DGX Station vs DGX Spark: Which NVIDIA Desktop AI System Do You Need?

DGX Spark and DGX Station share a name and a software stack, but they are not two sizes of the same product. Spark is a compact developer appliance that is priced like a high-end laptop and runs from a 240W power supply. Station is a deskside server with data center silicon, a dedicated 20A circuit requirement and a price that, as of September 2026, is quoted in six figures. The right choice depends on the size of the models you need to run, how fast you need them to run, and whether a conventional RTX PRO workstation would do the job for less. This guide sets out the specs, current pricing and availability, and where each one fits.

What is the difference between DGX Spark and DGX Station?

DGX Spark uses the GB10 Grace Blackwell superchip: a 20-core Arm CPU and a Blackwell GPU sharing one pool of LPDDR5X memory. NVIDIA lists 128GB or 64GB of unified memory at 273 GB/s, up to 1 petaflop of FP4 AI performance, up to 4TB of NVMe storage, a ConnectX-7 200Gb/s port for linking units, and a 240W power supply in a box 150mm square and 50.5mm tall (NVIDIA). DGX Station uses the much larger GB300 Desktop Superchip: a 72-core Grace CPU with 496GB of LPDDR5X at 396 GB/s, joined over 900 GB/s NVLink-C2C to a Blackwell Ultra GPU with 252GB of HBM3e at 7.1 TB/s. NVIDIA quotes 748GB of coherent memory in total, a ConnectX-8 SuperNIC at up to 800Gb/s, support for seven MIG instances, and a 1,600W system power figure that requires a 20A circuit (NVIDIA).

• Memory: Spark 64GB or 128GB unified LPDDR5X; Station 252GB HBM3e plus 496GB LPDDR5X.
• Memory bandwidth to the GPU: Spark 273 GB/s; Station 7.1 TB/s on HBM3e.
• AI compute: Spark up to 1 PFLOP FP4; Station up to 20 PFLOPS FP4, as quoted by HP for its GB300 system (StorageReview).
• Networking: Spark ConnectX-7 at 200Gb/s; Station ConnectX-8 at up to 800Gb/s.
• Power: Spark 240W supply; Station 1,600W on a 20A circuit.

What models can each one run?

NVIDIA rates the 128GB Spark for models up to about 200 billion parameters and the 64GB version for about 100 billion, with two linked 128GB units reaching 256GB and models of around 400 billion parameters (NVIDIA). Those figures assume heavily quantized weights (FP4), and fitting a model is not the same as running it quickly. Token generation for large language models is usually limited by memory bandwidth, and 273 GB/s is a fraction of what a discrete workstation GPU offers. Spark is best treated as a development target: build and test pipelines, fine-tune smaller models, and validate code that will later run on Blackwell servers. Station is a different class. With 748GB of coherent memory and HBM3e bandwidth on the GPU side, NVIDIA positions it for running models of up to a trillion parameters locally (NVIDIA), and its MIG support lets one Station be split between several developers. Our guide to why memory bandwidth limits inference explains why bandwidth, not capacity alone, sets the speed.

How much does DGX Spark cost in 2026?

NVIDIA raised the DGX Spark Founders Edition MSRP from $3,999 to $4,699 in late February 2026, citing industry-wide memory supply constraints (NVIDIA Developer Forums). On October 2, 2026, NVIDIA announced a 64GB version starting at $4,999 from Acer, ASUS, Dell, Gigabyte, HP and MSI, available from October 23. Two 64GB units can be cabled together through their ConnectX-7 ports to pool 128GB, and NVIDIA says the pair delivers up to 1.7x the performance of one system (NVIDIA). The uncomfortable detail is that the new 64GB model starts above the current 128GB Founders Edition price, which tells you how much of the bill is now memory. OEM GB10 systems, such as the Dell Pro Max with GB10, HP ZGX Nano, Lenovo ThinkStation PGX and ASUS Ascent GX10, use the same chip and differ mainly in storage, chassis, support terms and price.

