Receive DGX Spark as an operational AI platform? WZ-IT configures the hardware, Open WebUI, local model runtime, and an agreed model, and can take responsibility for ongoing operations. Explore the AI Cube
NVIDIA DGX Spark is the reference system for the GB10 platform. It combines a 20-core Arm processor, Blackwell GPU, and 128 GB of coherent unified memory in a compact computer. This accommodates local model classes that are impractical on ordinary workstation GPUs. A production AI service still requires a suitable software, security, and operating architecture.
Hardware and intended use
NVIDIA specifies up to 1 PFLOP FP4, a 4 TB self-encrypting NVMe SSD, 10 GbE, and ConnectX-7. DGX OS and NVIDIA's AI stack are preinstalled. Its ARM64 architecture must be considered for containers, Python dependencies, and custom extensions.
Since October 2026 NVIDIA offers DGX Spark in two memory configurations: the Founders Edition stays at 128 GB, while a 64 GB variant is available exclusively through OEM partners such as Acer, ASUS, Dell, Gigabyte, HP, and MSI. According to NVIDIA, the 64 GB systems start at USD 4,999, ship from 23 October 2026, and target models up to about 100 billion parameters (NVIDIA blog). For fine-tuning, NVIDIA quotes models up to 70 billion parameters on a 128 GB system.
The phrase “up to 200 billion parameters” is not a performance guarantee. A model may fit in memory yet remain too slow with long context or several concurrent users. Selection therefore needs measurements with representative input.
From developer system to business platform
A production setup includes:
- documenting and updating firmware, DGX OS, drivers, and container runtime;
- named administrators, SSH keys, and least privilege;
- defined network, DNS, TLS, firewall, user, and administration paths;
- a model runtime and model selected for language, quality, and throughput;
- Open WebUI with users, groups, model permissions, and persistent storage;
- backup coverage for models, chat data, knowledge spaces, and configuration;
- monitoring of the system, services, storage, and model endpoint;
- documented updates, restore procedures, and support access.
What staff use afterwards
Users work through a familiar browser interface. Depending on permissions, they can choose approved local models, work with files, and maintain personal or shared knowledge spaces. Approved cloud models can be added without replacing the local platform.
Automated connections to Nextcloud, SharePoint, a DMS, or business applications are separate integration projects, as are custom RAG pipelines, APIs, and MCP servers.
When DGX Spark is the right choice
The reference system is a good fit when an organisation explicitly wants NVIDIA's DGX platform and a clearly documented vendor path. It may be excessive for one workstation without central users or operating requirements. One device may be insufficient for high concurrency, redundant service, or very large models.
Application, data volume, expected concurrency, and support model should therefore be defined before procurement. This determines whether one DGX Spark as an AI Cube is sufficient, a second node is useful, or a larger GPU platform is the better starting point.
AI Cube in the DGX Spark class
The AI Cube is WZ-IT's complete offer in this hardware class. It includes hardware, a preconfigured AI platform, agreed local model, hardening, functional testing, initial setup, and support. It is generally ready within two weeks after configuration approval.
A second AI Cube can serve more users or larger models. NVIDIA quotes models up to 700 billion parameters for four clustered 128 GB DGX Spark systems; such a cluster requires a switch and does not apply to every OEM variant. Connecting AI Cubes with ConnectX-7 explains request balancing and distributed inference.
