Nvidia announces DGX Spark 64GB Grace Blackwell desktop for local AI at $4,999

Oct 2, 2026

Nvidia announces a new 64GB configuration of its GB10-powered DGX Spark desktop AI system, retailing at $4,999 to offer a more affordable option for local AI development. The launch comes as skyrocketing memory prices force Nvidia to raise the price of the 128GB model to $6,950. The new system retains the Grace Blackwell superchip, delivering up to 1 petaFLOP of FP4 AI compute, and is designed to run 30B-class open models locally, bypassing metered cloud API fees.

Launch and availability of the DGX Spark 64GB

▪The DGX Spark 64GB configuration will be offered exclusively through hardware partners including Acer, Asus, Dell, Gigabyte, HP, and MSI, with availability starting October 23, 2026.
▪Nvidia announced a 64GB configuration of its GB10-powered DGX Spark desktop AI system on October 2, 2026, retailing for approximately $4,999.

Hardware architecture of the GB10 platform

▪The GB10 platform utilizes unified memory shared over an NVLink-C2C connection, providing five times the bandwidth of PCIe Gen 5 to eliminate data copying between system RAM and VRAM.
▪The GB10 platform pairs a Blackwell GPU featuring fifth-generation Tensor Cores that deliver up to 1 petaFLOP of FP4 AI compute with a 20-core MediaTek Grace Arm processor.

Memory bandwidth performance

▪The DGX Spark 64GB maintains a memory bandwidth of 273 GB/s by utilizing lower-capacity LPDDR5x memory modules rather than reducing the total number of modules.
▪The 273 GB/s memory bandwidth of the DGX Spark 64GB limits its concurrent decode throughput, making it unsuitable as a chat server for 100 users.

Local model execution and fine-tuning capabilities

▪While a full BF16 build of Meta's Muse Glimmer model requires over 55GB of memory, a quantized build of approximately 17GB is recommended for practical local execution on the DGX Spark 64GB.
▪The DGX Spark 64GB can accommodate QLoRA fine-tuning of a 70-billion parameter model, achieving nanochat distributed fine-tuning speeds of approximately 18,400 tokens per second.
▪The DGX Spark 64GB is designed to run 30-billion to 35-billion parameter open AI models locally, such as Qwen 3.8 27B and Meta's distilled Muse Glimmer model.

Advantages of running local AI agents

▪The DGX Spark 64GB runs on a standard wall outlet without requiring specialized cooling or server room infrastructure, facilitating continuous, always-on local agent operations.
▪Running AI agents on local hardware like the DGX Spark avoids metered cloud API fees, which is increasingly significant as token consumption has grown 14-fold since early 2026.

Clustering capabilities

▪Nvidia will roll out new software tools to DGX OS in October 2026 to facilitate cluster configurations and inference deployments.
▪Two DGX Spark 64GB units can be clustered using built-in ConnectX-7 networking and NVIDIA Sync to deliver up to 1.7 times the performance of a single 128GB DGX Spark.

Price increases for the GB10 platform

▪The GB10 platform has experienced pricing creep since its unveiling as the Projects Digits concept, rising from an expected $3,000 to a $4,000 launch price, and now to $6,950 for the 128GB model.
▪Nvidia increased the price of the 128GB DGX Spark to $6,950 on October 2, 2026, representing an increase of nearly 75 percent from its price this time last year due to rising memory costs.

RTX Spark pricing

▪Nvidia plans to offer versions of the GB10 chip with lower memory capacities for price-conscious buyers of RTX Spark systems.
▪The ongoing memory shortage is expected to drive up the retail prices of Nvidia's upcoming RTX Spark notebooks and mini PCs scheduled for release in fall 2026.

Debatable claims

▪A 64GB memory configuration is insufficient for local AI development
▪A primary use case for desk-side agentic AI is in software development to reduce high monthly cloud spending on AI tools
▪Nvidia's superior AI performance justifies its price premium over AMD
▪Nvidia's DGX Spark price hikes are justified by rising memory costs