NVIDIA RTX Spark laptops: 128GB on Windows, and why to wait a quarter

What RTX Spark is

RTX Spark is NVIDIA’s name for its N1X laptop chip: a 20-core Grace Arm CPU and a 6,144-core Blackwell RTX GPU on one package, sharing up to 128GB of LPDDR5X unified memory. A smaller variant has an 18-core CPU, a 5,120-core GPU and tops out at 64GB. Laptop TDP is 45 to 80W. It runs Windows 11 on Arm.

For local AI, the one line that matters: this is the first Windows laptop where a 70B model fits in memory next to an NVIDIA GPU. Until now that meant a Mac, or an AMD Strix Halo machine without CUDA.

The laptops

Pre-orders opened October 7, 2026 and the first units reached US buyers October 16. None of these are sold on Amazon as of this writing; they’re on the makers’ stores, Best Buy and Micro Center. There are no affiliate links on this page.

Laptop	Memory options	Notes
ASUS ProArt P16	24, 32, 36, 48, 64, 128GB	16-inch 4K OLED, 1.77kg, up to 99Wh. 64 and 128GB mainly ASUS eShop and Best Buy; 128GB also Micro Center
ASUS ProArt P14	same line-wide list	14-inch 3K OLED, 1.48kg, 90Wh. ASUS’s lightest ProArt
Microsoft Surface Laptop Ultra	24GB base, up to 128GB	15-inch 120Hz touch, under 4.5lb. One 128GB unit reported a 110GB GPU memory ceiling in Windows
Dell XPS 16 Creator Edition	32, 64, 128GB	LPDDR5X-9400, 99Wh, 2.02kg. Best Buy first, Dell.com later
HP OmniBook Ultra 16	up to 128GB	99Wh, HP claims up to 17 hours
Lenovo Yoga 9n 2-in-1	up to 64GB	The 18-core / 5,120-core chip
MSI Prestige N16 Flip AI+	32GB (launch SKU)	20-core / 6,144-core chip, tandem OLED, 99.9Wh. Ships November 6

Starting configurations (24 or 32GB) sit in the $1,000 to $3,000 tier. 128GB configurations weren’t priced at pre-order and will be No ceiling. Prices are moving with memory supply, so check the maker’s store rather than any number you read in October.

What it means for running models

Memory. 128GB holds a 70B model at 4-bit (about 40GB) with room to spare, or gpt-oss-120B in its native MXFP4 (about 59GB). Windows reserves some; expect roughly 110GB usable by the GPU on a 128GB machine. 24GB is 7B-to-14B territory.

Bandwidth, which decides generation speed. NVIDIA quoted about 300GB/s in June and hasn’t published a laptop figure since. Working it out from Dell’s listed LPDDR5X-9400 on a 256-bit bus gives 300.8GB/s. Nobody has measured it on a shipping laptop. For comparison: Strix Halo laptops are 256GB/s, the MacBook Pro M5 Max is 614GB/s. So expect text generation close to a Strix Halo machine and roughly half a MacBook Pro M5 Max.

Prompt processing, which depends on compute. This is where RTX Spark should win. The desktop DGX Spark (same architecture, 273GB/s) reads a 2,048-token prompt into gpt-oss-120B at about 2,400 tokens a second and generates at about 59, measured with llama.cpp. A dense 70B at 4-bit on that bandwidth works out to under seven tokens a second. Laptop numbers will be lower at 45 to 80W; treat the DGX Spark figures as a ceiling, not an estimate.

The catch: Windows on Arm

Every one of these runs Windows on Arm, and the local-AI toolchain there is weeks old. CUDA Toolkit 13.4 is the first with native Windows-on-Arm64 support. NVIDIA’s porting page links Arm64 builds of llama.cpp and Ollama, and Microsoft used an Arm64 llama.cpp build with the CUDA backend for its own benchmarks. What’s not verified as of October 10: whether Ollama’s Arm64 build uses the GPU on these machines (one early report says CPU only), whether LM Studio and ComfyUI have Arm64 builds that see the GPU, and how x86 tools behave under emulation. If your workflow is “install LM Studio, click a model,” check that it works on Arm before you order.

Should you buy one now?

Not yet, unless you’re a developer who wants CUDA on Arm and expects to build things yourself. Reasons to wait a quarter:

No independent review has measured tokens per second on any of these laptops.
The bandwidth number is derived, not published.
The software stack is a developer preview in practice, whatever the labels say.
128GB pricing wasn’t public at pre-order, and memory prices are rising.

If you need a 128GB laptop today, the MacBook Pro M5 Max generates text about twice as fast and the software is mature. If you need Windows and x86 compatibility today, a Strix Halo laptop (HP ZBook Ultra G1a, ASUS ROG Flow Z13) does the same job as RTX Spark at the same bandwidth with tools that already work. RTX Spark’s case is CUDA, prompt-processing speed and Windows in one machine; that case gets made or unmade by the first real benchmarks.

Three 128GB laptop platforms

	RTX Spark laptop	MacBook Pro 16 M5 Max	Strix Halo laptop
Memory bandwidth	~300GB/s (derived, unmeasured)	614GB/s	256GB/s
Text generation	about Strix Halo, about half the Mac	fastest	about RTX Spark
Prompt processing	likely fastest (Blackwell, FP4)	good	weakest of the three
Software today	CUDA on Arm, immature	MLX, Metal, Ollama, LM Studio, all mature	x86 Windows or Linux, llama.cpp Vulkan/ROCm, mature
OS	Windows 11 on Arm	macOS	Windows or Linux, x86
On sale	October 16, 2026	yes	yes