DGX Spark Buyer’s Guide 2026: Models, Pricing and Which One to Buy

Buying an NVIDIA DGX Spark can seem complicated because while NVIDIA sells its own DGX Spark model, they also license the same core GB10 Grace Blackwell platform to be manufactured and sold by several OEMs, including ASUS, Dell, Gigabyte, HP, Lenovo, MSI and Acer, who all offer different specifications and pricing.

However, the most important thing to note is that all these Spark OEM models use the same NVIDIA GB10 Superchip and unified memory architecture. The differences lie in storage, cooling, software, support, price, and availability. 

Until recently, all models also came with the same 128GB memory, but NVIDIA has now introduced a lower-cost Spark with 64 GB.

For a current overview of the market, see our NVIDIA DGX Spark price and model comparison.

Different NVIDIA DGX Spark Models

The term “DGX Spark” can refer to computers sold directly by NVIDIA or OEM machines from ASUS, HP, Dell, Lenovo, Gigabyte, MSI and Acer. However, these OEM machines are technically GB10-based systems rather than DGX Sparks.

That said, for simplicity, we use DGX Spark in this guide to refer to both the NVIDIA system and all other OEM GB10 machines.

DGX Spark = any machine with an NVIDIA GB10 Grace Blackwell Superchip and 128GB or less memory. 

Across the standard 128 GB models, you get:

• NVIDIA GB10 Grace Blackwell Superchip.
• NVIDIA Blackwell GPU.
• 20-core Arm CPU.
• 128 GB or 64GB unified LPDDR5X memory.
• 273 GB/s memory bandwidth.
• Up to 1 PFLOP of sparse FP4 AI performance.
• NVIDIA ConnectX-7 networking at up to 200 Gbit/s.
• 10 GbE.
• Wi-Fi 7.
• NVIDIA AI software stack.
• 1 TB, 2 TB or 4 TB NVMe storage, depending on the OEM and configuration.

PI3g Spark Buying Tip: We sell a full range of DGX Spark systems, but we do not recommend our customers paying more for any particular 128 GB GB10 system expecting radically greater core AI performance than another one. What Spark model is best to buy has more to do with what you need in terms of storage, cooling, bundled software, support, physical design, and availability.

To explain the actual differences between models and where each one makes the most sense, the section below gives you a short introduction to each major DGX Spark model and what each is best for.

Top DGX Spark 2026 Choice: Gigabyte AI TOP ATOM

The Gigabyte AI TOP ATOM is one of our best-selling Spark models due to its combo of price, storage options, and software. It is relatively easy to set up and is competitively priced. 

It uses the same 128 GB GB10 platform as the other systems, with 1 TB and 4 TB configurations available. Gigabyte also offers 4 TB versions using either PCIe Gen4 or PCIe Gen5 storage.

The ATOM ships with Gigabyte’s AI TOP Utility, which provides tools for downloading models, inference, RAG and machine learning while making it relatively easier to get started running local models compared to the other models here.

Price is another attraction. Our October 2026 pricing data identified the 4 TB Gigabyte AI TOP ATOM as the lowest-cost 4 TB GB10 option we can supply. Our partner site, BuyZero, currently lists Gigabyte AI TOP ATOM configurations from €5,033.61 excluding VAT.

Buy the Gigabyte AI TOP ATOM from BuyZero.

HP ZGX Nano G1n AI Station

The HP ZGX Nano G1n is a good choice for buyers planning to run demanding workloads for extended periods. If you need a Spark for sustained, continuous use, this is the model to consider.

That is because HP uses an active cooling system with two integrated fans. We recommend this design for users who want to perform sustained operations under load, such as using the Spark as the basis for an AI product or service with constant uptime.

The HP ZGX Nano G1n is available with 2 TB or 4 TB NVMe storage and ships with the NVIDIA DGX OS, but HP adds its own ZGX Toolkit, too, which gives you a set of tools for easier model downloads, inference, RAG and machine learning.

Buy the HP ZGX Nano G1n AI Station at BuyZero.

NVIDIA DGX Spark Founders Edition

The original NVIDIA DGX Spark is the platform’s reference model. However, please note that this model will soon be discontinued and therefore only very limited stocks remain. 

.

It comes with 128 GB of unified memory and 4 TB of storage. PNY-branded DGX Sparks are the same original NVIDIA hardware, not a separate OEM design.

This is the simplest choice for buyers who want the NVIDIA-branded machine rather than an OEM interpretation of the GB10 platform. Also, in our opinion, it looks the coolest. 

See the NVIDIA DGX Spark Founders Edition at BuyZero.

