NVIDIA DGX Spark — Personal AI Computer (128 GB)

Description

NVIDIA DGX Spark brings the NVIDIA AI software path to a compact personal computer with 128 GB of unified memory. This D-Central listing is for the 4 TB configuration at $9,449 CAD before applicable tax and shipping. NVIDIA designs the DGX Spark hardware, the GB10 Grace Blackwell Superchip and DGX OS. D-Central’s role is Canadian sourcing, order-specific verification, first-boot configuration in Quebec and a practical handoff for local inference. We are an independent reseller and integrator; we do not manufacture the system, NVIDIA software or the models you choose to run.

What this configured DGX Spark order includes

• One NVIDIA DGX Spark with the GB10 Grace Blackwell Superchip, 128 GB LPDDR5x unified system memory and 4 TB NVMe storage.
• Order confirmation covering the exact supplied hardware, included accessories, configuration request, delivery destination and then-current procurement estimate before D-Central commits the hardware order.
• First boot and baseline DGX OS checks, plus installation or validation of Ollama and a private browser-based interface when those components remain compatible with the installed DGX OS release.
• A written handoff identifying installed versions, basic access steps and a model-fit discussion based on the model files, quantization, context target and workload you disclosed.
• Quebec-based configuration and support for the delivered setup scope. Custom application development, data migration, production monitoring, RAG ingestion, fine-tuning and enterprise platform support are separate work unless they appear in the written order.

The price is for the configured product described here; it is not a claim that D-Central has inventory on hand. Availability, delivery timing and the exact order cutoff are confirmed in writing. We do not substitute a 1 TB unit, another DGX product or a third-party “equivalent” without your approval.

Official platform specifications

According to the NVIDIA DGX Spark hardware documentation, the GB10 combines a 20-core Arm CPU with Blackwell GPU architecture. The platform uses a 256-bit, 128 GB LPDDR5x unified memory pool with stated memory bandwidth of 273 GB/s. NVIDIA lists 1 TB and 4 TB NVMe configurations; this product page is specifically for 4 TB. Connectivity documented by NVIDIA includes 10 GbE, ConnectX-7, Wi-Fi 7, Bluetooth 5.4, four USB-C ports and HDMI 2.1a. NVIDIA lists the unit at approximately 150 × 150 × 50.5 mm and 1.2 kg, with a 140 W GB10 TDP.

Those are NVIDIA specifications, not D-Central benchmarks. Port use, monitor support, cable requirements, network integration and sustained application performance depend on the software release and your environment. Use the supplied power adapter: NVIDIA’s known-issues documentation specifically cautions against third-party adapters. Confirm any rack, KVM, USB device, display or network requirement with us before the hardware order is committed.

Why 128 GB unified memory matters

Local inference is constrained by more than a model’s advertised parameter count. The weights, quantization format, context window, KV cache, runtime overhead, concurrent requests and multimodal components all consume memory. DGX Spark’s unified pool lets the CPU and GPU work from a large common memory space instead of forcing every accelerated workload into a conventional 24 GB or 48 GB discrete-GPU frame buffer. That creates useful room for larger quantized models, longer contexts and experimentation that would otherwise require multiple cards or CPU offload.

It does not mean every model below a particular parameter number will run well, or at all. NVIDIA positions DGX Spark for models up to 200 billion parameters and two linked systems for larger workloads. D-Central treats that as a platform capability statement, not a universal speed or fit guarantee. A model’s license, architecture, precision, runtime support and actual files still decide what works. Before ordering, send the exact repository or model name, quantization, expected context and concurrency. We can discuss fit, but we will not invent tokens-per-second figures for a workload we have not measured on the final configuration.

DGX OS, Arm64 and software compatibility

DGX OS for Spark is NVIDIA’s Ubuntu-based operating environment with the drivers and platform settings prepared for the hardware. NVIDIA documents Docker integration, access to NVIDIA NGC software and tools such as JupyterLab and NVIDIA Sync in the DGX Spark software guide. D-Central can prepare an open local-inference starting point, but NVIDIA owns and maintains the base platform and its update path.

The CPU architecture is Arm64. That boundary matters. Software distributed only as an x86-64 binary will not become compatible because the machine has an NVIDIA GPU. Native Windows applications are not part of this listing, and DGX Spark is not sold here as a Windows workstation. Containers also need an Arm64-compatible image and a stack compatible with the installed NVIDIA driver. Current NVIDIA documentation points users to NGC containers for actively supported builds of frameworks and inference tools. If your workflow depends on a specific PyTorch, CUDA, TensorRT-LLM, vLLM, NIM, Docker image, Python wheel or proprietary plugin, provide its version and architecture requirements before ordering.

NVIDIA also documents that GPU passthrough is not supported on DGX Spark. Enterprise assistance for NVIDIA AI Enterprise requires the relevant NVIDIA entitlement; it is not included merely because this system was configured by D-Central. Model licenses, API subscriptions and commercial software licenses remain the buyer’s responsibility. We can install compatible open components within the agreed setup, but we cannot grant rights that belong to NVIDIA, a model publisher or another software vendor.

