# NVIDIA RTX Spark arrives on desktop with 128 GB for local AI

> The first compact machines with the N1X superchip go on sale with up to 128 GB of unified memory and 1 PFLOP in FP4. Which models fit and what is missing.
> By José Maldonado · Technology · 2026-11-19

The first compact computers with RTX Spark reached stores on November 18, 2026, with systems from Acer, Asus, Dell, Gigabyte, HP, Lenovo and MSI. [Nvidia puts the starting price at 4,399 dollars](https://uk.pcmag.com/ai/168443/nvidia-and-microsoft-bring-rtx-spark-to-compact-desktops) for the 64 GB configuration, so that figure should be read as the floor of the range. What changes the category is memory: up to 128 GB unified under the desk.

That memory decides what can be run. A large local model is ruled out by capacity, not by computing power. A laptop with a high-end dedicated graphics card carries between 16 and 24 GB of video memory, and a 70 billion parameter model quantized to 4 bits takes up around 40 GB: it does not fit. The N1X superchip shares a single space between CPU and GPU, and the problem stops being the split and becomes the size of the model.

At Ignite, Microsoft positioned these machines as the entry step in its hybrid strategy and reserved the DGX Station for Windows for trillion-parameter models. Between the two there is a price gap, and that is where the buying decision lies.

## What the N1X superchip includes

The RTX Spark N1X combines a Blackwell GPU with 6,144 RTX cores and a 20-core Grace CPU, designed by Nvidia and manufactured by MediaTek, the same alliance the laptops already used weeks earlier. [Nvidia sums up the RTX Spark N1X in three figures](https://blogs.nvidia.com/blog/rtx-spark-desktop-local-ai-available-now/): 1 PFLOP in FP4, up to 128 GB unified and 120 billion parameter models running locally.

Of the three figures, the one that shapes the purchase is memory. The petaflop in FP4 looks good on the spec sheet, but it does not say how many tokens per second you will see: the low-precision format only describes how the calculations are done. Power consumption does not appear in the sales pitch either, and Nvidia does not detail it with the same insistence as capacity.

The package is the same as the laptops, and the jump to desktop serves to standardize the high configuration: 20 CPU cores, 6,144 GPU cores and 64 or 128 GB of memory, with no cut-down versions.

## Which models really fit

The calculation uses two figures: parameters and bits per parameter. A 70 billion model at 4 bits comes close to 40 GB and leaves room for a long context window on a 128 GB machine. A 120 billion one at the same precision is around 70 GB and still fits, with less headroom for the conversation state. If the model does not fit, no configuration saves it.

The trillion-parameter ones, the ones Microsoft mentions for the DGX Station, do not fit here. The DGX Station marks the ceiling of the range and does not promise that a desktop can run a server model.

The other half of the calculation is bandwidth. Unified memory solves capacity, not speed, and in local inference the bottleneck is usually how many tokens per second the GPU can process. Nvidia has not published performance per watt for these machines, and without that figure any promise of speed is an estimate.

## Software is half the product

[GitHub Copilot Fusion already routes to local models](https://blogs.windows.com/windowsexperience/2026/11/17/microsoft-ignite-2026-building-windows-for-the-next-era-of-work/) on machines with RTX Spark. The feature is in preview and available in the Copilot app, on the command line and in Visual Studio Code. The commercial logic is simple: if the model runs on the machine, the cost per token disappears and the bill moves to the hardware.

If you manage a fleet with Windows 11, check one detail before trying the preview: [version 26H2 has accumulated bugs a week after its rollout](https://observatorioblockchain.com/en/technology/windows-11-26h2-fallos-una-semana/), and a workstation that updates itself is not a good test bench.

The missing piece is the ecosystem. Microsoft's strategy relies on its MAI models locally and on the usual formats, and the machine works if a developer runs what they already use without rewriting anything. Nvidia is pushing with [the Nemotron for Science package open to scientific agents](https://observatorioblockchain.com/en/ai/nemotron-for-science-agentes-cientificos/). [Foundry Local documentation explains how to load a model](https://learn.microsoft.com/en-us/windows/ai/foundry-local/get-started) on the machine itself without special permissions, which is the boring work that decides whether the chip gets used.

## What it costs and what it compares with

Four thousand four hundred dollars is the starting point, not the final price: the 64 GB configuration stays at that figure, the amount varies somewhat by manufacturer and the 128 GB ones go up from there.

The honest comparison is between categories: the compact, the multi-card workstation and the cloud subscription. The traditional workstation wins on raw power and loses on consumption, space and noise. The cloud wins on flexibility and loses on privacy, and its bill is read on the spot: [Gemini 4 Hydrogen charges 1.50 dollars per million input tokens and 7.50 per million output](https://observatorioblockchain.com/en/ai/gemini-4-hydrogen/), with flat pricing across the entire context window. The local machine wins on latency and data control, and loses on everything a frontier model does better.

There is a factor that weighs more than it seems. [The memory shortage that SK hynix places through 2030](https://observatorioblockchain.com/en/technology/escasez-memoria-2030-sk-hynix-hardware/) makes more expensive precisely the component that gives the machine its meaning, and whoever buys today pays the premium of a tight market. Whoever waits may find the same capacity for less, or may not find it.

## Who it serves today

The profile that justifies the expense is not the one who wants a fast computer. It is the developer who tests models without uploading data outside, the small team that runs repetitive inference over its own documents and prefers a fixed cost to a monthly bill, and whoever works with data that cannot leave the office. For that case, local memory is the only answer.

The home user who just wants a machine for gaming and browsing does not need it: for that there are cheaper options and unified memory adds nothing if the software does not take advantage of it.

The question that decides the purchase is not how many cores the chip has: it is which model will be run on day one and whether it fits. With the answer in front of you, 128 GB is either a useful figure or a brochure number, and the distance between the two is measured in thousands of euros. A 4,399 dollar machine with 64 GB leaves little room to experiment.

## One more step in the shift of computing

The arrival of these machines closes a year in which the discourse moved from training to inference. If the model runs on the desktop, the cloud stops being the center of the bill and becomes the backup for the hard stuff.

It remains to be seen whether the software keeps up. A chip with 128 GB is a necessary condition and not a sufficient one, and the history of local computing is full of capable hardware that nobody knew how to take advantage of. This time the operating system maker and the chip maker are pushing in the same direction, and November 18, 2026 is the first date on which both products are sold together. What is new is that coincidence, not the petaflop.
