RTX Spark brings 128 GB of unified memory to Windows AI PCs
Laptop preorders are open ahead of an October 16 release, while compact desktops follow in November. The platform combines high-memory local AI hardware with a new Windows environment for background agents.
By
Marcus Delaney
NVIDIA opened RTX Spark laptop preorders on October 7, 2026, with systems arriving October 16 and compact desktops following in November.
RTX Spark release dates and manufacturers
The first wave covers more than one machine or form factor. Acer, ASUS, Dell, HP, Lenovo, Microsoft, MSI and Gigabyte are expected to offer RTX Spark systems ranging from thin laptops to compact desktops. Neither geographic availability nor model-by-model pricing has been disclosed.
| RTX Spark product | Order or release timing | What is established |
|---|---|---|
| Laptops | Preorders opened October 7, 2026; availability scheduled for October 16 | Models are planned across eight manufacturers |
| Compact desktops | Scheduled for November 2026 | Small desktop systems extend the platform beyond portable PCs |
Unite.AI reports those dates and identifies the participating PC manufacturers. Exact configurations may therefore differ even though the systems share the RTX Spark platform name.
Why unified memory matters for local AI
At the upper end, RTX Spark combines a Blackwell RTX GPU with as many as 6,144 cores and a Grace CPU with as many as 20 cores. The two components are connected through a 600 GB/s link, while the maximum configuration provides 128 GB of unified memory and an advertised one petaflop of FP4 AI performance.
Unified memory is central to the proposition because a local model must fit within available memory before its raw compute performance becomes useful. NVIDIA says RTX Spark can run models containing up to 120 billion parameters with a context of one million units. These are platform limits presented by NVIDIA, not assurances that every application or model will reach the same speed or capacity.
Microsoft’s Surface Laptop Ultra provides a concrete example of the top specification, with up to 128 GB of unified memory and up to one petaflop of AI compute. That is much more memory than a conventional discrete GPU can usually access directly, making the system relevant to developers, researchers and creators whose models or datasets exceed ordinary laptop GPU memory. Pricing, battery life and individual model specifications have not been announced in the supplied release details.
What users could run on RTX Spark
RTX Spark is positioned as both a local AI development system and a Gaming PC platform rather than hardware restricted to one workload.
- Local model execution: large models can remain on the PC instead of requiring remote execution for every task, subject to the software’s compatibility and the selected system’s memory.
- CUDA development: RTX Spark runs NVIDIA’s CUDA platform natively. NVIDIA says models, tools and workflows can move between RTX Spark and DGX Station without being rewritten, which could give developers a common path from a portable machine to a larger workstation.
- 1440p gaming: NVIDIA advertises more than 100 frames per second in AAA games at 1440p using DLSS 5, Reflex and G-SYNC. That figure is a vendor performance claim rather than a universal result across games and settings. Readers comparing rendering technologies can also see how Vulkan affects performance in Dota 2.
The practical beneficiaries are therefore users who need unusually large local memory for AI work but still want Windows software and laptop or compact-desktop form factors. Smaller models and lighter inference tasks may not require this hardware class, while workloads above its memory ceiling can still require a larger workstation or remote infrastructure.
Windows agents gain an operating system-controlled environment
The software change is Microsoft Execution Containers, or MXC, which Microsoft has moved into general availability. MXC provides an operating system-level environment where AI agents can continue running in the background under Windows control. This differs from a conventional assistant that acts only while its application is open and active.
Persistent operation could support agents that monitor or complete longer tasks without remaining in the foreground. The important architectural change is that Windows supplies the execution environment rather than leaving every developer to create an independent mechanism for background operation.
The NVIDIA Blog describes MXC as infrastructure intended to let Windows secure, observe and govern agent activity. That describes Microsoft’s control model, but it does not establish that every agent or use case will be secure. Actual behavior will also depend on the agent, its permissions and the software built around the container.
DGX Station for Windows serves a different tier
DGX Station for Windows is a deskside AI supercomputer rather than another mainstream RTX Spark PC. It brings GB300 Grace Blackwell hardware into the Windows ecosystem and raises the local resource ceiling substantially. Linux AI toolchains remain accessible through Windows Subsystem for Linux, allowing teams to retain Linux-oriented workflows while using the Windows-based system.
| Capability | RTX Spark | DGX Station for Windows |
|---|---|---|
| Form and audience | Laptops and compact desktops for local development, creation and Gaming | Deskside system for substantially larger professional AI workloads |
| Maximum memory | Up to 128 GB unified memory | Up to 748 GB coherent memory |
| Advertised FP4 AI compute | Up to 1 petaflop | Up to 20 petaflops |
| Advertised local model scale | Up to 120 billion parameters | Up to trillion-parameter scale |
| Linux tooling | Native CUDA platform is specified | Linux AI toolchains available through WSL |
| Availability | Laptops October 16; desktops in November 2026 | Fourth quarter of 2026 |
ASUS, Dell Technologies, GIGABYTE, HP, MSI and Supermicro are expected to offer DGX Station for Windows. Its 748 GB memory ceiling is nearly six times RTX Spark’s 128 GB maximum, while its advertised FP4 compute reaches 20 times the smaller platform’s peak. Those differences place it closer to institutional and advanced professional development than to a general-purpose Windows laptop.
Featured image. Source: Pexels. Credit: Daniil Komov. License: Pexels License.
