Nvidia launches a 64GB version of the DGX Spark, with memory capacity halved, starting at $4,999; the 128GB version is priced at $6,950.

As AI models become increasingly lightweight, video memory (VRAM) is becoming increasingly expensive. Nvidia is addressing this contradiction with a product that "reduces memory usage."

On Friday, October 2nd, Eastern Time, NVIDIA announced a 64GB unified memory version of its desktop AI computer, the DGX Spark, starting at $4,999. It will be available on October 23rd from OEMs including Acer, ASUS, Dell, Gigabyte, HP, and MSI. The new machine retains the GB10 Grace Blackwell super chip, DGX OS, the complete NVIDIA AI software stack, and ConnectX-7 networking capabilities, but the unified memory has been reduced from the existing 128GB to 64GB.

Intriguingly, the starting price of this "watered-down" DGX Spark is actually higher than the initial price of $3,999 for the 128GB version last year; meanwhile, Nvidia has further increased the price of the 128GB Founders Edition to $6,950. Media reports indicate that Nvidia attributes this price increase to limited memory supply and rising costs.

This means that, with the continued pressure on memory supply from the demand for AI servers, "reducing memory" has become a way to lower the price threshold for local AI devices, but it does not mean that the devices themselves have actually become cheaper.

64GB is enough? Nvidia targets local AI agents.

The core logic behind Nvidia's adjustment is that more and more open-source models are now able to run in smaller memory capacities.

NVIDIA stated that with the improvement of open-source model capabilities and the continuous reduction in model size, 64GB of unified memory is sufficient to support a range of local AI applications. The new DGX Spark can run models with up to 100 billion parameters locally on the device and supports workloads such as AI agents, inference, fine-tuning, data science, and edge development.

Tom's Hardware also points out that some of the latest high-performance, intensive models can already run on around 32GB of memory, although they are still limited in scenarios with larger context windows. This means that the 128GB of memory originally intended for "running large models locally" is not necessary for all developers.

This is also the practical basis for NVIDIA's launch of the 64GB version: for users who mainly perform local inference and develop AI agents, rather than large-scale model training or fine-tuning, 64GB can cover a considerable portion of their workload.

In its announcement, Nvidia even summarized this trend as "the practicality of local AI is increasing as technology continues to evolve," and as AI agents move from experimentation to routine development, the need for locally running models is increasing.

At the same time, on-premises deployment also offers advantages in terms of privacy and cost. Developers can process their own data and run AI agents directly on DGX Spark without having to call cloud models for every task.

The same GB10, 64GB version has no performance reduction.

From a hardware architecture perspective, the 64GB version is not a completely new chip product.

Nvidia stated that the new machine still uses the GB10 Grace Blackwell Superchip and retains the complete DGX OS and AI software stack. Media reports indicate that the 64GB version also retains the original 20-core Arm CPU and 273GB/s shared memory bandwidth. Therefore, for models capable of holding 64GB of RAM, the basic computing power of both products has not changed despite the halved memory capacity.

In other words, the focus of this change is not "cutting computing power", but rather cutting some memory capacity that users may not need .

The 64GB version still supports mainstream inference frameworks such as llama.cpp, Ollama, vLLM, and LM Studio, and comes pre-installed with NVIDIA Agent Toolkit, CUDA-X AI library, and open models such as Nemotron.

Nvidia's strategy is clear: if developers currently only need 64GB, they should purchase a machine with a lower capacity; if the model continues to expand in the future, they can then increase memory and computing power through clustering.

Two 64GB units can be combined to create a 128GB system, resulting in a performance increase of up to 70%.

Another key aspect of this DGX Spark update is NVIDIA's further enhancement of multi-machine collaboration.

The 64GB version also features a built-in ConnectX-7 network interface, allowing two devices to be directly connected via a QSFP cable. NVIDIA Sync Cluster Assistant automatically configures the network, creating a local AI cluster. NVIDIA claims that two 64GB DGX Sparks can form a 128GB memory pool, supporting models with up to 200 billion parameters.

In NVIDIA's Qwen 3.8 27B test, the performance of a cluster consisting of two 64GB machines can reach up to 1.7 times that of a single machine.

NVIDIA will also launch the NVIDIA Sync Model Launcher at the end of October to further simplify the deployment process of models on single machines or in clusters. For example, developers can use this tool to run Qwen 3.8 27B and connect it to programming tools such as OpenCode.

This means that DGX Spark's product logic is shifting from "a desktop AI supercomputer" to "a local AI node that can be scaled up gradually".

Of course, two machines do not equate to a single physical 128GB DGX Spark; whether a specific model can run across nodes still depends on the software and workload. Therefore, for tasks requiring large amounts of memory to run on a single machine, the 64GB version still presents significant limitations.

Given the shortage of video memory, the 64GB version at $4999 isn't really "cheap."

What really deserves attention is the price.

NVIDIA's official starting price for the 64GB DGX Spark is $4,999. This version will not be offered in an NVIDIA Founders Edition; instead, it will be sold entirely through OEM partners . These partners include Acer, ASUS, Dell, Gigabyte, HP, and MSI. Specific configurations and prices may vary.

In comparison, when the DGX Spark was first launched in 2025, the official price of the 128GB version was $3,999; media reports indicate that Nvidia subsequently raised the price of the 128GB Founders Edition to $4,699 in February 2026, and now it has been further adjusted to $6,950.

Therefore, if we only compare the historical initial prices, today's 64GB version is not only not cheaper than the 128GB version, but is actually $1,000 more expensive, an increase of 25%.

However, compared to Nvidia's current official price of 128GB, the 64GB version at $4,999 is still $1,951 cheaper than $6,950, but at the cost of halving the memory capacity.

PC Watch, citing relevant information, stated that the price increase for the 128GB version is related to memory supply constraints and rising costs; Tom's Hardware pointed out that the actual selling price of a 128GB GB10 system on the market has already reached approximately $7,000 to $9,000.

Therefore, another implication of this product adjustment is that AI computing power is shifting towards local devices, but memory has become a significant cost bottleneck in this trend.

From "heap memory" to "on-demand expansion"

From a product strategy perspective, Nvidia is not simply making a "low-spec version" of DGX Spark this time.

In the past, one of the core selling points of DGX Spark was its 128GB unified memory, which allowed developers to run large parameter models on desktop devices; now, with the advancement of model compression and quantization technologies, some models no longer require such a large memory capacity.

Therefore, NVIDIA's solution is as follows: 64GB single machine can meet the needs of mainstream local AI inference, and when memory requirements increase further, it can be expanded through multi-machine clusters.

The 64GB version will be officially launched on October 23, with a starting price of $4,999. At the same time, the 128GB version will be priced at $6,950, making this new product particularly special – it is both an attempt by NVIDIA to lower the barrier to entry for local AI hardware and a reflection of the current contradiction in the AI industry of "strong demand for computing power and tight memory supply".

Risk warning and disclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.