NVIDIA and Microsoft opened pre-orders for RTX Spark Windows laptops on October 7, 2026, with shipping set for October 16, and the paperwork that came with the launch contains the first hard numbers on how much power the N1X chip draws. According to StorageReview's breakdown of the published specs, laptop versions of the chip are rated 45 to 80W, the desktop version 140W, and every system is sold with the same "up to 1 petaflop" FP4 headline.
This guide focuses on what we have not covered yet: the power envelopes, how Windows splits the unified memory pool, what actually fits, and what the Q4 DGX Station for Windows adds. For the launch itself, see our posts on Microsoft's hybrid local-and-cloud agents and the $5,999 Surface RTX Spark Dev Box.
TL;DR: the spec table and what it means
| Question | Answer |
|---|---|
| When do laptops ship? | October 16, 2026; pre-orders opened October 7 |
| When do desktops ship? | November; the Surface RTX Spark Dev Box is $5,999 with 128GB |
| Laptop power envelope | 45 to 80W TDP |
| Desktop power envelope | 140W TDP for NVIDIA's N1X listing |
| Maximum unified memory | 128GB LPDDR5X (64GB on the smaller 5,120-core laptop chip) |
| GPU-addressable memory on a 128GB unit | 110GB in Microsoft's Task Manager example |
| Cheapest system | Surface Laptop Ultra from $2,599 |
| DGX Station for Windows | Coming in Q4 with GB300, up to 748GB coherent memory |
What are the N1X power envelopes?
NVIDIA's updated table separates laptops from desktops. Laptops get two variants: a 6,144-core Blackwell GPU with a 20-core Grace CPU and up to 128GB of memory, and a 5,120-core GPU with an 18-core CPU capped at 64GB. Both are rated 45 to 80W. Desktops get only the larger chip, up to 128GB, rated 140W.
Two details are worth noting. First, 140W is the same TDP NVIDIA publishes for the GB10 in DGX Spark, which fits NVIDIA's earlier framing that the two are essentially one system tuned for different platforms. We covered the original positioning in the N1X Computex coverage. Second, a laptop N1X runs at roughly a third to just over half of the desktop rating, yet all systems carry the same peak FP4 headline. Microsoft's footnotes describe that number as theoretical performance using sparsity, and NVIDIA does not publish a dense FP4 figure for N1X.
The practical reading: peak numbers will not separate these machines, thermals will. A 35W spread inside the laptop range, before any vendor tuning, is the earliest evidence that two RTX Spark laptops with the same chip can perform differently under sustained load.
A laptop outline with a small green chip on its base, illustrating RTX Spark N1X power limits in a thin laptop
Desktop wrinkle: 140W versus 100W
Microsoft's own Surface RTX Spark Dev Box page quotes a 100W thermal envelope for its aluminum chassis, 40W under the TDP NVIDIA lists for the desktop N1X. StorageReview notes the two may be measuring different things, a chip rating versus a chassis design point, but suggests buyers comparing compact desktops in November should ask each vendor what sustained power the box holds. That is good advice for any small-form-factor AI machine. Our config guide for the Surface Laptop Ultra goes into which memory tier is worth paying for.
How does Windows divide 128GB of unified memory?
The CPU and GPU share one pool, but Windows presents it like a discrete card. In a screenshot from Microsoft and GitHub's engineering post, a Surface Laptop Ultra with the 6,144-core chip reports 74.9GB of dedicated GPU memory, 35.5GB of shared GPU memory, 110GB of total GPU memory and 51.5GB of system memory. The GPU ceiling is therefore about 86 percent of the installed 128GB, and model weights, the KV cache that holds an agent's growing context, and runtime overhead all have to fit under it.
That matters most for smaller configurations. NVIDIA's marketplace lists 17 laptop configurations at pre-order, with memory options of 24GB, 32GB, 48GB and 128GB. On the 24GB and 32GB models, StorageReview estimates the room is for roughly 27B-class models at 4-bit and not much above. Every large-model claim in the launch material assumes the 128GB configuration.
