AI ComparisonsNov 19, 20269 min read

DGX Spark vs DGX Station: $6,950 to Hold a Model, $99k to Serve It

DGX Spark vs DGX Station: the $6,950 Spark holds 200B-parameter models but reads at 273 GB/s. The Station's GPU memory is 26 times faster, at about $99k.

On this page
  1. DGX Spark vs DGX Station at a glance
  2. How much does a DGX Spark cost now?
  3. Which matters more, memory size or memory speed?
  4. Which open models fit on a DGX Spark or a DGX Station?
  5. Can you link two DGX Sparks instead of buying a Station?
  6. What does a DGX Station cost, and what does it need?
  7. Does a DGX Spark or DGX Station keep data in Canada?
  8. What we did not test
  9. FAQ
  10. Sources

On 11 November 2026, NVIDIA's Marketplace lists the DGX Spark at $6,950, up from $3,999 at launch. The DGX Station has no NVIDIA price; Exxact's starts at $98,893.30. The Spark holds 128 gigabytes and reads them at 273 gigabytes a second. The Station's 252 gigabytes of GPU memory read about 26 times faster. So the Spark can hold a big model, but it writes answers slowly.

If one person is trying models, fine-tuning small ones or building a prototype, buy the Spark.

If a team needs answers at reading speed from a model larger than about 100 gigabytes, the Station is the desk box that does it.

If the model is Kimi K3.1 Pro, neither: its 1.56 terabytes of weights fit on no desk box.

DGX Spark vs DGX Station at a glance

DGX SparkDGX Station
ChipGB10 Grace BlackwellGB300 Grace Blackwell Ultra
Memory128 GB, shared by CPU and GPU252 GB GPU memory + 496 GB CPU memory = 748 GB
How fast it reads that memory273 GB/s7.1 TB/s (GPU memory), 396 GB/s (CPU memory)
AI compute, NVIDIA's figure at FP4Up to 1 petaflop (with sparsity)Up to 20 petaflops
Linking unitsConnectX-7, two ports up to 200 Gb/s eachConnectX-8, two ports at 400 Gb/s each
Power240 W power supply1,600 W, on a 20-amp circuit
Size150 × 150 × 50.5 mm, 1.2 kgDesktop tower
Price, read 11 Nov 2026$6,950 (NVIDIA Marketplace, 4 TB)No NVIDIA price; Exxact from $98,893.30

GPU memory is the memory the graphics processor reads while it runs a model. Specs come from NVIDIA's DGX Spark and DGX Station product pages, read on 11 November 2026. FP4 is a 4-bit number format; NVIDIA quotes compute at FP4 because it gives the largest figure.

Early coverage of the Station quoted 784 gigabytes of memory. That total used an earlier 288-gigabyte GPU memory spec. NVIDIA's page now lists 252 gigabytes, which ServeTheHome noted on 20 March 2026 is about 12% less than the server version of the chip.

How much does a DGX Spark cost now?

The 128-gigabyte Spark costs $6,950 on NVIDIA's US Marketplace, read on 11 November 2026. NVIDIA blamed memory supply for the February increase, and Implicator reported its own unit out of stock in October.

Date128 GB DGX Spark priceSource
October 2025 (launch)$3,999NVIDIA developer forum
Week of 23 Feb 2026$4,699NVIDIA developer forum, citing "worldwide constraints in memory supply"
October 2026$6,950Implicator, 4 October 2026
11 Nov 2026$6,950NVIDIA Marketplace, read today

A cheaper 64-gigabyte tier started shipping on 23 October 2026 from $4,999, sold only through partners (Acer, ASUS, Dell, GIGABYTE, HP and MSI), per Implicator. It reads memory at the same 273 gigabytes a second, so it is the same speed with half the room. NVIDIA rates it for models up to 100B parameters, against 200B for the 128-gigabyte unit.

Partner boxes built on the same chip sit beside it on the Marketplace. The ASUS Ascent GX10 costs $5,999 with a 1-terabyte drive or $7,999 with 4 terabytes. The MSI EdgeXpert costs $8,499.99. Our DGX Spark alternatives post covers them and the non-NVIDIA options.

Which matters more, memory size or memory speed?