MSI EdgeXpert MS-C931

The MSI EdgeXpert is an extremely portable DGX Spark implementation, weighing around 1.2 kg and occupying roughly 1.19 litres of space (very small). It is an attractive option for anyone who wants to take an AI system to a customer site, laboratory or field location.

MSI offers 1 TB and 4 TB versions with the same 128 GB unified memory and 273 GB/s bandwidth as other GB10 systems.

Buy the MSI EdgeXpert at BuyZero.

ASUS Ascent GX10

ASUS calls its GB10 system the Ascent GX10. It uses the standard NVIDIA GB10 platform with 128 GB unified memory and is available in 1 TB, 2 TB and 4 TB storage configurations. It runs NVIDIA DGX OS and the same NVIDIA AI software stack found across the platform.

Pi3g is an ASUS IoT partner and supplies the Ascent GX10 through BuyZero.

The Ascent GX10 makes sense for buyers with an existing preference for ASUS hardware or support, or where the required ASUS configuration has better availability than competing GB10 models. 

.

Its AI performance should be broadly comparable to other 128 GB GB10 systems, so check the current price carefully before paying a premium.

Buy the ASUS Ascent GX10 at BuyZero.

Lenovo ThinkStation PGX

The Lenovo ThinkStation PGX is particularly interesting at the budget end of the 128 GB market. 

In October 2026, the 1 TB ThinkStation PGX without an operating system could be purchased at  €5,038 excluding VAT, making it the lowest-cost 128 GB GB10 system in our current list. However, prices are rising at present. 

There are 1 TB, 2 TB and 4 TB configurations available.

The lowest cost is the no-OS 1 TB model, which is attractive if you are comfortable configuring the machine yourself and your main goal is gaining access to the GB10 compute platform for the lowest possible price.

The trade-off is that you get less storage vs alternative models and must do more setup work than with many of the preconfigured alternatives.

See our current DGX Spark price comparison for Lenovo ThinkStation PGX pricing and availability.

Dell Pro Max with GB10

Dell’s Spark system is the Dell Pro Max with GB10, which is sold in 1 TB, 2 TB and 4 TB configurations and uses the same GB10 compute and 128 GB unified-memory architecture.

Our October pricing data put the 4 TB Dell configuration at €5,667 excluding VAT, making it competitive with the NVIDIA Founders Edition and MSI systems at the time of our pricing exercise.

For AI performance alone, there is little reason to expect the Dell GB10 to significantly outperform a cheaper system with the same processor and memory.

See our DGX Spark pricing guide for current Dell Pro Max with GB10 pricing.

Acer Veriton GN100

The Acer Veriton GN100 AI Mini Workstation is another 128 GB GB10 system and is available with up to 4 TB of storage.

Our October pricing data listed the 4 TB Acer at around €6,013 excluding VAT.

As with the Dell and Lenovo systems, the choice largely comes down to current pricing, storage needs, availability, and your preferred manufacturer, as AI performance is broadly similar.

Check our current DGX Spark price list for Acer availability.

What Can You Run On a DGX Spark?

The core Spark platform is rated for models up to roughly the 200 billion parameter class on one 128 GB Spark, with the exact limit depending on model architecture and quantisation. 

This means that two Sparks can support models up to roughly 405 billion parameters. 

We have sold Sparks to customers who use them for workloads including:

• Local LLM inference.
• Private AI assistants.
• RAG using company documents.
• Agentic AI applications.
• AI development and prototyping.
• LoRA and other fine-tuning workloads.
• Machine-learning development.
• Vision AI.
• Robotics development.
• Research and education.
• Testing applications locally before moving them to larger NVIDIA infrastructure.

The Spark’s 128 GB unified memory pool lets you load AI models that are too large to fit in the memory of many high-end workstations or gaming GPUs.

Our team has run the open-source Local Deep Research framework with Qwen 27B on DGX Spark hardware, creating a fully local research workflow rather than sending model inference to an external AI provider.

Read our DGX Spark Local Deep Research test.

Who Is the DGX Spark for?

A DGX Spark is suitable for anyone who wants to run LLMs locally, from hobbyists to multinational companies. 

Our customers who have bought the DGX Spark include:

• Developers who want large local models locally.
• Small companies building AI products.
• Private or sovereign AI deployments.
• RAG systems containing sensitive company data.
• Universities and research groups.
• Companies that want to do customer demonstrations of AI tools.
• Edge AI device manufacturers.
• Teams that want to reduce reliance on external AI APIs.

What Can a DGX Spark Not Do?