What fits? Parameter counts and bits per weight
Vendors quote different ceilings because they count bits differently. NVIDIA's page says models up to 120 billion parameters with a 1 million token context. Microsoft says models exceeding 120 billion and names MAI Code 1.1 Flash at 137 billion and DeepSeek V4 Flash at 284 billion as coming to RTX Spark. Dell quotes 200 billion.
| Model | Parameters | Footprint per the sources |
|---|---|---|
| MAI Code 1.1 Flash | 137B total, 6.8B active | 53GB at roughly 3.3 bits per weight; 75.5GB peak at 256K context |
| NVIDIA Nemotron (upcoming) | Over 70B | Just over 20GB at 2-bit, per Microsoft |
| DeepSeek V4 Flash | 284B | Footprint not stated |
| Qwen 3.8 Flash Next | 125B | Footprint not stated |
The best-documented case is MAI Code 1.1 Flash, which we broke down in our MAI Code 1.1 Flash local Windows guide. Microsoft reports the on-device build at 70.8 percent on SWE-Bench Verified against 72.6 percent for the BF16 cloud model. By StorageReview's arithmetic, a 284B model would need about 3.1 to 3.3 bits per weight or fewer to squeeze under a 110GB ceiling before any context loads, so treat the DeepSeek claim as plausible but not demonstrated in the published material.
How fast is it? Read the footnotes
Microsoft's throughput data comes from a Surface Laptop Ultra running a Windows Arm64 llama.cpp build with the CUDA backend and speculative decoding on a synthetic code-generation workload. StorageReview read the unlabeled chart at about 63 tokens per second at a 2K prompt, falling to roughly 39 at 256K. Prompt processing is quoted at 923.5 tokens per second at 64K context and 769.8 at 128K, which means a cold 64K prompt takes about 71 seconds before the first token. For agent work that prefill figure matters more than decode speed.
The Mac comparison needs care too. Microsoft and NVIDIA compare against a 16-inch MacBook Pro with M5 Pro and 64GB, claiming up to 2.1x faster time to first token, 4.3x faster image generation and 6.2x faster video generation. Footnotes say the tests ran in September on preproduction RTX Spark systems, no M5 Max was included, and no decode tokens-per-second comparison was published. These are vendor numbers, so wait for independent reviews after October 16.
Microsoft Execution Containers: the sandbox that went GA
Alongside the hardware, Microsoft declared Microsoft Execution Containers generally available on Windows 11. TechCrunch reports that Satya Nadella said the capability would reach all Windows 11 users. A developer writes a JSON policy that lists which files an agent can change or only read, what network traffic is allowed, and whether it can touch the desktop; the policy sits outside the workload, so generated code cannot grant itself more access.
| MXC backend | Platforms | Mechanism |
|---|---|---|
| Process container | Windows, macOS, Linux | AppContainer, Seatbelt, Bubblewrap |
| Session container | Windows | Separate account and desktop session |
| WSL container | Windows | Linux environment through WSL |
| MicroVM | Windows, Linux (experimental) | Hardware-enforced virtual boundary |
Small shapes moving safely inside a low green-walled square, illustrating an agent sandbox policy in Microsoft Execution Containers
The caveats are as important as the headline. StorageReview points out that agent identity in Entra and extended Agent 365 controls for local agents are still "coming soon," so containment shipped but fleet-level governance did not. GitHub's own description says Copilot uses the base process-container tier for shell commands, while built-in file tools run in the harness and remote MCP servers sit outside the local sandbox. If you run coding agents on a Windows machine, our guide on sandboxing Claude Desktop covers the same logic on another platform.
DGX Station for Windows: the Q4 desk-side option
NVIDIA's page now reads "Coming in Q4" for DGX Station for Windows, with Dell and HP named first. It uses the GB300 Grace Blackwell Ultra desktop superchip: 252GB of HBM3e GPU memory, 496GB of LPDDR5X on the Grace CPU, up to 748GB of coherent memory, 20 petaflops of FP4 with sparsity, 800Gb/s networking, and a 1,600W power draw that needs a 20A circuit. Microsoft describes uses ranging from a personal AI supercomputer for models up to about 1 trillion parameters to a shared "token factory" box. We compared the Linux version of that hardware in the Gigabyte W775 workstation versus DGX Spark guide.
Which RTX Spark should you buy?
- Casual local coding agent, small models: a 5,120-core laptop with 32GB or 48GB is enough for 27B-class models, but do not expect the 137B headline.
- 137B-class coding model or 100K+ context: buy the 128GB, 6,144-core configuration, and budget for the $2,599-and-up Surface Laptop Ultra tier or equivalents.
- Sustained workloads: wait for reviews that report sustained wattage; a 45W laptop and an 80W laptop are different machines.
- Clusters or Linux-first work: DGX Spark remains the option with ConnectX-7 networking; see our note on PAIR and Hermes Desktop on RTX Spark for the software path.
What is still unknown
- Dense FP4 throughput and sustained wattage for each laptop.
- Real-world prices outside the Surface line; only the $2,599 starting price and the $5,999 desktop are firm.
- Independent decode speed against Apple silicon with matching models and quantization.
- Whether the 284B and 200B claims run at usable context lengths on shipping hardware.
Specifications and dates reflect launch coverage as of October 9, 2026 and may change as OEMs publish final configurations.