Memory size decides whether a model fits; memory speed decides how fast it writes. To write each token, a piece of a word, the GPU reads the model's active weights from memory. A model that fits but sits in slow memory still answers, slowly.

LMSYS, the team behind the SGLang serving software, measured this on a Spark in its review of 13 October 2025. Llama 3.1 70B at 8-bit precision read a prompt at 803 tokens per second and wrote its answer at 2.7 tokens per second. A person reads faster than that. The same review measured the small gpt-oss-20b at 49.7 tokens per second on Ollama, one request at a time. Same box, different model weight per token.

Two things soften this. Mixture-of-experts models switch on only part of themselves for each token, so a 120B-parameter model can write far faster than a dense 70B one. And 4-bit formats halve the bytes again. The Spark is a fine box for those models. Large dense models are slow on it.

The Station has two speeds. Its 252 gigabytes of GPU memory reads at 7.1 terabytes a second. Its 496 gigabytes of CPU memory reads at 396 gigabytes a second, only about 1.5 times the Spark. NVIDIA's software can treat both as one pool. Any part of a model held in the CPU memory still reads at the slower speed. So the Station's fast tier is 252 gigabytes, not 748.

Which open models fit on a DGX Spark or a DGX Station?

Models up to about 100 gigabytes fit on a Spark with room to work. Models up to 252 gigabytes fit in the Station's fast memory. Nothing at Kimi K3.1 Pro's size fits on either. Weight size here is the Hugging Face download, read on 11 November 2026. A running model also needs memory for each conversation's context, so leave headroom.

ModelWeightsOne DGX Spark (128 GB)Two linked Sparks (256 GB)DGX Station (252 GB fast + 496 GB slow)
Qwen3-32B-FP8about 34 GBFitsFitsFits in fast memory
gpt-oss-120babout 65 GB of model filesFitsFitsFits in fast memory
MiniMax M3.1 Flashabout 252 GBNoNo headroom leftFits, spilling slightly into slow memory
DeepSeek-V4.1-Flashabout 510 GBNoNoFits, about half in slow memory
Kimi K3.1 Proabout 1.56 TBNoNoNo

MiniMax M3.1 Flash is the edge case. Its 252 gigabytes of weights equal two Sparks' 256 gigabytes with almost nothing left for context, so treat two Sparks as too small. NVIDIA rates four linked Sparks (512 gigabytes) for models up to 700B parameters, which makes room, at four boxes' cost. On a Station, MiniMax just overflows the 252-gigabyte fast tier.

Kimi K3.1 Pro needs a multi-GPU server. Our sibling guide on how much GPU memory you need for open models works through the sizes, and Kimi K3.1 Pro vs Kimi K3 covers the model itself. For the two mid-size models, see MiniMax M3.1 Flash vs DeepSeek V4.1 Flash.

Linking two Sparks doubles the room, not the speed. Each Spark has a ConnectX-7 network card, and NVIDIA sells a 0.4 m cable for $99. Two Sparks give 256 gigabytes, which NVIDIA rates for models up to 400B parameters.

Each box still reads its own memory at 273 gigabytes a second. The cable between them carries up to 200 Gb/s per port, roughly 25 gigabytes a second, which is far slower than either box's memory. Splitting a model across two boxes lets it fit. It does not turn them into a Station.

At Marketplace prices, two Sparks and a cable cost $13,999. That is about a seventh of a Station, for about a third of its total memory and none of its fast tier. NVIDIA also offers two- and four-unit bundles on request, without a listed price.

What does a DGX Station cost, and what does it need?

NVIDIA sets no list price. Partners quote their own, and the one we could read online starts at $98,893.30. That is Exxact's TensorEX DGX Station page, read on 11 November 2026. NVIDIA's page directs buyers to ASUS, Dell, Exxact, GIGABYTE, HP, MSI and Supermicro. ServeTheHome reported in March 2026 that most builders quote on request, with unconfirmed figures from $80,000 to $125,000.

The Station is a different class of machine to live with:

  • Power. NVIDIA lists 1,600 W total system power and a 20-amp circuit. Plan the circuit with your facilities team before it arrives. The Spark runs from a 240 W power supply.
  • Cooling and noise. A 1,600 W tower in a small office heats the room. Put it where the heat can go.
  • Availability. NVIDIA's own page still offers a sign-up for availability notices, so expect to wait on a partner quote.

