nvidia says that is enough to run "120B-parameter LLMs with up to 1 million tokens context" locally.

The chip ships alongside Nvidia's own Nemotron 3.5 Bolt, a 74-billion-parameter model built to fit in 24 GB, and day-one agent apps including GitHub Copilot, Hermes, OpenClaw and Perplexity.

"The PC is being reinvented." — Jensen Huang, founder and CEO, Nvidia

Agents, with guardrails

The launch came with security plumbing for AI agents, built with Microsoft. Windows gains new controls covering "identity, containment, policy and end-to-end security," and Nvidia's OpenShell runtime lets users set what agents are allowed to do. It can also route requests to local models based on a privacy policy, and disguise personal information before anything is sent to a cloud model.

That is exactly the layer worth asking about with any agent product — see our questions for Grok Bot.

"Our goal is to deliver unmetered intelligence to every home and every desk with Windows." — Satya Nadella, chairman and CEO, Microsoft

The fine print

  • Price: Nvidia's October 16 press release says laptops start at $2,599; its October 17 blog post says $2,399. Neither explains the difference.
  • Desktops: the May announcement promised laptops and compact desktops "this fall." The October launch covered laptops only, with no desktop date.
  • Video claims shifted: May's announcement said 12K video editing; October's says 8K.

Launch laptops include the ASUS ProArt P14 and P16, Dell XPS 16, HP OmniBook Ultra and OmniBook X 14, Lenovo Yoga Pro 9n, Microsoft Surface Laptop Ultra, and MSI Prestige N16 and N16 Flip AI+.

Update log
  • Published.