How much does DGX Station cost, and can you buy one now?

NVIDIA does not sell DGX Station itself; it is built by OEM partners. ServeTheHome reported in March 2026 that ASUS, Dell, Gigabyte, HP, MSI and Supermicro were offering systems, that the top of the price range it heard was $120,000 to $125,000, and that some configurations would land around $100,000 (ServeTheHome). Published list prices are rare. One exception is Dell's UK store, which lists a Dell Pro Max with GB300 configured with 16TB of NVMe and an RTX PRO 2000 display card at £151,306.95 excluding VAT (Dell UK). HP announced its ZGX Fury as orderable on September 9, 2026, in tower or 5U rack form with 2TB or 4TB of storage, without disclosing a price (StorageReview). Treat any number you see as a reported figure for a specific configuration and region, and get a written quote with a stated validity period.

Which DGX Station versions and OEM models are there?

• GB300 Linux systems: the core product from ASUS, Dell, Gigabyte, HP, MSI and Supermicro, shipping with Ubuntu and NVIDIA's AI developer tools.
• DGX Station for Windows: announced in June 2026 with the same GB300 hardware and the same OEM partners, scheduled for Q4 2026 (ServeTheHome).
• GB200-based variant: Supermicro offers a developer system with the older B200 GPU and 186GB of HBM3e, which ServeTheHome notes has stronger FP64 performance than B300 and may suit HPC users better.
• A 252GB note: early announcements quoted 288GB of HBM3e for the Station GPU; shipping specifications list 252GB, so check any older material you are using for sizing.

Is an RTX PRO 6000 workstation a better buy than either?

For many teams, yes. The RTX PRO 6000 Blackwell Workstation Edition carries 96GB of ECC GDDR7 at 1,792 GB/s with a 600W maximum board power, and supports up to four MIG instances, as does the Max-Q version (NVIDIA). One card has less memory than a 128GB Spark but more than six times its bandwidth, so models that fit run considerably faster. Two cards give 192GB, and a four-card tower reaches 384GB while running Windows or Linux, standard x86 applications and ISV-certified software. What a workstation does not give you is the Grace CPU's coherent memory or the Station's 748GB in one address space. Our guides to RTX PRO 6000 price and availability and how much VRAM you need go deeper.

Who should buy which?

• DGX Spark: individual developers and small teams building agents, testing pipelines that will deploy to Blackwell servers, or fine-tuning small models, where portability and the NVIDIA software stack matter more than speed.
• RTX PRO workstation: teams that need faster inference and fine-tuning on models that fit in 96GB to 384GB, need Windows or ISV applications, or want to upgrade GPUs over time.
• DGX Station: groups that must run very large models locally for data-control reasons, want to share one system through MIG, and have the power, cooling and budget for a deskside server.
• None of the above: if usage is occasional, renting cloud GPUs by the hour is usually cheaper. Our local vs cloud cost framework shows how to run the numbers.

What should you check before ordering?

For a Station, confirm that the room has a dedicated 20A circuit and can remove about 1.6kW of heat, that noise is acceptable for the location, and whether you want the tower or a rack-mount version. Ask the OEM for the exact storage, display GPU and support term in the quote, since those change the price substantially. For Spark, decide between 64GB and 128GB based on the largest model you expect to run, and remember that clustering two units adds a second purchase and a cable. In both cases, memory is the part of the bill moving fastest in 2026, so read our AI hardware shortage playbook before you lock a budget, and see whether DGX is worth it versus HGX if you are weighing a server instead.

Nexus Compute supplies new RTX PRO Blackwell GPUs, including the RTX PRO 6000 Max-Q, and complete AI workstations, sourced through authorized distribution with full manufacturer warranty. We check compatibility against your platform before dispatch, quote within 48 business hours, and offer purchase orders and net terms for qualified organizations. If a DGX Spark or Station is the better fit for your workload, we will tell you.