As AI models become increasingly lightweight, video memory keeps getting more expensive. NVIDIA is responding to this contradiction with a product that cuts memory.

On Friday, October 2, U.S. Eastern Time, NVIDIA (NASDAQ: NVDA) announced the launch of a 64GB unified memory version of its desktop AI computer DGX Spark, starting at $4,999, which will be brought to market on October 23 by OEM manufacturers including Acer, ASUS, Dell, Gigabyte, HP, and MSI. The new machine retains the GB10 Grace Blackwell Superchip, DGX OS, the full NVIDIA AI software stack, and ConnectX-7 networking capability, but its unified memory has been reduced from the existing 128GB to 64GB.

Interestingly, the starting price of this stripped-down DGX Spark is actually higher than last year's 128GB version's launch price of $3,999. At the same time, NVIDIA has further raised the price of the 128GB Founders Edition to $6,950. Media noted that NVIDIA attributed this price increase to constrained memory supply and rising costs.

This means that against the backdrop of AI server demand continuing to squeeze memory supply, providing less memory has become a way to lower the price barrier for local AI devices, but it does not mean the device itself has actually become cheaper.

Is 64GB enough? NVIDIA targets local AI agents

The core logic behind NVIDIA's adjustment is that a growing number of open-source models can already run within smaller memory capacities.

NVIDIA stated that with the improving capabilities of open-source models and shrinking model sizes, 64GB of unified memory is already 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 including AI agents, inference, fine-tuning, data science, and edge development.

Tom's Hardware also noted that some of the latest high-performance dense models can already run in around 32GB of memory, though they may still be limited in scenarios involving 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 mainly performing local inference and developing AI agents rather than large-scale model training or fine-tuning, 64GB can cover a considerable portion of workloads.

In its announcement, NVIDIA even summarized this trend as local AI becoming increasingly practical as technology continues to evolve, with demand for running models locally growing as AI agents move from experimentation to everyday development.

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

Same GB10, the 64GB version's performance has not been reduced

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

NVIDIA stated that the new machine still uses the GB10 Grace Blackwell Superchip and retains the full DGX OS and AI software stack. Media noted that the 64GB version also retains the original 20-core Arm CPU and 273GB/s shared memory bandwidth, so for models that can fit within 64GB of memory, the basic computing capability of the two products has not changed due to the halving of memory capacity.

In other words, the focus of this change is not cutting computing power, but cutting 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, and LM Studio, and comes pre-installed with NVIDIA Agent Toolkit, CUDA-X AI libraries, and open models such as Nemotron.

NVIDIA's calculation is also clear: if developers currently only need 64GB, they can first buy a lower-capacity machine; if model sizes continue to expand in the future, they can then add memory and computing power through clustering.

Two 64GB units can be combined into 128GB, with performance improved by up to about 70%

Another key point of the DGX Spark update is that NVIDIA has further strengthened multi-machine collaboration.

The 64GB version also has a built-in ConnectX-7 network interface. Two devices can be directly connected via QSFP cables, and NVIDIA Sync Cluster Assistant automatically completes network configuration to form a local AI cluster from the two machines. NVIDIA says 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, after two 64GB machines formed a cluster, performance could reach up to about 1.7 times that of a single machine.

NVIDIA will also launch NVIDIA Sync Model Launcher at the end of October to further simplify the deployment process of models on a single machine or cluster. 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 the product logic of DGX Spark is shifting from a desktop AI supercomputer to a local AI node that can be gradually expanded.

Of course, two machines do not equal a physically single 128GB DGX Spark, and whether a specific model can run across nodes still depends on software and workloads. Therefore, for tasks requiring large-memory single-machine operation, the 64GB version still has clear limitations.

Amid the memory shortage, the 64GB version at $4,999 is not truly cheap

What is truly noteworthy is actually the price.

NVIDIA's official starting price for the 64GB DGX Spark is $4,999, and this version will not have NVIDIA's own Founders Edition, but will be sold entirely through OEM partners. Partners include Acer, ASUS, Dell, Gigabyte, HP, and MSI, and specific configurations and prices may vary.

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

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

If compared with the current official NVIDIA 128GB price, the $4,999 64GB version is still $1,951 cheaper than $6,950, but the cost is that memory capacity is directly halved.

PC Watch cited related information saying that this price increase for the 128GB version is related to memory supply constraints and rising costs. Tom's Hardware noted that the actual market price of 128GB GB10 systems has even reached about $7,000 to $9,000.

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

From stacking memory to scaling on demand

From a product strategy perspective, NVIDIA is not simply making DGX Spark into a low-spec version.

In the past, one of the core selling points of DGX Spark was 128GB of unified memory, allowing developers to run large-parameter models on desktop devices. Now, with advances in model compression and quantization technology, some models no longer require such large memory capacity.

Therefore, NVIDIA's solution has become: a single 64GB machine to satisfy mainstream local AI inference, and when memory demand increases further, expand through multi-machine clustering.

The 64GB version will officially go on sale on October 23, starting at $4,999. At the same time, the 128GB version's price increase to $6,950 also makes this new product look particularly unusual. It is both NVIDIA's attempt to lower the barrier to local AI hardware and a mirror of the current contradiction in the AI industry: strong demand for computing power and tightening memory supply.