Surface Laptop Ultra preorders opened at $2,599 on October 7, giving Microsoft and Nvidia a high-end showcase for running large AI models and agents directly on Windows. Microsoft says the RTX Spark-powered laptop will become available October 16 through its store and selected retailers.

The commercial launch establishes four key details:

• The entry price is $2,599.
• Retail availability begins October 16.
• Configurations offer up to 128 GB of unified memory.
• Microsoft Execution Containers are generally available.

Surface Laptop Ultra Preorders Start at $2,599

Microsoft’s device announcement moves Surface Laptop Ultra from an earlier preview into a commercial release. The company is accepting orders now and lists October 16 as the start of availability, while commercial configurations are also open for preorder.

The laptop is built around Nvidia’s RTX Spark platform and targets developers, creators and teams that want substantial AI compute without relying on a remote data center for every task. Microsoft positions it as its most powerful Surface laptop rather than a mainstream replacement for lower-cost Surface models.

Reuters reported that the highest configuration reaches $5,899 with a 20-core processor, 128 GB of memory and one terabyte of storage. The entry configuration costs more than Apple’s base MacBook Pro, making price a major constraint on broad adoption.

Microsoft also opened preorders for the Surface RTX Spark Dev Box, a compact desktop aimed at AI developers. It costs $5,999 and is scheduled to begin shipping to U.S. customers in November. The desktop is designed for continuous local experimentation rather than mobile work.

RTX Spark Brings One Petaflop to a Laptop

Surface Laptop Ultra supports up to 128 GB of unified memory shared dynamically between its CPU and GPU. Microsoft says the hardware can run AI models exceeding 120 billion parameters locally, though usable model size depends on configuration, precision, context length and the memory consumed by other software.

Nvidia rates RTX Spark for up to one petaflop of sparse FP4 AI performance. That theoretical figure describes low-precision acceleration under specific conditions; it is not a general measure of application speed. Actual inference results depend on model architecture, software optimization, memory bandwidth and power limits.

Microsoft’s Windows announcement reports up to 2.1 times faster time to first token, 4.3 times faster image generation and 6.2 times faster video generation than a 64 GB MacBook Pro with M5 Pro. The comparisons used preproduction Windows PCs and company-selected models and workloads.

Those tests do not establish a universal advantage across every model or creative application. Microsoft and Nvidia used different benchmarks for language, image and video performance, and both companies state that results vary by device configuration and usage.

Related Research

• Nvidia Launches Open Agent Safety Platform With OpenShell and Sentry
• Axelera Launches Europa AI Chip for Dell and Supermicro Systems

Microsoft Execution Containers Limit Agent Access

The hardware announcement is paired with the general availability of Microsoft Execution Containers, or MXC. The system creates policy-controlled boundaries for model-generated code, plugins, tools or complete agents, limiting which files, network destinations and operating-system resources they can reach.

Policies remain outside the agent workload, so an agent cannot grant itself more access. Microsoft says developers can describe required resources through a unified configuration while MXC maps those rules to supported containment backends on Windows, macOS and Linux.

Windows 11 offers a session container that runs an agent under a separate account with an isolated desktop, clipboard, interface and input boundary. Other options include lightweight process containers, Windows Subsystem for Linux containers and an experimental microVM for higher-risk workloads.

Containment does not prove an agent is safe or correct. It reduces the damage a mistaken or compromised workload can cause only when policies are narrowly designed, the chosen backend provides sufficient isolation and organizations monitor what agents attempt to do.

Microsoft Pushes Hybrid AI Beyond the Cloud

Microsoft calls its broader strategy “hybrid intelligence,” with software deciding whether a task should run on the device or use cloud-scale models. Local execution can improve privacy, latency and cost control, while cloud services remain useful for workloads that exceed a laptop’s memory or compute capacity.

The company says upcoming Windows capabilities will use Microsoft Entra identities to distinguish agent activity from user activity and extend Agent 365 controls to local agents. Those features are separate from MXC’s current general availability and will roll out over time rather than arriving universally with the laptop.

Surface Laptop Ultra is one part of a larger RTX Spark PC lineup from Asus, Dell, HP, Lenovo and MSI. Those systems are also available for preorder and are scheduled to start shipping October 16, giving Nvidia multiple routes into a premium PC segment traditionally led by Intel and AMD processors.

The near-term test is whether developers and creators will pay workstation prices to keep more AI work local. Performance on sustained workloads, battery life, software compatibility and the effectiveness of agent containment will matter more than peak theoretical specifications once the first systems reach customers.