NVIDIA RTX Spark PCs Open for Preorder — a 128GB Dev Box Costs $5,999

A 128GB local-AI PC is now something you can buy, but the price is about $6,000, more than double an analyst’s estimate of around $2,899.

Products built on NVIDIA’s PC chip, RTX Spark, have gone up for preorder. Microsoft opened preorders for new Surface devices, and the 128GB unified-memory developer desktop, the “Surface RTX Spark Dev Box,” costs $5,999 in the U.S. For developers who want to run big AI models, such as the large language models (LLMs) behind chatbots like ChatGPT, on their own PC without the cloud, that is a new option, but not at the price people were hoping for.

What kind of chip is RTX Spark

RTX Spark is a chip for Windows PCs based on Arm (the chip design used in most phones and in Apple’s Macs), unveiled this year at Computex, a major computer trade show in Taiwan. NVIDIA designed it together with MediaTek and worked with Microsoft on Windows support. The reported specs:

GPU: 6,144 CUDA cores (the calculation units of NVIDIA GPUs) and 5th-generation Tensor cores (units built for AI math)
CPU: Grace CPU (NVIDIA’s Arm-based CPU), up to 20 cores
AI compute: up to 1 petaflop, or a quadrillion calculations per second. This is a theoretical maximum, based on FP4 (computing with numbers shrunk to 4 bits) with sparsity applied, a method that skips unneeded calculations.
Memory: up to 128GB of LPDDR5X (a fast, low-power memory standard) unified memory (CPU and GPU share one pool)

Specs follow Microsoft’s and NVIDIA’s official announcements. The memory bandwidth (speed) is not listed on NVIDIA’s product page either.

What you can buy now, and the prices

Product	Memory	Price (U.S.)	Ships
Surface Laptop Ultra	up to 128GB	from $2,599; top 128GB / 1TB configuration $5,899.99	from Oct. 16
Surface RTX Spark Dev Box	128GB + 2TB SSD	$5,999	November (Microsoft.com only)
ASUS ProArt P16	up to 128GB	U.S. price of the 128GB configuration not found in official material (64GB configuration listed at $4,399.99 on ASUS’s U.S. store)	Oct. 16

All prices are preorder prices. ASUS itself only states a U.S. starting price of $2,599.99.
The Dev Box costs $2,000 more than AMD’s Ryzen AI Halo ($3,999). Among 128GB-class machines, it is also higher than NVIDIA’s DGX Spark ($4,699).

Claims we could not confirm

“Fine-tune 100B+ models locally”: Microsoft’s official post says the devices support AI models exceeding 120 billion parameters. The phrase “fine-tune” (further training an existing model on your own data for a specific job) does not appear in it, though. On paper, a 100-billion-parameter model (parameters are the values a model learns) squeezed to 4 bits is about 50GB, so the weights fit in 128GB. But “fits” only means fits; real speed and the fine-tuning method (such as LoRA, which changes only part of a model) need to be verified per product.
“$2,899-class”: that number is a Morgan Stanley analyst’s pre-launch estimate for the 128GB configuration. The actual preorder price is $5,999.

What could change for developers

What follows is our estimate and advice based on the reports.

CUDA-based local development on Windows may become possible. But how well PyTorch (a popular AI development library) and local AI programs such as Ollama, llama.cpp and vLLM run on Windows on Arm (Windows running on Arm chips) will only be clear from post-launch reviews.
Check speed (bandwidth), not just capacity. Even at 128GB, the speed at which text appears depends a lot on bandwidth. This product’s bandwidth is not yet confirmed.
Judge it on price for what you get. Among 128GB-class machines, compare against DGX Spark ($4,699) and Ryzen AI Halo ($3,999), and ask whether you really need Windows.
It’s still only a preorder. The Dev Box is U.S.-only and ships in November, so real performance reviews will come about a month from now.

What to watch

RTX Spark’s memory bandwidth and local-LLM benchmarks, CUDA and PyTorch compatibility on Windows on Arm, and prices from other makers (Dell, HP, Lenovo, MSI).