RTX Spark Laptops: How Far Does Local AI Go?
Want to run large AI models on your own laptop, no internet required? RTX Spark is the laptop pitched at exactly that. NVIDIA is back in the laptop chip business — RTX Spark went on sale on October 16, 2026 — and the headline is not the graphics. It is the memory: up to 128GB of unified memory. But the $2,599 starting price and the price of a real configuration are different stories. This article separates the launch claims from independent testing to see how far local AI actually goes.
The short version
On sale Oct 16 — the N1X pairs a Grace CPU with a Blackwell RTX GPU on one chip, with up to 128GB of unified memory
The $2,599 starting price buys the 24GB base model. The 128GB configuration runs to $5,899.99
Independent testing runs 8B-class models comfortably; for larger models, memory and test conditions decide
What arrived
At the heart of RTX Spark is the N1X, a superchip that puts a Grace CPU and a Blackwell RTX GPU on one die. These are Windows laptops, but they run Windows on Arm, not x86.
The chip comes in two tiers. The higher tier has a 20-core CPU, a 6,144-core CUDA GPU, and up to 128GB of unified memory. The lower tier has an 18-core CPU, 5,120 CUDA cores, and up to 64GB. Laptop power runs 45–80W. One thing to note: NVIDIA does not print a memory bandwidth figure in its official spec sheet. The roughly 300GB/s figure in circulation is converted from the memory spec (LPDDR5X-9400) and used across press coverage.
The launch makers are ASUS, Dell, HP, Lenovo, Microsoft, and MSI — nine models in all (per the lineup listed in NVIDIA's release). On October 15, NVIDIA also unveiled its own model, Nemotron 3.5 Bolt 74B-A9B: a 74B MoE with 9B active parameters, quantized to 2-bit so it fits in 24GB of memory, built to run local agents on RTX Spark.
Read the price in three steps
The official starting price is $2,599. But at that price, the Surface Laptop Ultra base model pairs the lower-tier chip with 24GB of memory and a 512GB SSD.
Memory climbs fast. In Surface terms, 64GB costs $4,299.99 and 128GB costs $5,899.99. Going from 64GB to 128GB alone adds $1,600, and the 128GB model cannot be configured with a 2TB SSD (per the current Microsoft Store configuration table).
That said, the "starting line" depends on your purpose. The 24GB base model is not useless: NVIDIA's own Bolt was designed to fit in 24GB, and 8B-class models can run locally in 24GB too. Only when the goal is running large models with room to spare do the 64GB — and then 128GB — configurations come into view. That is why judging by the starting price alone misleads.
The numbers from independent testing
The launch numbers from NVIDIA and Microsoft are flashy: 1 petaflop at FP4, AAA games at 1440p above 100 frames, and image generation claimed 4.3x faster than a MacBook Pro with M5 Pro.
Conditions apply. A Microsoft Store footnote states the 1 petaflop is a theoretical figure using sparsity; the 100-frame claim assumes DLSS 5 and similar features, with no game titles or settings published; and the 4.3x comparison is the maker's own measurement on a preproduction 64GB unit.
Independent measurements read differently. On a top-tier Surface Laptop Ultra, Procyon benchmarks put Llama 3.1 8B at roughly 49–57 tokens per second and Phi-3.5-mini at 72–74 (measured by Windows Central and Digit).
A Qwen 27B-class model was reported at a median of about 25 tokens per second under Ollama (measured by NAG on an ASUS ProArt P16 with 128GB; the article does not state the quantization, so the figure cannot be pinned to a specific quantization setting). Windows Central rated the graphics at RTX 4070–5060 laptop class on 3DMark, and a Premiere Pro export — on a beta app build — was slower than an RTX 5060 laptop.
In short: 8B-class models run comfortably. 27B-class models were reported around 25 tokens per second within the machines and conditions tested. Token speed is heavily influenced by memory bandwidth, but bandwidth alone does not decide it — model size, quantization, the software stack, and the workload all matter.
What the same money buys elsewhere
| Product | Memory | Bandwidth | Starting price |
|---|---|---|---|
| Surface Laptop Ultra 128GB (RTX Spark) | 128GB | ~300GB/s (as reported, not officially specified) | $5,899.99 |
| AMD Strix Halo mini PC 128GB | 128GB | 256GB/s | ~$3,449 and up (Presenc survey) |
| Mac Studio M5 Max | Up to 128GB (higher GPU option) | 614GB/s (higher option; base 460GB/s) | From $2,499 (36GB base) |
The Mac Studio has twice the bandwidth, but it is a desktop and has no CUDA. The development ecosystem still treats NVIDIA's CUDA as the standard, which matters for local AI work. The Mac Studio's 128GB memory upgrade was reported at $2,000 (AppleInsider). AMD's Strix Halo is x86, so existing Windows software compatibility is better and it costs less, but its lower bandwidth can show up as a speed difference on large models. Qualcomm's Snapdragon X2 line also reaches 128GB of memory, but its local LLM performance has to be compared model by model, runtime by runtime, under stated test conditions.
RTX Spark's place is clear: a laptop option that offers CUDA and up to 128GB of unified memory. The price is the cost.
Before you buy
First, this is Windows on Arm. Some programs cannot run even through emulation at launch, and some Windows AI features (Recall among them) are not supported yet. Second, battery: roundup coverage of the early reviews reports web browsing figures of 10 hours 15 minutes to 43 minutes (as measured by Tom's Guide and Tom's Hardware), shorter than a MacBook Pro, with loud fan noise noted. Those figures come through second-hand roundup reporting, not from reading the original reviews directly.
Third, cost depends on usage, model, and period. The reason to buy this laptop is not speed — it is keeping your data off outside servers, and using it without an internet connection. Fourth, Korean pricing and release timing are not yet confirmed.
NVIDIA sells this not as an "AI PC" but as a personal AI computer you carry. The $2,599 starting price is only the price of the smallest configuration; the real price is set the moment you pick the memory your purpose needs. Still, the fact itself stands: 74B models have started running on a laptop.
Sources
- NVIDIA official press release (Oct 16, 2026 — on sale, six makers, nine models, from $2,599, up to 128GB, Nemotron 3.5 Bolt 74B-A9B 24GB)
- Apple Newsroom (Aug 25, 2026 — Mac Studio with M5 Max and M5 Ultra, up to 128GB, $2,499)
- NVIDIA official product page (both chip tiers, TDP, memory) · NVIDIA blog (Oct 13, 2026)
- Microsoft Store Surface Laptop Ultra configurations and pricing (including the sparsity footnote for the 1-petaflop figure)
- Independent measurements: Windows Central · Digit · NAG
- Battery figures arrive via second-hand roundup reporting — not read in the original reviews: teckhustlers roundup
- Comparison details: MacRumors · AppleInsider (128GB upgrade price) · Presenc (Strix Halo survey)
Not stated as fact in this article: Korean pricing and timing; the claim that 200B-parameter models can run (hearsay only — Microsoft's official wording goes up to 120B); an official source for the 300GB/s bandwidth figure; 8-bit figures for Qwen 27B; the original text of the battery reviews. All sources were checked on Oct 17, 2026.
Puzz's one-line summary
The substance of RTX Spark is CUDA and up to 128GB of unified memory in a laptop. Judge it by the price of the memory configuration your purpose needs — not by the starting price. Ribbit.