CryptoGPUs
Special orderShips in 2–5 days
More from NVIDIA2 in AI workstations

NVIDIA DGX Spark 128GB

$5,331.51Was $5,612.12−5%

Black Friday price: you save $280.61 until Dec 1. No code needed.

Paid in Bitcoin, Monero, USDT or another coin you choose.

Free insured shipping. US delivery in about 5 to 11 business days. Shipping details

Overview

The DGX Spark is NVIDIA's own desktop GB10 system, a 150mm square box that pairs a 20-core Arm CPU with a Blackwell GPU over 128GB of coherent LPDDR5X. It ships with NVIDIA's DGX OS and the CUDA stack installed, so Ollama, vLLM and PyTorch run without any driver work.

Capacity is the reason to buy it. A 70B model at 4-bit needs roughly 40GB, which leaves plenty of room for long context, and NVIDIA rates a single unit for inference on models up to about 200B parameters at FP4. Fine-tuning is supported up to 70B parameters. Two units can be joined over the built-in ConnectX-7 200Gbps port to reach 256GB and models up to about 400B.

The trade-off is memory bandwidth. At 273GB/s, dense 70B models generate only a few tokens per second, while mixture-of-experts models feel much faster because they read fewer weights per token. This configuration carries up to 4TB of NVMe storage, a 240W power supply and a declared 29dB(A) noise level under full GPU load.

  • GB10 Grace Blackwell, 128GB LPDDR5X
  • 273GB/s unified memory bandwidth
  • 200Gbps ConnectX-7 clustering port
  • 150mm square, 1.2kg, 240W

Specifications

Performance
SuperchipNVIDIA GB10 Grace Blackwell Superchip, up to 1 petaFLOP FP4
Processor20-core Arm, 10 Cortex-X925 + 10 Cortex-A725
GraphicsBlackwell GPU, 5th-gen Tensor Cores, 4th-gen RT Cores
Memory128GB LPDDR5X, coherent unified system memory, 256-bit
Memory bandwidth273GB/s
Storage4TB NVMe M.2, self-encrypting (platform supports up to 4TB)
Local AI capacityInference up to about 200B parameters at FP4, fine-tuning up to 70B (NVIDIA figures)
ClusteringUp to 4 systems over ConnectX-7; 2 units give 256GB and about 400B parameters
Connectivity
Ethernet1x 10GbE RJ-45
High-speed networkNVIDIA ConnectX-7 NIC, 200Gbps
WirelessWi-Fi 7, Bluetooth 5.4
USB4x USB Type-C
Video output1x HDMI 2.1a, up to 3x DisplayPort over USB-C
Media engines1x NVENC, 1x NVDEC
Power and size
Power supply240W external adapter, GB10 TDP 140W
Dimensions150 x 150 x 50.5mm
Weight1.2kg
Noise29dB(A) sound pressure at max GPU load, 13dB(A) idle (NVIDIA declared values)
Software
Operating systemNVIDIA DGX OS with the NVIDIA AI software stack
Preinstalled runtimesOllama, vLLM, PyTorch with CUDA, NVIDIA Agent Toolkit, Nemotron models

Questions and answers

Still unsure? Visit the help center or message us.

Which local models fit in 128GB?

A 70B model at 4-bit uses about 40GB and runs with long context, and 100B to 120B class mixture-of-experts models fit at 4-bit. NVIDIA rates a single unit for inference on models up to about 200B parameters at FP4.

How fast is a dense 70B model?

Token generation is limited by the 273GB/s memory bandwidth, so expect single-digit tokens per second on dense 70B models. Mixture-of-experts models with a small active parameter count run noticeably faster.

Can I link two units together?

Yes. Every unit has a ConnectX-7 200Gbps port, and two systems connect directly with a QSFP cable to pool 256GB. NVIDIA says up to four systems can be clustered.

Is this the same as the partner GB10 boxes from Dell, HP or ASUS?

They share the GB10 chip and 128GB of memory, but chassis, storage size and software bundle differ by maker. This product is NVIDIA's own design with 4TB of storage.

What operating system does it run?

It ships with NVIDIA DGX OS and the AI software stack preinstalled, and NVIDIA publishes step-by-step playbooks for agents, fine-tuning and inference. NVIDIA lists DGX OS as the only operating system.

From the blog

All articles
Local AI

How much VRAM for local LLM work in 2026: a practical sizing guide

A practical 2026 sizing guide: add up weights, KV cache and overhead, check real GGUF sizes for current open models, and see what fits on 16GB to 512GB.

Local AI

Best computer for local AI: DGX Spark vs Mac Studio vs Strix Halo

NVIDIA DGX Spark, Mac Studio M5 Ultra and Ryzen AI Max+ 395 compared on memory, bandwidth, real benchmarks and price, plus when a GPU tower is the better buy.

Pairs well with

Picked to go with this product. Check sizes and connectors for your build before you order.