DGX Spark vs DGX Station GB300: specs, prices and which models each one runs

The DGX Station is a 1,600 W deskside machine built on NVIDIA's GB300: 252 GB of HBM3e at 7.1 TB/s for the GPU plus 496 GB of LPDDR5X for the CPU, 748 GB of coherent memory in all. The DGX Spark is a 240 W box with 128 GB of LPDDR5x at 273 GB/s, about 26 times less GPU bandwidth. On 7 November 2026 Exxact listed a Station from $104,830 and Dell's configured Pro Precision with GB300 was $160,195, against $6,950 for a 128 GB Spark; NVIDIA sells the Station only through partners. Eight linked Sparks reach 1 TB for about $60,000, but in NVIDIA's benchmark they decode a 27B model at 52.6 tok/s, far below what one Station's bandwidth allows.

Specs from NVIDIA's DGX Station and DGX Spark pages, read 7 November 2026; full Spark details are on the specs page. NVIDIA's Spark FP4 figure is stated with sparsity; the Station page does not say, so I do not compute a compute ratio. The Spark price is the US list price tracked on the price page.

How big is the gap between a DGX Spark and a DGX Station?

About 26 times in GPU memory bandwidth and about 6 times in memory capacity, and bandwidth is what sets decode speed. Each generated token reads the active weights once, so the ceiling is bandwidth divided by bytes per token. A worked example with a dense 70B model at 4-bit (37 GiB, or 39.7 GB, of weights):

DGX Spark:   273 GB/s  / 39.7 GB = 6.9 tok/s ceiling
             measured: 6.0 tok/s (Llama 3.3 70B AWQ, vLLM 0.27.1, 2 Sep 2026)
DGX Station: 7,100 GB/s / 39.7 GB = 179 tok/s ceiling (weights in HBM)
             no published Station figure found; engines usually land well below the ceiling

The Spark figure is GPUwerk's own measurement on the benchmarks page; the Station line is arithmetic only. I could not find a published single-user tok/s for the DGX Station on a named model and engine as of 7 November 2026, so treat 179 tok/s as an upper bound, not a forecast.

The Station's 748 GB is not one fast pool. Only the 252 GB of HBM3e runs at 7.1 TB/s. The 496 GB of LPDDR5X behind the Grace CPU runs at 396 GB/s, about 1.45 times a Spark, and the GPU reaches it over the 900 GB/s NVLink-C2C. By my reading of those figures, any weights placed in LPDDR5X are read at no more than 396 GB/s, so a model that spills out of HBM loses most of the Station's advantage for the spilled part. Dell's "up to 1 trillion parameter models" claim depends on that slower tier.

Which models does each one run?

A Spark holds 120B-class models at 4-bit; a Station holds 200B to 400B-class models in its fast memory and roughly a 1T model across both tiers. Fit by weight size, before KV cache:

"Yes" means the weights fit with some room for context; KV cache and runtime memory come on top. The GLM-5.3 and 1T rows are my arithmetic on NVIDIA's stated sizes and capacities, not reported Station runs. NVIDIA's vLLM playbook covers both machines and describes the Station's memory as "Large HBM + Grace DRAM" without naming models.

The practical sweet spot for a Station is the 150 to 250 GB band: models that need two or more Sparks but fit entirely in HBM, where the Station is both simpler and much faster. Above 252 GB a Station still runs the model, but the gap to a Spark cluster narrows. Newer large MoE releases are listed with their sizes in the guide to large MoE models on multiple Sparks.

What does a DGX Station cost, and who sells it?

From about $105,000, through seven partners. NVIDIA's page lists ASUS, Dell, Exxact, Gigabyte, HP, MSI and Supermicro, tells buyers to "Contact a partner to order your system", and still carries a form to be notified "when DGX Station becomes available". What each partner showed on 7 November 2026:

US prices before tax. Dell's landing page says it is moving both its GB10 and GB300 machines under the Dell Pro Precision name and lists Windows 11 as "coming soon" for the GB300. NVIDIA's playbook site has a DGX Station section, including a two-Station link over two 400G ConnectX-8 rails, which suggests units are in customers' hands; I found no NVIDIA shipping announcement to confirm volume availability.

Per gigabyte of memory, the Spark is far cheaper; per GB/s of bandwidth, the Station is. My arithmetic:

The eight-Spark total uses $6,950 per Spark and the switch and cable prices from the four- and eight-Spark cluster guide. Prices checked 7 November 2026.

When are four or eight Sparks an alternative?

When you need capacity more than speed, and can live with manual cluster setup. NVIDIA has supported eight-node Spark clusters since 22 October 2026, with 1 TB across the nodes. Its benchmark shows what that buys: Qwen3.8-27B NVFP4 at 52.58 tok/s for one user on eight switched nodes, against 11.86 on one, and the full 470 GB GLM-5.3 at 31 tok/s with multi-token prediction. Training scales better: nanochat pre-training ran 7.79 times faster on eight nodes.

• Pick Sparks for running a 400B to 1T MoE model privately for a few users at modest speed, for the lowest cost per gigabyte, or for learning multi-node setups. Four Sparks are about $28,000 to $30,000 with cabling and give 512 GB, rated by NVIDIA for models up to 700B parameters.
• Pick a Station when per-user speed matters, when the model fits in 252 GB, when several developers share it (MIG splits the GPU into up to seven instances), or for fine-tuning, where the Station's compute and HBM are in another class. It is one machine with one memory system, so there is no tensor-parallel tuning.

Power is closer than it looks. Eight Spark supplies are rated at 1,920 W in total, though the power guide puts each Spark at 60 to 90 W during inference; the Station's 1,600 W needs a 20 A circuit, and HP specifies a dedicated 200 to 240 V line. A single Spark runs from a normal outlet and is quiet (see noise and cooling); ASUS quotes about 50 dB for its Station at full load.

Which should you buy?

For one developer working with models up to about 120B parameters, a Spark: it costs a fifteenth of a Station and runs the same CUDA stack. For a team that serves 150 to 250 GB models interactively or fine-tunes large models on site, the Station is the better machine, and the comparison with datacenter GPUs in B300 vs B200 vs H200 is the more relevant one. If you only need a Station-sized model a few hours a week, rented HBM by the hour is cheaper than either; the Spark vs H200 comparison works through that trade. You can test the Spark side on GPUwerk for $0.79 per hour, or $1.79 for a cabled pair, before pricing a six-figure workstation.

FAQ

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

• NVIDIA DGX Station specifications, NVIDIA
• NVIDIA DGX Spark specifications, NVIDIA
• Dell Pro Precision with GB300 (FCT6263), Dell
• Dell Pro Max NVIDIA AI Dev, Dell
• Exxact Valence NVIDIA DGX Station VWS-158270643, Exxact
• HP ZGX Fury AI Station, HP
• Super AI Station, Supermicro
• ASUS ExpertCenter Pro ET900N G3, ASUS
• DGX Station playbooks, NVIDIA build.nvidia.com
• Connect Two Nodes for Distributed Workloads, NVIDIA build.nvidia.com, 2026-10-15
• Serve LLMs with vLLM on One Device, NVIDIA build.nvidia.com, 2026-10-19
• Running Frontier-Class AI Models Across 8 NVIDIA DGX Spark Nodes, NVIDIA Technical Blog, 2026-10-22
• openai/gpt-oss-120b files, Hugging Face
• nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 files, Hugging Face