While the AI model is becoming increasingly “lightweight,” the cost of video memory (VRAM) is rising. NVIDIA is addressing this contradiction with a product that aims to reduce the amount of memory required.

On Friday, October 2nd, Eastern Time, NVIDIA announced the launch of a desktop computer AI with 64GB of unified memory, priced at $4,999. It will be available in the market on October 23rd through manufacturers such as Acer, ASUS, Dell, Gigabyte, HP, and MSI. The new model retains the GB10 Grace Blackwell superchip, DGX OS, a complete NVIDIA AI software stack, and ConnectX-7 networking capabilities, but the amount of unified memory has been reduced from the current 128GB to 64GB.

Interestingly, the starting price of this “reduced-size” DGX Spark is actually higher than the initial price of $3,999 for the 128GB version last year; at the same time, NVIDIA has also raised the price of the 128GB Founders Edition to $6,950. The media points out that NVIDIA attributes this price increase to limited memory supply and rising costs.

This means that, against the backdrop of the AI servers continuously straining the memory supply, "providing less memory" has become a way to lower the price threshold for local AI devices, but it does not necessarily mean that the devices themselves have actually become cheaper.

Is 64GB enough? NVIDIA targets local AI intelligents?

The core logic behind NVIDIA's adjustment this time is that an increasing number of open-source models are now capable of running with smaller memory capacities.

NVIDIA states that with the improvement of open-source model capabilities and the continuous reduction in model size, 64GB of unified memory is now sufficient to support a range of local AI applications. The new DGX Spark can run models with up to 100 billion parameters locally on devices 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, memory-intensive models can now run with around 32GB of memory, although they may still be limited in scenarios with larger context windows. This means that the 128GB of memory originally reserved for "local running of large models" is not necessary for all developers.

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

In its announcement, NVIDIA even summarized this trend as "as technology continues to evolve, the practicality of local AI is increasingly enhancing." As AI intelligents continue to move from experimental stages to daily development, the demand for locally running models is on the rise.

At the same time, local deployment also has attractions in terms of privacy and cost. Developers can directly process their own data and run AI agents on DGX Spark, without having to call cloud models for each task.

The same GB10, the performance of the 64GB version has not been reduced.

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

NVIDIA stated that the new models still use GB10 Grace Blackwell Superchip, and retain the complete DGX OS and AI software stacks. The media noted that the 64GB version also retains the original 20-core Arm CPU as well as a shared memory bandwidth of 273GB/s. Therefore, for models that can accommodate 64GB of memory, the basic computing power of both products has not changed despite the reduction in memory capacity by half.

In other words, the focus of this change is not to reduce computing power, but to cut down on a portion of memory capacity that some users may not need.

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

NVIDIA's strategy is also clear: if developers currently only need 64GB of memory, they should purchase a machine with less capacity first; if the scale of models needs to be expanded in the future, then additional memory and computing power can be added through clustering.

Two 64GB units can be combined to form 128GB, with a performance increase of up to about 70%.

DGX Spark Another highlight of this update is that NVIDIA has further enhanced multi-machine collaboration.

The 64GB version also comes with the ConnectX-7 network interface built-in. Two devices can be directly connected using a QSFP cable, and with the help of NVIDIA Sync Cluster Assistant, the network configuration is automatically completed, forming a local AI cluster between the two machines. NVIDIA claims that two 64GB DGX Spark units can create a 128GB memory pool, supporting models with up to 200 billion parameters.

In the Qwen 3.8 27B test provided by NVIDIA, when two 64GB machines are clustered together, the performance can reach approximately 1.7 times that of a single machine.

NVIDIA will also launch NVIDIA Sync Model Launcher at the end of October, further simplifying 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 integrate it with programming tools such as OpenCode.

This means that the product logic of DGX Spark is transitioning from a “single desktop AI supercomputer” to “locally scalable AI nodes” that can be gradually expanded.

Of course, two machines do not equal one physical 128GB DGX Spark unit. Whether a specific model can run across nodes still depends on the software and the workload. Therefore, for tasks that require large amounts of memory to run on a single machine, the 64GB version still has obvious limitations.

With a shortage of video memory, the 64GB version priced at $4,999 is not really “cheap”

What really deserves attention is actually the price.

NVIDIA's official starting price for the 64GB version is $4,999, and this version will not come with NVIDIA's own Founders Edition; instead, it will be sold entirely through OEM partners. The partners include Acer, ASUS, Dell, Gigabyte, HP, and MSI. The specific configurations and prices may vary.

For comparison, when DGX Spark was first launched in 2025, the official price of the 128GB version was $3,999; media reported that NVIDIA later raised the price of the 128GB Founders Edition to $4,699 in February 2026, and after this further adjustment, it reached $6,950.

Therefore, if we only compare the historical launch prices, today's 64GB version is not only not cheaper than the 128GB version, but actually costs $1,000 more, representing a 25% increase.

However, compared to the current official price of 128GB from NVIDIA, the 64GB version priced at $4999 is still $1951 cheaper than the $6950 version, but the trade-off is that the memory capacity is halved directly.

PC Watch cites relevant information stating that the price increase for the 128GB version this time is related to limited memory supply and rising costs; Tom's Hardware points out that the actual selling price of 128GB GB10 systems on the market has even reached around $7,000 to $9,000.

Therefore, another implication of this product adjustment is that: AI computing power is moving towards local devices, but memory has become an important cost bottleneck in this trend.

Shift from "stack memory" to "on-demand scaling"

From a product strategy perspective, NVIDIA's approach this time is not simply to create a "low-end version" of DGX Spark.

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

Therefore, the solution proposed by NVIDIA became: a single machine with 64GB of memory is sufficient for mainstream AI inference, and when the memory requirement increases further, it can be expanded through a multi-machine cluster.

The 64GB version will be officially launched on October 23rd, with a starting price of $4,999. At the same time, the price of the 128GB version has been raised to $6,950, which makes this new product particularly special – it is not only NVIDIA's attempt to lower the hardware threshold for AI, but also a reflection of the current contradiction in the AI industry, where there is a high demand for computing power and a tightening supply of memory.