DGX Spark Explained

The starting price of the 64 GB NVIDIA DGX Spark is $4,999 from participating OEM partners beginning 23 October 2026, while NVIDIA publishes no price for the 128 GB model. At Layer3Labs, we build and operate AI systems inside other people's businesses. The DGX Spark is a desktop AI computer built on the NVIDIA GB10 Grace Blackwell Superchip.

NVIDIA calls the DGX Spark a complete platform for local autonomous agents and a Desktop Agent Computer. NVIDIA says local memory and compute reduce the need for cloud token generation. The DGX Spark delivers up to 1 petaFLOP of AI performance at FP4 precision.

What a DGX Spark Costs and What It Is For

The 64 GB DGX Spark starts at $4,999 from participating OEM partners beginning 23 October 2026. Participating partners include Acer, ASUS, Dell, Gigabyte, HP, and MSI. NVIDIA publishes no price for the 128 GB model.

On the official NVIDIA product page, the 64 GB configuration is marked as coming soon with a notify-me form. Pricing for the 128 GB model is unpublished. Prospective buyers should check with the OEM partners listed. Our Perplexity Portable Computer pricing guide recorded a $4,699 listing for the DGX Spark on NVIDIA's marketplace as of 27 August 2026, which is a dated snapshot and not a current price.

Workloads listed by NVIDIA include agentic AI, prototyping, fine-tuning, inference, data science, and edge applications. Edge applications listed on the product page include Isaac, Metropolis, and Holoscan. These represent NVIDIA's listed workloads, not our test results.

Hardware Specifications of DGX Spark

The DGX Spark runs on the NVIDIA GB10 Grace Blackwell Superchip with a 20-core Arm CPU. That processor combines 10 Cortex-X925 cores and 10 Cortex-A725 cores. Memory comes as coherent unified system memory in 64 GB or 128 GB configurations.

The 64 GB configuration uses LPDDR5X memory, while the 128 GB configuration uses LPDDR5x. NVIDIA lists a 256-bit interface and 273 GB/s bandwidth. Storage options extend up to 4 TB NVMe M.2 drives featuring self-encryption.

Networking hardware includes a 10 GbE RJ-45 port, a ConnectX-7 NIC at 200 Gbps, Wi-Fi 7, and Bluetooth 5.4. Physical connectivity provides four USB-C ports alongside an HDMI 2.1a video output. The power supply rating is 240 watts.

The GB10 TDP is 140 W. A Reddit post titled "From 1x3090 to 20 DGX Sparks: my house fuses were the first bottleneck" exists. We did not verify it. The chassis measures 150 mm by 150 mm by 50.5 mm and weighs 1.2 kg. Declared sound power is 35 dB under maximum GPU stress at 25 C and drops to 19 dB at idle.

64 GB and 128 GB Configurations

The 64 GB configuration of the DGX Spark is available exclusively through participating OEM partners starting 23 October 2026. Those OEM partners are Acer, ASUS, Dell, Gigabyte, HP, and MSI. The 128 GB model shares the same GB10 superchip, DGX OS, and software stack.

The 64 GB configuration supports models up to 100 billion parameters on device. The 128 GB configuration supports inference up to 200 billion parameters.

It also supports fine-tuning up to 70 billion parameters.

Clusters of Up to Four DGX Spark Units

ConnectX networking allows users to link up to four DGX Spark systems. Two units connect directly with a QSFP cable. The NVIDIA Sync Cluster Assistant detects connected units, validates device configuration, and configures the ConnectX-7 network.

NVIDIA says every node runs the same software stack. In NVIDIA's Qwen 3.8 27B test, two clustered 64 GB systems delivered up to 1.7x the performance of a single system. NVIDIA says two clustered units provide twice the memory bandwidth.

Two 64 GB units pool memory to 128 GB to support models up to 200 billion parameters. Clustering two 128 GB units pools memory to 256 GB, supporting models up to 400 billion parameters. A cluster of four 128 GB units reaches 512 GB pooled memory, supporting models up to 700 billion parameters.

For more on networking, read our guide to clustering DGX Spark. A Reddit post titled "For dual DGX spark users; GLM 5.3 flash got a 50%+ performance boost" exists. We did not verify it.

Listed Workloads and Software Playbooks

The preinstalled software includes the NVIDIA AI software stack, DGX OS, NVIDIA Agent Toolkit, and CUDA-X AI libraries. NVIDIA also provides preinstalled Nemotron open models. Supported runtimes include Ollama, vLLM, and PyTorch with CUDA.

In the NVIDIA blog post dated 2 October 2026, the get-started steps also name llama.cpp and LM Studio. NVIDIA lists playbooks on build.nvidia.com/spark, including NemoClaw, OpenClaw, Hermes Agent, and OpenShell. Additional playbooks coming soon, per NVIDIA, include Serve LLMs With vLLM, Run OpenClaw With a Local LLM, and Connect Multiple DGX Sparks for Distributed Workloads.

The NVIDIA Sync Model Launcher, coming at the end of October 2026, downloads and launches Qwen3.8 27B on one DGX Spark or a cluster. NVIDIA says the model launcher makes the model accessible from laptops and sets up OpenCode to use it.

NVIDIA highlights running an AI agent around the clock and powering AI applications on an everyday PC while the DGX Spark handles inference. These use cases represent NVIDIA's listed workloads, not our test results.

Cases Where DGX Spark Is the Wrong Answer

Small operations without technical staff are a poor fit. In our opinion, local hardware fits when query volume is high or a policy rules out sending documents to a cloud API. Our answer would change if hosted prices dropped sharply.

Our local AI hardware calculator estimates the GPU, VRAM and RAM a model needs for a given model size and use case. Layer3Labs runs a free AI workflow audit.