NVIDIA DGX Spark Price (2026): What It Costs, Where to Buy, and Is It Worth It?

Launch price and today's price

NVIDIA opened orders for DGX Spark in October 2025 at $3,999. Because memory is fixed at 128 GB on every unit, prices differ mainly by storage (1 TB, 2 TB or 4 TB), cooling, chassis and warranty. Regional prices include local taxes, so UK and EU prices are higher in headline terms. Memory market conditions have also moved prices since launch, so always check a live listing before budgeting.

Founders Edition vs partner units

All GB10 systems share the chip and memory, so AI performance is broadly similar. Differences show up in noise, sustained clocks and service options.

Total cost of ownership

DGX Spark draws around 240 W at the wall at most, so running cost is modest. Over three years the main cost is the purchase price. Compared with renting a cloud GPU, the break-even point depends on how many hours a day you use it — use our Local vs Cloud calculator with your own numbers.

• Purchase price: the dominant cost.
• Electricity: a few hundred kWh per year even under heavy use.
• Storage add-ons: an external NVMe drive for model libraries.
• Optional second unit for 2-node clustering over ConnectX-7.

Is DGX Spark worth it?

It is worth it if you need 128 GB of CUDA-capable memory on your desk, want NVIDIA's DGX software stack, and prototype models or agents daily. It is not the fastest option for pure token speed — its ~273 GB/s memory bandwidth is lower than high-end discrete GPUs or top Mac Studio configs — and it is not a Windows PC. If you want Windows and portability, look at RTX Spark systems instead.

DGX Spark vs renting cloud GPUs

A simple break-even view: divide the purchase price by what you would otherwise spend on cloud GPU hours or API tokens each month. Cloud prices vary widely by provider and GPU, so plug in your own numbers on the TCO calculator.

Who should and shouldn't buy one

DGX Spark is a specialist machine. It is excellent value for some people and the wrong buy for others.

• Good fit: AI developers prototyping on the NVIDIA stack before deploying to bigger DGX/cloud systems.
• Good fit: an always-on, private agent host for a small team.
• Poor fit: gaming or general Windows use — it runs Linux (DGX OS).
• Poor fit: training large models from scratch — bandwidth and compute are workstation-class, not data-centre-class.

Clustering two units: total cost

Two DGX Spark units can be linked over their built-in ConnectX-7 networking to pool 256 GB for larger models. Budget for two units plus a suitable high-speed cable; NVIDIA has positioned two-unit clustering as the supported configuration.

How RTXsparks sources and checks these numbers

RTXsparks is independent and not affiliated with NVIDIA or any OEM. Prices come from manufacturer launch announcements and retailer listings (US MSRP before tax unless stated). Specs come from official product pages. Performance figures are either our own lab runs, labelled community submissions, or clearly marked estimates derived from memory bandwidth. Anything a vendor has not publicly confirmed is labelled 'unconfirmed'. See the methodology page for the full SCS, ARS and TLS scoring rules.

• Prices re-checked weekly during the launch window.
• Estimates are always labelled as estimates — never presented as measured results.
• Corrections are welcome via the benchmark submission page.

Founders Edition vs partner models

NVIDIA sells its own Founders Edition at $3,999. Partner versions from OEMs use the same GB10 chip and 128 GB memory; they differ mainly in storage, case design, cooling and support. Check each partner's listing for current pricing — we only quote numbers the vendor has published.

Total cost of ownership in year one

The sticker price is not the full cost. A typical first-year budget looks like this (estimates, US).

Buyer's checklist before you order

Most regret with these machines comes from buying too little memory. Run through this list before checking out — it applies to every RTX Spark laptop and to DGX Spark.

• Write down the largest model you want to run and look up its memory need in our VRAM guide.
• Add 20–30% headroom for context (KV cache), the OS and your other apps.
• Memory is soldered on every Grace-Blackwell system — you cannot add more later.
• Check the return window: launch-window units often have 14–30 day returns depending on the retailer.
• Confirm the exact chip configuration (core count) — OEMs ship more than one.
• Budget for fast external storage if you plan to keep a large model library.