October 10, 2026 · Hardware · ~9 min read

DGX Station vs a DGX Spark cluster — price, memory and speed compared

If you're looking up the NVIDIA DGX Station price, your budget is in the order of a hundred thousand dollars, and you probably have one question: wouldn't it be better to buy a few smaller DGX Spark units and link them into a cluster? The answer starts with something worth saying plainly: they aren't in the same class.

A DGX Station is one machine with huge, very fast memory. A DGX Spark cluster is several small units that together give a lot of memory for a fraction of the price — but the memory is slower, and the units talk over a network. A cluster wins on capacity per dollar; a Station wins on speed. The rest of this article is the numbers that help you decide which matters more in your case.

Specs and prices were checked on October 9–10, 2026; each comes with a date and a source, because prices for this hardware are moving with memory prices.

The NVIDIA DGX family in brief

Under the DGX brand NVIDIA sells several classes of hardware that are easy to mix up:

• DGX Spark — a palm-sized computer with the GB10 chip and up to 128 GB of memory, for the desk. Partner machines such as the ASUS Ascent GX10 use the same chip.
• DGX Station — a tower workstation with the GB300 chip and 748 GB of memory, sold by NVIDIA's partners.
• DGX servers and rack systems — data-center hardware with many GPUs in one chassis or rack.

This article compares the first two, because that's the choice a company faces when it wants to run large models in-house, without a server room.

DGX Station GB300: price and specs

According to NVIDIA, the DGX Station has the GB300 Grace Blackwell Ultra Desktop Superchip: a 72-core Grace CPU and a Blackwell Ultra GPU, linked by NVLink-C2C at 900 GB/s. In total it has 748 GB of coherent memory, made of two parts with very different speeds:

• 252 GB of HBM3e next to the GPU — 7.1 TB/s of bandwidth;
• 496 GB of LPDDR5X next to the CPU — 396 GB/s.

It also has a ConnectX-8 SuperNIC with up to 800 Gb/s, which can link two stations, up to 20 petaFLOPS of AI compute, and — relevant in an office — a total system power of 1,600 W, requiring a 20 A circuit. NVIDIA says the Station supports models of up to a trillion parameters.

DGX Station GB300 price. NVIDIA doesn't publish an official price and doesn't sell its own version — you order it from partners: ASUS, Dell, Exxact, Gigabyte, HP, MSI and Supermicro. In the US, the lowest public listing on October 9, 2026 was about $92,550 (Supermicro's eStore, in stock); Exxact's base configuration was $96,250 and MSI's XpertStation WS300 $99,999 (out of stock at Newegg). A European distributor listed nine models on October 10, 2026, from €104,800 to €205,000 (the page doesn't clearly say whether VAT is included). Lead times quoted across OEMs ranged from about two to sixteen weeks.

Two or four DGX Spark units: how much memory they give

A single DGX Spark has 128 GB of memory with 273 GB/s of bandwidth and a ConnectX-7 200 Gb/s network card. Two units are linked directly with one cable and give twice the memory; four need a 200 Gb/s switch and give four times the memory of one. NVIDIA covers these setups in its official playbooks.

We run such a cluster every day: on two units we serve MiniMax M2.7 — a 230-billion-parameter model — at 33–36 tokens per second. How the cluster is wired, how to scale it and what we measured is on our DGX Spark cluster page.

A note for buyers: since October 23, 2026, NVIDIA's partners also sell a 64 GB DGX Spark. For a cluster aimed at large models, the 128 GB version is the one that counts — the differences between versions and partner machines are in our ASUS GX10 vs DGX Spark comparison.

DGX Station vs DGX Spark — 2 and 4 units in one table

We calculated the cluster prices from two October 2026 US prices per unit: an ASUS GX10 with 1 TB at $5,999 (a US retailer, October 10) and the 128 GB DGX Spark tier at about $6,950 (ServeTheHome, October 3). Per gigabyte of memory, four units come to roughly $47–54, and the cheapest Station listing to about $124 — more than twice as expensive per gigabyte. For that difference, the Station gives you speed a cluster doesn't have.

Which models fit — capacity vs speed

Capacity decides whether a model runs at all. Memory bandwidth decides how fast it answers. A few examples, with sizes in the 4-bit formats in which such models are usually run locally:

• Models up to about 120B parameters fit on a single 128 GB DGX Spark. You don't need a cluster.
• DeepSeek V4 Flash (284B) — about 156–175 GB in common 4-bit formats (community estimates): too big for one Spark, fits on two with room for context. On a Station it fits entirely in the fast HBM3e.
• MiniMax M2.7 (230B) — runs on our two units at 33–36 tokens per second.
• DeepSeek V4 Pro (1.6T) — the official weights are already mostly 4-bit and still take at least about 800 GB. That's more than four Sparks and more than the Station's 748 GB; it would only fit with heavier compression, which usually costs answer quality.