A practical local-inference handoff

Our default objective is a machine that arrives with a documented starting point, not a mysterious black box. After receipt, D-Central checks the supplied identity and storage configuration, starts the vendor OS, records the relevant versions, applies the agreed baseline and validates that the requested interface can launch. If Ollama and the selected private UI are part of the order, we validate their basic local path with a small compatible test model. That check proves the installation path; it is not a benchmark and does not certify your undisclosed production workload.

Your handoff can also record the local network address, administrative boundary, update approach and how to remove the test model. Credentials are delivered through the agreed channel. You remain responsible for user access, backups, confidential data, firewall rules, Internet exposure and model governance after acceptance. If the system must be fully offline, say so before ordering: dependencies and model files need a planned transfer method, and some third-party packages or license checks may expect Internet access.

When DGX Spark is the right path

Choose this platform when CUDA ecosystem alignment, a large unified memory pool, compact physical size and NVIDIA’s integrated DGX software path matter more than x86 compatibility or component-level expansion. It is a strong laboratory, development and private-inference platform for teams that want one accountable baseline instead of choosing a motherboard, GPU, driver branch and enclosure independently.

A discrete-GPU tower can be the better answer when the target model fits entirely in available VRAM and maximum throughput, replaceable GPUs, conventional x86 software or PCIe expansion is the priority. An AMD Ryzen AI Max+ system can be the better desk machine when x86 and Windows-or-Linux flexibility matter more than CUDA. Compare the paths in our DGX Spark versus custom AI build guide, or review the AMD Strix Halo 128 GB desktop and custom GPU inference rig. These are architecture choices, not a claim that one machine wins every workload.

Compatibility boundary before you order

• Supported baseline: the NVIDIA-supplied DGX Spark hardware and DGX OS path, Arm64-compatible packages, and software explicitly agreed in the order confirmation.
• Must be confirmed: the exact model and quantization, context and concurrency target, container architecture, required framework versions, storage growth, display/KVM devices, network authentication, offline operation and any proprietary integration.
• Not represented as included: native Windows, x86-only binaries, GPU passthrough, guaranteed throughput, guaranteed fit for every “200B” model, production SLA, data migration, ongoing managed service, enterprise NVIDIA support entitlement or rights to third-party models and software.
• Environmental responsibility: the buyer provides suitable power, ventilation, network security and a location within NVIDIA’s documented operating limits. Any non-standard adapter or physical integration must be approved before use.

Ordering, payment and returns

The listed price is $9,449 CAD before applicable taxes and shipping. The order becomes configuration-specific once D-Central starts sourcing, configuration or service work. We confirm the exact 4 TB item, included accessories, requested setup, destination, shipping charge and current procurement estimate in the order record. No delivery date or stock status is promised by this page. Changes requested after sourcing begins may require a revised price and may not be possible.

This is configured, made-to-order hardware. Under D-Central’s Return and Refund Policy, configured builds, special orders and made-to-order work are non-refundable once sourcing, assembly, configuration or service work has started, except where the unit arrives defective or dead on arrival. The policy provides 30 days from delivery to report a defect or DOA for review; that is not a change-of-mind return window. Contact support with the order number and evidence, and wait for return authorization before shipping anything. The item must remain complete with included accessories and must not be damaged by misuse, installation error, modification or improper handling. Initial shipping is generally non-refundable, and return shipping responsibility follows the approved written instructions. The current policy and the written order confirmation govern; this page does not create an additional warranty or return promise.

DGX Spark FAQ

Is D-Central the manufacturer or an authorized NVIDIA service provider?

No such status is claimed. NVIDIA manufactures and documents DGX Spark. D-Central independently sources the specified system, performs the agreed Quebec setup and supports that handoff. NVIDIA entitlements and manufacturer processes remain NVIDIA’s.

Will a 200B model definitely run?

No. NVIDIA’s platform positioning is not a blanket compatibility or performance guarantee. Send the exact model, files, quantization, runtime, context and concurrency. Model licensing and software support also have to align.

Can I install Windows or run my x86 workstation applications?

This listing is for the Arm64 DGX OS platform, not a Windows/x86 workstation. Do not order it for an x86-only dependency unless that dependency’s publisher provides a supported Arm64 path that we confirm in writing.

Does the price include ongoing AI administration or custom RAG work?

No. It includes the hardware and the written first-boot configuration scope. Managed updates, application integration, RAG ingestion, fine-tuning, backups and production operations require a separate scope.

Can the unit be delivered fully offline?

Yes, an offline handoff can be planned, but it must be specified before ordering. We need to agree on dependency, container and model-file transfer, and we cannot remove Internet or license requirements imposed by third-party software.

What happens if it arrives damaged or does not start?

Contact D-Central promptly with the order number, photos or video and a clear fault description. The posted return policy allows a defect or DOA report within 30 days of delivery for review. Do not return the unit until D-Central provides written instructions.

What should I send for a compatibility check?

Send the model repository or exact file, quantization, desired context, expected simultaneous users, preferred runtime, required integrations, network policy, display needs and any hard latency target. We will identify known boundaries and state what the quoted setup will actually validate.

Sources and review: NVIDIA hardware, DGX OS, software, known issues and support documentation; D-Central Return and Refund Policy. Reviewed 25 August 2026. Specifications and software support can change; the written order confirmation controls the supplied configuration.