LLMs

Microsoft and NVIDIA Open RTX Spark Preorders: Surface Laptop Ultra From $2,599, Up to 128GB for 120B+ Local Models

Oct 8, 2026·6 min read·By Rahish Saifi

Cover: Microsoft and NVIDIA RTX Spark preorders — Surface Laptop Ultra from $2,599, up to 128GB for 120B+ local models, ships Oct 16 — AI with Rahish

At a Microsoft Windows AI and Surface event at Dogpatch Studios in San Francisco on Wednesday, October 7, 2026, Microsoft and NVIDIA turned RTX Spark from a June teaser into something you can preorder. NVIDIA CEO Jensen Huang and Microsoft CEO Satya Nadella discussed running AI agents on Windows PCs in a fireside chat hosted by Sriram Krishnan.

The buyer-facing news: RTX Spark laptops and Surface hardware are open for preorder now, with many systems available October 16. Microsoft’s parallel pitch is hybrid intelligence—local models and agents on the PC, cloud when you need scale, and policy guardrails when software touches files and networks.

What Microsoft and NVIDIA announced

In Building Windows for hybrid intelligence, Microsoft EVP Pavan Davuluri frames Windows as an OS for agents: long-running software that reads documents, calls tools, and stays inside enterprise policy. NVIDIA’s October 7 post focuses on RTX Spark PCs and model sizes that did not fit on prior laptops.

Software shipped alongside the hardware story: Microsoft Execution Containers (MXC) are generally available on Windows 11, new local model packs are headed to RTX Spark, and Microsoft previewed DGX Station on Windows. On devices, Microsoft’s devices blog opened preorders for Surface Laptop Ultra and the Surface RTX Spark Dev Box.

What’s inside RTX Spark

NVIDIA describes RTX Spark as a Blackwell RTX GPU with up to 6,144 cores plus an up to 20-core Grace CPU, connected at 600 GB/s. The platform advertises up to 1 petaflop of FP4 AI performance (Microsoft’s footnote: theoretical FP4 with sparsity) and up to 128GB unified memory, with the full NVIDIA CUDA stack.

NVIDIA says RTX Spark can run models such as Qwen 3.8 Flash Next, a 125B model, locally without sending data to the cloud. For gaming, NVIDIA claims AAA titles at 1440p over 100 fps with DLSS 5, Reflex, and G-SYNC. Microsoft adds that Call of Duty is coming to RTX Spark in 2027, and that Gears of War: E-Day showcases RTX Spark.

RTX Spark laptops opened for preorder October 7 and ship October 16. NVIDIA says compact desktops go on sale in November; Microsoft lists RTX Spark dev boxes and mini desktops for later this year.

Surface Laptop Ultra price and configs

Surface Laptop Ultra is a 15-inch touchscreen built around RTX Spark. Microsoft says it can run AI models exceeding 120 billion parameters locally, with up to 128GB unified memory. Preorders start at $2,599 MSRP on Microsoft.com and select retailers worldwide, with availability from October 16.

Davuluri’s line from NVIDIA’s coverage is worth quoting directly: “We built Surface Laptop Ultra around NVIDIA RTX Spark. With up to 128 gigs of unified memory and up to a petaflop of AI compute, you can run models on this laptop that simply don't fit on a traditional machine.”

The full 128GB memory ceiling is tied to the top CPU SKU—not the $2,599 base unit. Map model sizes to configs before you standardize on hardware for a team.

Config	CPU	Memory	Storage	Price (MSRP)
Base	18-core	24GB unified	512GB	$2,599
Top	20-core	128GB unified	1TB	$5,899.99

Pricing detail from The Verge (October 7, 2026): the full 128GB memory ceiling requires the 20-core model. The Surface RTX Spark Dev Box lists at $5,999 MSRP, preorder exclusively on Microsoft.com in the U.S., shipping in November.