DGX Spark is a powerful desktop AI system, but it is not a miniature replacement for every GPU server. Its main limitation is its performance when running very large models. A 70B Q8 model may fit, for example, but our own testing of Llama 3.3 runs at around 3 tokens per second, which is not very fast.

The second limitation is model size. A single 128 GB Spark is aimed at models up to roughly 200 billion parameters, dependent on quantisation. Models beyond that point require multiple systems or another hardware platform such as the DGX Station.

The Spark is at its best when you want a large amount of AI-addressable memory in a very small local system.

Can You Connect Multiple DGX Sparks?

Yes. You can start with one DGX Spark machine, validate your workloads, then add another if you need greater model capacity.

Each 128 GB Spark includes NVIDIA ConnectX-7 networking with up to 200 Gbit/s bandwidth. Two systems can be connected to work with models up to around 405 billion parameters, depending on quantisation and model architecture.

However, there are limits to how far this approach makes economic sense. Linking more systems requires more complex networking, and going much beyond two can make the setup less attractive than moving to larger AI infrastructure. 

MSI officially supports a three-EdgeXpert configuration using the two ConnectX-7 interfaces on each machine. For a two-machine setup, remember that the required QSFP interconnect cable is not normally included with an individual Spark.

DGX Spark Buying Tip: Pi3g can provide the interconnect cables free on request when purchasing three DGX Sparks under our current multi-unit offer. See our DGX Spark discount guide.

DGX Spark vs DGX Station, Which Do You Really Need?

The DGX Spark is designed for individuals, developers, and small businesses running smaller AI workloads locally, with up to 128 GB of unified memory.

The DGX Station GB300, on the other hand, moves into a completely different performance class with 748 GB of coherent memory and is technically capable of running models up to the 1 trillion parameter class.

For an individual developer, small company, local RAG application or experimental AI project, Spark can give you far more practical compute than you need.

For a team working with very large models, demanding inference throughput, large-scale local AI or shared infrastructure, DGX Station is much more capable.

Read our full DGX Spark vs DGX Station comparison.

NVIDIA DGX Spark FAQs

Below are answers to questions we see again and again. 

How much does an NVIDIA DGX Spark cost? 

As of October 2026, the DGX Spark market covers a fairly wide price range.

The new 64 GB Spark starts at around €3,150 excluding VAT.

For 128 GB systems in our current pricing data, prices range from around €5,038 to €8,096 excluding VAT, depending on the manufacturer, SSD size, operating system and current supply.

See our continuously updated NVIDIA DGX Spark price guide for current pricing and lead times.

What about the 64 GB DGX Spark?

NVIDIA introduced a lower-cost DGX Spark with 64 GB of memory in October 2026, available through Pi3g from around €3,150 excluding VAT.

A 64 GB machine does not have the same model capacity as the 128 GB versions covered throughout most of this guide, so buyers should not treat them as interchangeable.

For smaller local models and AI development, the lower purchase price may make the 64 GB system attractive.

If your reason for buying Spark is the ability to load very large models locally, the extra memory in the 128 GB systems remains one of their main advantages.

Check our DGX Spark pricing page for the latest 64 GB availability.

Can you get a discount on a DGX Spark?

Yes.

There are three main routes available through Pi3g.

1. NVIDIA Inception discount. Eligible startups can apply for NVIDIA’s Inception hardware discount.

The program covers both DGX Spark Founders Edition systems and OEM GB10 devices. Pi3g can process the approved discount through the manufacturer and distributor when preparing your quote.

The hardware is intended for the startup’s own development, demonstration, testing, evaluation and proof-of-concept work rather than resale.

Read our full guide to NVIDIA DGX Spark Inception discounts.

2. NVIDIA Education discount

Qualified universities, research institutions and other eligible educational organisations can apply for NVIDIA education pricing.

Pi3g can submit the required organisational information through the distribution and manufacturer process, then provide a quote containing the approved discount.

The exact evidence required varies, but can include an institutional email address, university purchase order or research grant documentation.

3. Pi3g volume discounts for orders above ten DGX Sparks.

For projects starting with three units, we can currently include the required interconnect cables free on request using the offer described in our DGX Spark discount guide.

Is it possible to rent a DGX Spark?

Yes, you can rent a physical DGX Spark or rent a DGX Spark remotely. 

Learn more about that here. 

Where to Buy a DGX Spark? 

Pi3g sells the full range of DGX Spark and GB10 systems and can help compare models before you commit.

Many systems can be ordered directly through our sister site BuyZero’s AI hardware store.

If you already know the models and workloads you want to run, send them to us, and we can help determine which Spark configuration makes sense, whether you need multiple systems, or whether moving directly to a DGX Station would be a better fit.