Running a 120-billion-parameter AI model on a portable machine has, until now, required meaningful trade-offs in speed, size, or cost. Microsoft and Nvidia want to change that. Preorders for the Surface Laptop Ultra opened October 7, 2026, with units available starting October 16. The machine is built on Nvidia’s RTX Spark platform, starts at $2,599, and the flagship configuration costs $5,899.99. Microsoft is framing this launch as the beginning of an “agentic PC” era, a claim worth examining alongside the AI chips and hardware itself.
What the Hardware Actually Is
RTX Spark brings together a Blackwell GPU, a Grace CPU, and a shared memory pool in a portable form factor, with Microsoft claiming up to one petaflop of AI compute.
RTX Spark combines an Nvidia Blackwell GPU, an Nvidia Grace CPU built on Arm architecture, and a unified memory pool that CPU and GPU workloads share rather than compete over. Microsoft claims local inference support for models up to 120 billion parameters.
The confirmed configurations are:
- Base: 18-core CPU, 5,120-core GPU, 24GB unified memory, 512GB storage, starting at $2,599
- Top config: 20-core CPU, 6,144-core GPU, 128GB unified memory, 1TB storage, at $5,899.99
- Display: 15-inch HDR panel, up to 2,000 nits brightness, under 4.5 lbs, under 18mm thick
- Ports: three USB-C, USB-A, HDMI, SD card reader, headphone jack
All specifications and performance figures come from Microsoft and Nvidia, not independent benchmarks. The 120-billion-parameter figure is a capability ceiling, not a daily-use guarantee. Quantization, memory bandwidth, and application optimization will determine what actually runs well.
What “Agentic” Actually Means Here
Microsoft’s approach is hybrid: keep routine and sensitive tasks on the device, then route heavier workloads to cloud infrastructure.
The architecture is not purely local. The intended model keeps privacy-sensitive and latency-critical tasks on device while escalating more demanding work to cloud services. GitHub’s HydraFusion is designed to route coding tasks between local and cloud models along these lines, though it was announced as an experimental preview and is not yet broadly available.
Microsoft Execution Containers are sandboxed environments that limit which files and networks an agent can reach, with policies enforced at runtime. Supporting features also include action-oriented Windows Search commands, Copilot access to local context, and simplified local model installation.
The unresolved questions are significant. Agent reliability, clear user consent flows, permission auditing, and enterprise compliance are not yet settled. The “agentic PC” label is a product strategy and an industry prediction; the software ecosystem has not yet validated it. Exploring AI-powered tools available today can help gauge how far the ecosystem has actually come.
Windows on Arm and Gaming
The Surface Laptop Ultra is an Arm-architecture Windows machine, and gaming support at launch is limited.
RTX Spark runs on Nvidia’s N1X family, placing the Surface Laptop Ultra within the Windows on Arm ecosystem rather than conventional x86. According to The Verge, Call of Duty compatibility is scheduled for 2027. Anti-cheat compatibility and driver maturity remain practical concerns for buyers prioritizing gaming.
A Growing Platform
The RTX Spark lineup extends beyond the Surface Laptop Ultra, with a developer workstation and third-party systems on the way.
Microsoft also opened preorders for the Surface RTX Spark Dev Box, a desk-side developer system priced at $5,999 and scheduled to ship in November 2026. Asus, Dell, HP, Lenovo, and MSI are building RTX Spark systems as well, expanding the platform beyond Microsoft’s own hardware.
Who This Is Actually For
The Surface Laptop Ultra is built for developers, AI researchers, and creators with GPU-heavy workloads, not general productivity users.
Developers running local models, AI researchers, and creators doing GPU-intensive work have a credible case for this machine, consistent with Microsoft’s stated target audience. For users whose daily work is email, browser tabs, and light creative tasks, the premium is difficult to justify , especially compared to alternatives like the Mac Mini. The “agentic” bet pays off only if the software, the permissions framework, and the developer ecosystem follow the hardware. That part remains to be seen.
