NVIDIA announced on October 2nd the 64GB integrated memory model of its desktop AI system, the "DGX Spark." It will be released in the US on October 23rd through six partners: Acer, ASUS, Dell, Gigabyte, HP, and MSI, with prices starting from $4999 (approximately ¥790,000).

The DGX Spark is a compact computer designed to run large-scale AI models locally, without relying on cloud infrastructure. The new 64GB version features the same "GB10 Grace Blackwell" superchip as the conventional 128GB version, offering up to 1 PFLOP (FP4) of AI compute performance and 273GB/s of memory bandwidth. This configuration maintains core processing performance while halving the memory capacity.

According to NVIDIA, the 64GB version can execute AI models with up to 100 billion parameters locally. While this doesn't match the 128GB version's maximum of 200 billion parameters, it can be utilized for developing AI agents using relatively smaller language models or for AI processing of confidential data without sending it externally.

The pricing, however, is a point of concern. The 64GB version's $4999 is $1000 higher than the 128GB version's launch price of $3999. NVIDIA had previously raised the MSRP of the 128GB "Founders Edition" to $4699 in February 2026, citing a global memory shortage. This new 64GB version, however, exceeds even that revised price by $300.

The 128GB version continued to see further price increases, reaching $6950 (approximately ¥1,100,000) in October. While the 64GB version is a $1951 cheaper option when compared to current prices, it's a rather complex feeling to consider that the 128GB version was available for $3999 just a year ago. This scenario, where a model with half the memory capacity costs more than its former higher-end counterpart, underscores the impact of rising memory prices driven by the surging demand for AI.

For users who find 64GB of memory insufficient, NVIDIA also proposes a solution: connecting two DGX Spark units. By linking two units using the built-in ConnectX-7 network, users can access a total of 128GB of memory, supporting models with up to 200 billion parameters.

At approximately ¥790,000, the price point is steep, even when compared to high-performance consumer PCs. While this expands options for developers serious about local AI, it remains a luxury machine for hobbyists simply looking to experiment.

Made with AI