RTX Spark N1X: The Windows Numbers That Matter Are TDP, Memory Ceiling, and Prefill

On October 7, NVIDIA and Microsoft opened Windows pre-orders for RTX Spark laptops (availability October 16; compact desktops in November) and previewed DGX Station for Windows for Q4. The useful release is not the “1 petaflop FP4” slide. It is the first public split of N1X power envelopes, how Windows presents the unified memory pool, and a vendor decode/prefill curve for a 137B coding MoE on a laptop.

I care about this stack the same way I care about Halo Station and GB10 deskside boxes: if you run local agents or size a small lab, you need watts, addressable bytes, and tokens per second — not another “personal AI PC” paragraph.

Two chips, three power envelopes

NVIDIA’s product table now names the Windows superchip N1X and separates laptop from desktop. StorageReview transcribed the table that matches the NVIDIA specs page:

The desktop 140W figure is the same chip-level TDP NVIDIA lists for GB10. Laptop N1X therefore runs at roughly 32%–57% of that desktop rating while every SKU still markets “up to 1 petaflop” of sparse FP4. Microsoft’s Surface footnotes call that number theoretical FP4 with sparsity; NVIDIA still does not publish a dense FP4 figure for N1X.

Microsoft’s Surface RTX Spark Dev Box page quotes a 100W thermal envelope in an aluminum chassis — 40W under NVIDIA’s desktop TDP. Treat that as a chassis design point until review units settle sustained power under a long agent session.

Memory bandwidth is another platform delta. Dell’s XPS sheet lists LPDDR5X at 9,400 MT/s; on a 256-bit interface that is about 300.8 GB/s, matching NVIDIA’s June ~300 GB/s quote and roughly 10% above DGX Spark’s 273 GB/s. CPU↔GPU is NVLink-C2C at 600 GB/s. Media: 1× 9th-gen NVENC and 1× 6th-gen NVDEC; display path HDMI 2.1b / DisplayPort 2.1b.

OEM pre-order lists already show how much of the market sits below the headline: of fifteen marketplace configs StorageReview checked, seven used the 5,120-core part, with memory at 24 / 32 / 48 / 128GB. ASUS opened ProArt P16 / P14 pre-orders October 7 (US units from October 16), starting at $2,599.99, with 64GB and 128GB mostly through ASUS eShop / Best Buy / Micro Center for the top bin.

Windows does not give the GPU all 128GB

Microsoft’s Surface footnote is explicit: unified memory is dynamically shared, and the maximum addressable by the GPU is less than the installed total. A Task Manager capture on a Surface Laptop Ultra with the 6,144-core part (reported by StorageReview from Microsoft/GitHub engineering) showed:

• 74.9GB dedicated GPU memory
• 35.5GB shared GPU memory
• 110GB total GPU memory
• 51.5GB system memory

Dedicated + system ≈ 126.4GB. Windows is presenting the pool the way it presents a discrete card: a fixed dedicated allocation, with borrow from system memory up to a cap. On that unit the GPU ceiling is 110GB, about 86% of 128GB. Weights, KV cache, OS, apps, and the inference runtime all fight for that ceiling.

Parameter-count marketing spreads for that reason. NVIDIA says models up to 120B with 1M context and cites Qwen 3.8 Flash Next at 125B. Microsoft says models exceeding 120B, then names MAI Code 1.1 Flash at 137B and DeepSeek V4 Flash at 284B. Dell has quoted 200B. The spread is bits-per-weight, not magic silicon.

Microsoft’s MAI Code 1.1 Flash numbers are the most complete vendor package:

Prefill is the agent number. A cold 64K prompt is ~71 s before the first output token; cold 128K is ~170 s. Keeping the model loaded does not make TTFT free.

MacBook comparison footnotes (Microsoft-commissioned llama.cpp Qwen3.5 27B Q4_K_M TTFT; NVIDIA ComfyUI FLUX.2 / LTX tests at NVFP4) used 64GB preproduction RTX Spark systems against an M5 Pro 64GB, not M5 Max, and did not publish a matched decode tokens-per-second table. Useful as marketing, not as a bake-off.