For a server-room alternative at a similar memory size, see DGX Spark vs RTX PRO 6000 and RTX PRO 6000 vs H100.

Does a DGX Spark or DGX Station keep data in Canada?

Yes, by design: the model runs on the box, so prompts and outputs stay wherever the box sits. Put either one in a Canadian office or data centre, and inference stays in Canada with no cloud region to choose.

The trade-offs are the ones that come with owning hardware:

  • You own the box. Purchase, warranty, power and upkeep are yours, and so are serving software, patches and logs.
  • One box is not redundant. A desk machine that reboots takes the model with it. Production needs a second unit or a fallback, which our guide to on-premise, client cloud and managed cloud compares.
  • The Station needs a room. The power and cooling above apply. The Spark plugs into a normal outlet.
  • Prices are in US dollars. The prices above are from US listings; Canadian pricing and stock vary by partner.

If you would rather rent, the big three clouds offer L4, A100 and H100 GPUs in Canadian regions; our GPU cloud Canada guide lists them with prices read on 11 November 2026. In our experience the H100 and H200 there usually need a one- to three-year commitment, which makes an owned box easier to justify.

XY Space deploys open-weight models on dedicated GPUs in the client's own Canadian cloud account or on hardware they own, sizes the hardware to the model, and builds the workflow on top. How we deploy covers the build, and the Canadian residency guide covers the rules.

If you need a model running with inference kept in Canada, bring one job to a free 30-minute call.

What we did not test

We have not run a model on either box. The speed figures are LMSYS's, measured with October 2025 software, and serving software changes often. We did not measure the Station's speed when a model spills into its CPU memory; the CPU memory speed is NVIDIA's spec. Station prices are partner quotes that change, so ask for a current one.

FAQ

Is the DGX Station worth it over two DGX Sparks?

For serving a team, yes, if the model is over 100 gigabytes. Two Sparks hold 256 gigabytes but each reads memory at 273 gigabytes a second. The Station reads its 252 gigabytes of GPU memory at 7.1 terabytes a second. For one person testing models, two Sparks at $13,999 cost about a seventh as much and fit models up to 400B parameters, per NVIDIA.

Can a DGX Spark run a 200B-parameter model?

NVIDIA rates the 128-gigabyte Spark for models up to 200B parameters, which assumes a 4-bit format. It will load and answer. How fast depends on how much of the model is active per token: a mixture-of-experts model writes far faster than a dense one. LMSYS measured a dense 70B model at 2.7 tokens per second on the Spark.

How much does the DGX Spark cost in November 2026?

$6,950 for the 128-gigabyte, 4-terabyte unit on NVIDIA's US Marketplace, read on 11 November 2026. It launched at $3,999 in October 2025 and rose to $4,699 in February 2026. A 64-gigabyte tier from partners starts at $4,999. Partner 128-gigabyte boxes range from $5,999 to $8,499.99 on the same Marketplace.

Does the DGX Station have 748 GB or 784 GB of memory?

748 gigabytes, per NVIDIA's product page: 252 gigabytes of GPU memory and 496 gigabytes of CPU memory. The 784-gigabyte figure in early coverage came from an earlier spec with 288 gigabytes of GPU memory. Only the 252-gigabyte GPU tier reads at 7.1 terabytes a second; the rest reads at 396 gigabytes a second.

Can I run Kimi K3.1 Pro on a DGX Station?

No. Its weights download at about 1.56 terabytes, double the Station's 748 gigabytes of total memory. It needs a multi-GPU server, either in a cloud account or a data centre. MiniMax M3.1 Flash, at about 252 gigabytes, and DeepSeek-V4.1-Flash, at about 510 gigabytes, both fit on a Station.

Sources

Written by

Cho Yin Yong

Principal AI Solutions Engineer, XY Space

Principal AI Solutions Engineer at XY Space. University of Toronto lecturer for five years, co-author of two patents, winner of two competitive AI awards, and nine years of regulated engineering leadership.

More from Cho Yin Yong

Share this article

Work with us

We build the systems these posts describe, and we'll tell you in the first call whether yours is worth building.

Start a project
Work with us

Book a call.We'll come back with specifics.

Start with the map of your organization, or with the one job that hurts. Measured in hours and money, and everything we build stays yours.

Loading form…