Nvidia RTX Spark Reaches Windows PCs as Microsoft Opens Surface Laptop Ultra Preorders

Microsoft and Nvidia on October 7, 2026 opened preorders for the first Windows PCs built on Nvidia's RTX Spark chip, led by the Surface Laptop Ultra at $2,599. Both companies pitch RTX Spark, which pairs an Nvidia Blackwell RTX GPU with an Nvidia Grace CPU and up to 128GB of unified memory, as a way to run large AI models on a PC instead of in the cloud.

The launch, at a San Francisco event with Microsoft CEO Satya Nadella and Nvidia CEO Jensen Huang, brings the Arm based design behind Nvidia's desktop AI chips to premium laptops. Microsoft first showed the Surface Laptop Ultra on May 31, 2026 without a price or date, in a Surface blog post, and says the two companies introduced RTX Spark in June.

What did Microsoft and Nvidia announce?

Nvidia said on October 7, 2026 that RTX Spark laptops are on preorder and available from October 16, with compact desktops in November. Microsoft lists preorders from ASUS, Dell, HP, Lenovo, MSI and Surface; Nvidia also names Acer and Gigabyte.

According to Nvidia's announcement, RTX Spark combines a Blackwell RTX GPU with up to 6,144 cores and a Grace CPU with up to 20 cores, connected at 600 GB/s. Nvidia rates it at one petaflop of AI performance in FP4, a low precision number format used for AI inference; a Microsoft footnote says that figure is theoretical and relies on sparsity. The chip runs Nvidia's CUDA software stack.

The Surface Laptop Ultra has a 15 inch touchscreen and weighs under 4.5 pounds, according to Microsoft's Surface blog. The $2,599 base model has 24GB of memory and 512GB of storage, and the top model with 128GB, a 20 core CPU and 1TB costs $5,899.99 (The Verge). Microsoft's Windows chief made memory the selling point:

Which AI models can RTX Spark run locally?

Microsoft says Surface devices with RTX Spark and 128GB of memory can run AI models with more than 120 billion parameters locally. Unified memory is one pool shared by CPU and GPU, though Microsoft notes the GPU can address less than the total.

In its Windows blog post, Microsoft named the first models tuned for these PCs: its own MAI Code 1.1 Flash (137 billion total parameters, 6.8 billion active) compressed to 3 bit precision with a 256,000 token context window, an upcoming Nvidia Nemotron model with more than 70 billion parameters that fits in just over 20GB, and the 284 billion parameter DeepSeek V4 Flash. Nvidia points to Qwen 3.8 Flash Next, a 125 billion parameter model.

How does RTX Spark compare with Apple's MacBook Pro?

Microsoft claimed on October 7, 2026 that Windows PCs with RTX Spark are up to 2.1 times faster to the first token of an AI answer, 4.3 times faster at AI image generation and 6.2 times faster at AI video generation than a 16 inch MacBook Pro with Apple's M5 Pro chip. No independent test results were available at launch.

Per Microsoft's footnotes, preproduction 64GB RTX Spark machines faced a 64GB MacBook Pro in September 2026, in tests commissioned by Microsoft (llama.cpp) and run by Nvidia (ComfyUI). The comparison uses the M5 Pro, not the higher end M5 Max that Apple also sells in the 16 inch MacBook Pro (The Verge).

The Verge puts the $2,599 entry model close in price to a similar 14 inch MacBook Pro with M5 Pro, which has 1TB of storage.

Where does the $5,999 Dev Box fit?

The Surface RTX Spark Dev Box is a $5,999 mini PC with 128GB of unified memory that ships in the US in November, sold only on Microsoft.com, Microsoft said on October 7, 2026. It ships with Windows 11 Pro and developer tools such as GitHub Copilot preinstalled (The Verge).

The Verge calls it pricier than the DGX Spark mini PC that Nvidia launched in 2025 at $3,999. Nvidia raised the Founders Edition price to $4,699 in February 2026, according to an Nvidia forum notice, and Guru3D reports the 128GB Founders Edition now costs $6,950. At that reported price, Microsoft's 128GB Windows box costs about $950 less than Nvidia's own.

Above both, Microsoft previewed DGX Station for Windows, a deskside system on Nvidia's GB300 chip with 748GB of memory; until now DGX Station ran only on Linux, Nvidia said.

How do the new Windows agent features use local hardware?

Microsoft made Microsoft Execution Containers (MXC), sandboxes that limit which files and networks an AI agent can reach, generally available on Windows 11 on October 7, 2026. It says Copilot on Copilot+ PCs will gain access to local files, actions across Windows and local models in the coming months.

Microsoft says OpenAI's Codex, GitHub Copilot, OpenClaw and LM Studio already support MXC, and that Meta's Muse agent is coming to Windows. GitHub's HydraFusion, which routes each coding task to a suitable model, gains local models in an experimental preview later in October.

What we don't know yet

• How RTX Spark performs in independent tests against the MacBook Pro and AMD's chips.
• How much of the 128GB the GPU can use for models in each configuration.
• Prices and ship dates outside the US; the Dev Box is US only for now.
• What DGX Station for Windows will cost and when it ships.