Can the DGX Spark train large language models?

It can fine-tune and run models up to about 200 billion parameters, but its memory bandwidth makes it much slower than a data-center GPU for heavy training.

How does it compare with a cloud GPU rental?

It gives you a private, always-available machine with no per-hour billing, but raw throughput is lower than a rented high-end GPU.

Does it run normal desktop apps?

It runs DGX OS on an Arm processor, so AI and Linux tooling work well, while some x86-only software will not run natively.

Can two units be connected?

Yes, two DGX Spark systems can be linked over their high-speed network ports to work with larger models.

The unit, an external 240W power supply and the preinstalled NVIDIA AI software stack; storage comes as 1 TB or 4 TB NVMe.

Gamers and general desktop users, since it is built for AI development rather than everyday computing.

Where is the NVIDIA DGX Spark cheapest?

At Alibaba.com, for $7,500.00, as of our latest check.

Where can I buy the NVIDIA DGX Spark?

4 stores we checked sell it, including Newegg, Walmart, Alibaba.com, Amazon.

NVIDIA DGX Spark

Best price now

$7,500.00at Alibaba.comChecked 1 h ago

Compared across 4 stores. Prices are checked against each store and labelled with when.

Visit Alibaba.com

NVIDIA DGX Spark

Best price now

$7,500.00at Alibaba.comChecked 1 h ago

Compared across 4 stores. Prices are checked against each store and labelled with when.

Visit Alibaba.com

Ask our agent about this product

It checks today's price at every store, compares the alternatives and answers in about a minute.

Price facts

Lowest price today: $7,500.00 at Alibaba.com (checked Nov 11, 2026).
Checked at 4 stores.

Our take on the NVIDIA DGX Spark

The DGX Spark is a small desktop AI computer built around NVIDIA's GB10 Grace Blackwell Superchip, rated at about 1 petaFLOP of AI performance. It pairs a 20-core Arm CPU with 128 GB of unified LPDDR5x memory, so it can run and fine-tune models up to roughly 200 billion parameters locally. It ships with the NVIDIA AI software stack and DGX OS, aimed at developers who want a quiet, low-power alternative to renting cloud GPUs. Against mini-PC rivals it wins on unified memory capacity and CUDA tooling, but its memory bandwidth of 273 GB/s is far below a data-center GPU, so training throughput is modest.

Best forAI developers prototyping and fine-tuning models locallySmall teams that want a private, desk-side alternative to cloud GPUsEdge and robotics work needing CUDA on ArmStudents and researchers learning the NVIDIA AI stack

Pros

128 GB unified memory runs large models that consumer GPUs cannot fit
Full CUDA and NVIDIA AI software stack out of the box
Compact, low-power design with an external 240W supply
Wi-Fi 7, 10GbE and USB4 for fast local networking
Two units can be linked to work with larger models

Cons

273 GB/s memory bandwidth limits training speed versus data-center GPUs
Arm CPU means some x86 desktop software will not run natively
Best suited to AI workloads, not general desktop use
Model support tops out around 200 billion parameters

NVIDIA DGX Spark prices at every store

Alibaba.comNew
In stock
$7,500.00+ $25.10 shippingChecked 1 h ago
Visit Alibaba.com
WalmartNew
In stockSold by Joes Tech ShopReturns: Sold by a Walmart Marketplace seller; returns go to the seller under the seller's returns policy.
$8,250.00Free shippingChecked 55 min ago
Visit Walmart
NeweggNew
In stock
$9,399.00Free shippingChecked 21 min ago
Visit Newegg
AmazonNew
Availability unknown
As an Amazon Associate, Deal.ai earns from qualifying purchases.
$7,799.99Free shippingChecked 1 h ago
Visit Amazon

Store links may earn us a commission. Some prices via Klarna.

NVIDIA DGX Spark price history

Lowest price we've seen: $7,500.00 on Nov 11 at Alibaba.com. Tracked across 3 stores since Nov 11.

What reviewers say

Hands-on coverage treats the DGX Spark as a niche developer tool: reviewers like the memory capacity and CUDA support in a tiny box, but note that its bandwidth and Arm platform keep it well behind full GPUs for heavy training.

What people mention

MemoryPraisedReviewers highlight the 128 GB unified pool as the reason it can hold models far bigger than consumer GPUs allow.
BandwidthCriticisedCommentators note 273 GB/s is low next to data-center GPUs, so training and inference are slower.
SoftwarePraisedThe preloaded NVIDIA AI stack and DGX OS are seen as a real convenience for getting started.
Arm compatibilityMixedThe Arm CPU is fine for AI tooling but some x86 desktop software needs workarounds.
Size and powerPraisedIts compact, low-power design is called out as unusual for this much AI capability.

About the NVIDIA DGX Spark

It has a 20-core Arm processor (10 Cortex-X925 and 10 Cortex-A725), 6,144 CUDA cores and 128 GB of LPDDR5x unified system memory with 273 GB/s of bandwidth.

Storage is an NVMe M.2 SSD with self-encryption, and a 240W external power supply is included.

Processor: 20-core Arm (10 Cortex-X925 + 10 Cortex-A725)
Memory: 128 GB LPDDR5x unified, 273 GB/s
Storage: 1 TB or 4 TB NVMe M.2, self-encrypting
CUDA cores: 6,144
Ports: 4x USB Type-C, 1x HDMI 2.1a
Weight: 1.2 kg (2.6 lbs)
Power supply: 240W external, included
Color: Gold

Specifications

Processor: 20-core Arm processor (10 Cortex-X925 + 10 Cortex-A725)Source
Memory: 128 GB LPDDR5x unified system memorySource
Memory bandwidth: 273 GB/sSource
Storage: 1 TB or 4 TB NVMe M.2 with self-encryptionSource
Ports

Questions and answers

Can the DGX Spark train large language models?: It can fine-tune and run models up to about 200 billion parameters, but its memory bandwidth makes it much slower than a data-center GPU for heavy training.
How does it compare with a cloud GPU rental?: It gives you a private, always-available machine with no per-hour billing, but raw throughput is lower than a rented high-end GPU.
Does it run normal desktop apps?: It runs DGX OS on an Arm processor, so AI and Linux tooling work well, while some x86-only software will not run natively.
Can two units be connected?: Yes, two DGX Spark systems can be linked over their high-speed network ports to work with larger models.
What is in the box?: The unit, an external 240W power supply and the preinstalled NVIDIA AI software stack; storage comes as 1 TB or 4 TB NVMe.
Who is it not for?: Gamers and general desktop users, since it is built for AI development rather than everyday computing.
Where is the NVIDIA DGX Spark cheapest?: At Alibaba.com, for $7,500.00, as of our latest check.
Where can I buy the NVIDIA DGX Spark?: 4 stores we checked sell it, including Newegg, Walmart, Alibaba.com, Amazon.