MXC: containment is GA; identity is not

The same event made Microsoft Execution Containers (MXC) generally available on Windows 11. Policy sits outside the workload: files (read/write/deny), network, UI, and process launch settings in JSON, mapped onto backends:

Modes: Enforcement, Learning (block + JSON activity report), Permissive (allow + record). Activity reports for process containers on Windows only. Entra agent identity and Agent 365 fleet controls for local agents are still “coming soon” — until then, security tools see agent activity under the signed-in user even when the container limits reach.

GitHub Copilot’s documented integration uses the base process container for shell / local MCP / language servers; built-in file tools stay in the harness (policy-checked, not OS-isolated); remote MCP is outside the local sandbox. NVIDIA wired OpenShell into MXC for file/inference policy, network controls, credentials, and OCSF audit. Agents already listed with MXC support include Copilot, Codex, OpenClaw, Replit, LM Studio, and Unsloth; Claude Code and Hermes are among those coming.

For cluster people this is the Windows equivalent of a pod security boundary: useful when the model can shell out, incomplete until identity separates the agent from the human.

DGX Station for Windows: GB300 on the enterprise desktop

The Q4 preview is not N1X. It is the GB300 Grace Blackwell Ultra deskside box under Windows, with Linux toolchains through WSL. StorageReview’s table from NVIDIA’s page:

Microsoft frames two uses: a personal frontier box (models up to ~1T parameters in their footnote) and a “Windows token factory” for 32+ concurrent agents on one deskside node. First named systems: Dell Pro Precision with GB300 and HP ZGX Fury AI Station; partners also include ASUS, Gigabyte, MSI, Supermicro.

Open question from June still open: which serving stacks run native on Windows here vs under WSL, and what WSL costs on a 72-core Grace host with two 400Gb/s ports. That is the first lab ticket when a Windows unit ships.

What I would put on the whiteboard

1. Buy the memory bin you need. 24–48GB SKUs are 27B-class at 4-bit territory once OS + KV land. The 120B+ claims assume 128GB and aggressive quantization.
2. Budget the GPU ceiling, not the sticker RAM. Plan against ~110GB addressable on the Surface-class 128GB capture until OEM docs publish their own split.
3. Measure prefill first for agent work. Microsoft’s 923.5 / 769.8 tok/s prompt numbers matter more than the ~60 tok/s decode headline.
4. Do not confuse N1X desktops with DGX Spark clustering. Surface Dev Box lists USB-C / HDMI / Ethernet and no ConnectX-7; GB10 DGX Spark still has the 200Gb/s NIC path for two-box 405B experiments.
5. Treat MXC as process sandbox GA, not full agent identity. Wire least-privilege policies in Learning/Permissive before Enforcement.

Surface Laptop Ultra starts at $2,599; Surface RTX Spark Dev Box is $5,999 with 128GB (US Microsoft.com, ships November) — $1,000 above the $4,999 64GB DGX Spark with twice the memory and none of the high-speed clustering ports. OEM laptops from Acer, ASUS, Dell, HP, Lenovo, MSI, Gigabyte fill the rest of the October 16 wave.

If you already run GB10 or argue HBM budgets on deskside boxes, this is the October 7 drop to update the comparison table — with TDP ranges, Windows memory accounting, and prefill written in permanent marker next to the sparse FP4 headline.

Related: AMD’s Threadripper Halo Station: Spec Sheet, HBM Math, and the DGX Station Bet It Is Making · Scality AI Inference Factory: Object Storage Over RDMA as the Shared KV Cache · Upscale Token Fabric: SkyFabriX at 115.2 Tbps Meets Spectrum-X Scale-Out Under One OS · Astera Leo X: Fabric-Attached DDR for KV Cache — 62% Faster Warm TTFT on H200 Math.

Anpoo Sivanadi, Staff Software Engineer