And speed? We don't have our own DGX Station measurements, so we don't quote tokens per second. The spec shows the scale, though: the Station's HBM3e runs at 7.1 TB/s, about 26 times the 273 GB/s of one Spark. Memory bandwidth is the main driver of generation speed, although it doesn't translate one-to-one. There's a caveat on the Station side too: that advantage applies to models that fit in the 252 GB of HBM3e. Whatever doesn't fit goes to LPDDR5X at 396 GB/s — slower, though still faster than a Spark.

When to choose a Station, when a cluster

Practical criteria, in the order worth checking:

1. Budget. If spending around $100,000 isn't realistic, the answer is simple — and it doesn't mean giving up large models, only slower answers.
2. How many users at once. For a handful of people asking questions now and then, tens of tokens per second is usually enough. With dozens of simultaneous requests or large document batches, the Station's fast memory starts to matter.
3. Speed or capacity. If the key is running a big model at all — a cluster. If the key is response time with a model that fits in 252 GB — a Station.
4. Growing step by step. A cluster can start with one unit, add a second, then two more — without spending the whole budget at once. A Station is bought in one go.
5. Power and space. 1,600 W on a 20 A circuit is a requirement to check in an ordinary office before buying. Sparks plug into regular sockets.

Frequently asked questions

How much does a DGX Station cost?

NVIDIA doesn't publish an official price and doesn't sell its own version; you buy it from partners such as ASUS, Dell, Exxact, Gigabyte, HP, MSI and Supermicro. In the US, the lowest public listing on October 9, 2026 was about $92,550 (Supermicro), with others at about $96,000–100,000. A European distributor listed models from €104,800 to €205,000 on October 10, 2026.

DGX Spark vs DGX Station: what's the difference?

DGX Spark is a palm-sized computer with the GB10 chip and up to 128 GB of memory at 273 GB/s. DGX Station is a tower with the GB300 chip and 748 GB of memory, 252 GB of it HBM3e at 7.1 TB/s, drawing 1,600 W. Several Sparks linked into a cluster give a lot of memory for much less money, while a Station is much faster.

How much memory does a DGX Station have?

748 GB of coherent memory: 252 GB of HBM3e at 7.1 TB/s next to the GPU and 496 GB of LPDDR5X at 396 GB/s next to the CPU.

Is a DGX Spark cluster faster than a DGX Station?

No. A cluster gives more memory per dollar, but each Spark's memory runs at 273 GB/s, while the Station's HBM3e runs at 7.1 TB/s — about 26 times faster. For models that fit in the Station's 252 GB of HBM3e, the Station will answer much faster. We haven't measured a Station ourselves.

Can a DGX Station run a trillion-parameter model?

NVIDIA says it supports models of up to a trillion parameters. In practice it depends on the format: DeepSeek V4 Pro (1.6 trillion parameters) needs at least about 800 GB even with mostly 4-bit weights — more than the Station's 748 GB.

Sources

• NVIDIA, DGX Station — specifications — GB300, 748 GB of coherent memory (252 GB HBM3e at 7.1 TB/s + 496 GB LPDDR5X at 396 GB/s), NVLink-C2C 900 GB/s, ConnectX-8 up to 800 Gb/s, up to 20 PFLOPS, 1,600 W on a 20 A circuit, models up to a trillion parameters, partners. Checked October 10, 2026.
• NVIDIA, DGX Spark — specifications — 128 GB, 273 GB/s, ConnectX-7 200 Gb/s, 240 W power supply; DGX Spark Playbooks — linking units with a cable or through a switch.
• BetterClaw, DGX Station price overview (prices checked October 9, 2026) — Supermicro $92,550.85, Exxact $96,250, MSI $99,999, no official NVIDIA price, lead times.
• buyzero.de, NVIDIA DGX Station GB300 — EUR prices of nine partner models, €104,800–205,000. Checked October 10, 2026.
• ServeTheHome, NVIDIA DGX Spark 64GB Launched and Big 128GB GB10 Price Increases (October 3, 2026); Excaliber PC — ASUS GX10 1 TB at $5,999, October 10, 2026.
• DeepSeek, DeepSeek-V4-Pro model card — 1.6T parameters, FP4 + FP8 weights; localmodel.run, DeepSeek-V4-Flash GGUF sizes — 4-bit file sizes (estimates).
• ailone — our MiniMax M2.7 measurements on two units (see the cluster page).

As of October 10, 2026. Prices come from public offers on the day we checked, change often and don't include extra drives, cards or service — confirm them with the seller before you buy. Model sizes are estimates for 4-bit formats; actual memory needs also depend on context length and the number of simultaneous users.