Microsoft also published vendor-commissioned comparisons against a MacBook Pro 16-inch with M5 Pro on preproduction hardware tested in September 2026. Per Microsoft’s footnotes, both the RTX Spark PCs and the MacBook Pro in the comparison were 64GB configurations, and the RTX Spark units were preproduction. The company claims up to 2.1× faster time to first token, 4.3× faster AI image generation, and 6.2× faster AI video generation—treat those as vendor claims on preproduction hardware, not independent benchmarks.

Other RTX Spark laptops you can preorder

Microsoft lists these RTX Spark PCs as available to preorder: ASUS ProArt P16 and P14, Dell XPS 16 Creator Edition (first XPS with RTX Spark), HP OmniBook Ultra 16, Lenovo Yoga 9n 2-in-1, MSI Prestige N16 Flip AI+, and Microsoft Surface Laptop Ultra. NVIDIA also names Acer and Gigabyte among system makers.

Surface Laptop Ultra anchors Microsoft’s story, but the OEM spread shows RTX Spark is a platform launch—not a single halo SKU. Check memory tiers per model; not every configuration ships with 128GB.

DGX Station comes to Windows

Microsoft and NVIDIA previewed the first deskside AI supercomputer for Windows: DGX Station with the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip, up to 748GB coherent memory, and up to 20 petaFLOPS FP4, supporting models up to roughly 1 trillion parameters locally. Systems such as the Dell Pro Precision with GB300 and HP ZGX Fury AI Station are slated for Windows later this year.

Microsoft describes DGX Station as a team “token factory” that can serve 32 or more simultaneous agents. Historically DGX Station ran Linux only; Windows support is new, with WSL still available. For a compact Grace Blackwell dev box at a different price point, see our write-up of the NVIDIA DGX Spark 64GB line.

MXC: a sandbox for AI agents on Windows

Microsoft Execution Containers (MXC) are now generally available on Windows 11. MXC provides policy-driven containment so organizations define which files and networks an agent may access, enforced at runtime. The feature integrates with Microsoft Agent 365 and Intune.

Agents already supporting MXC include OpenAI Codex, GitHub Copilot, OpenClaw, Replit, LM Studio, NVIDIA OpenShell, and Unsloth AI. Microsoft lists Anthropic Claude Code, Perplexity, Manus, Raycast, Box, and others as coming. Meta’s Muse for Windows personal agent is planned as a native app with MXC.

For security teams, MXC is the practical hook: agent access becomes a policy decision at runtime instead of a one-off exception per tool.

Local models and hybrid intelligence

Microsoft highlights MAI Code 1.1 Flash (137B total / 6.8B active parameters) running locally at 3-bit precision (~80% smaller) with 256K context. Upcoming on RTX Spark: an NVIDIA Nemotron model (70B+, 2-bit, just over 20GB) and DeepSeek V4 Flash (284B).

GitHub HydraFusion routing will tap local models on Windows, with an experimental preview in the GitHub Copilot app, Copilot CLI, and VS Code later in October. llama.cpp support is coming to Windows ML.

Copilot on Copilot+ PCs gains local context, local actions, and local models (with permission), rolling out in coming months. Windows Search on the taskbar picks up actions such as “switch to dark mode” or “mute,” rolling out to Windows Insiders on the experimental channel now.

Why it matters for AI builders

Three threads stand out if you ship agents on Windows:

Preorders are the news. June introduced RTX Spark; October 7 added prices and ship dates, with laptops arriving October 16.
Memory unlocks model choice. Up to 128GB unified memory on RTX Spark laptops is what makes 120B+ local models plausible on a portable machine—when you buy the right SKU.
Policy and routing are catching up. MXC is GA; HydraFusion and Copilot local features are rolling. Hardware can land before every software path is finished, so plan staged rollouts.

Size the fleet against your workload: RTX Spark laptops for individuals, the Surface dev box for a fixed desk, DGX Station when a team needs shared inference on Windows. Re-read Microsoft’s footnotes on FP4 claims and the M5 Pro benchmarks before you treat marketing charts as buying guides.