How much does the NVIDIA DGX Spark cost?

NVIDIA priced DGX Spark from $3,999 USD at launch. OEM variants from Acer, ASUS, Dell, HP, Lenovo and MSI sit in the same range, typically $3,999–$4,499 depending on storage.

DGX Spark vs RTX 5090 — which is better for AI?

An RTX 5090 has higher raw FP16 throughput but only 32 GB of VRAM. DGX Spark gives you 128 GB of unified memory, which means you can load 70B–200B parameter models that simply will not fit on a 5090. For training tiny models or gaming the 5090 wins; for serving and fine-tuning large LLMs locally, DGX Spark wins.

Can DGX Spark fine-tune large language models?

Yes. With 128 GB of unified memory and full CUDA / NCCL support you can LoRA-fine-tune models up to roughly 70B parameters out of the box, and chain two DGX Sparks together over the built-in ConnectX-7 link to fine-tune up to ~405B-class models.

Is DGX Spark the same as 'RTX Spark'?

Most people searching for 'NVIDIA RTX Spark' mean DGX Spark — the personal AI supercomputer announced as Project DIGITS and shipped as DGX Spark. It ships with a Blackwell GPU from the same silicon family as RTX 50-series, but it is branded DGX, not RTX.

NVIDIA DGX Spark Review 2026 — The RTX-Powered Personal AI Supercomputer

Q: What is NVIDIA DGX Spark?

NVIDIA DGX Spark is a desktop-sized personal AI supercomputer built on the GB10 Grace Blackwell Superchip. It pairs a Blackwell-generation GPU with a 20-core Arm CPU and 128 GB of unified LPDDR5X memory, so you can run, fine-tune and serve LLMs up to ~200B parameters locally.

What is NVIDIA DGX Spark?

DGX Spark is the productised version of Project DIGITS, NVIDIA's personal AI supercomputer announced at CES 2025 and shipping through 2026. It puts a full DGX-class software stack — CUDA, NCCL, NIM microservices, NeMo — onto a desktop unit you can plug into a standard wall socket. Under the hood is the GB10 Grace Blackwell Superchip: a Blackwell-generation GPU fused to a 20-core Arm CPU over NVLink-C2C, sharing 128 GB of unified LPDDR5X memory.

DGX Spark specs at a glance

GPU: NVIDIA Blackwell, 5th-gen Tensor Cores, FP4 support
CPU: 20-core Arm (10× Cortex-X925 + 10× Cortex-A725), via NVIDIA Grace
Memory: 128 GB unified LPDDR5X, ~273 GB/s
Storage: 4 TB NVMe SSD
Networking: ConnectX-7 SmartNIC (200 Gb/s) for chaining two units
Compute: ~1 PFLOP of FP4 AI performance
OS: DGX OS (Ubuntu-based) with NVIDIA AI Enterprise
Price: from $3,999 USD

DGX Spark vs RTX 5090 vs Mac Studio

The honest comparison: an RTX 5090beats DGX Spark on raw FP16 throughput and gaming, but tops out at 32 GB of VRAM. A Mac Studio M3 Ultra matches DGX Spark on unified memory (up to 512 GB) and uses less power, but its AI software ecosystem outside Apple's own MLX framework is years behind CUDA. DGX Spark's pitch is the combination: 128 GB of GPU-addressable memory plus the full NVIDIA AI stack the rest of the industry already targets.

If you mostly run inference on 7B–13B models, an RTX 5090 is faster and cheaper. If you want to load Llama 3.1 70B, Mixtral 8×22B, DeepSeek-V3 or fine-tune them locally without renting a GPU, DGX Spark is the only thing in its price bracket that lets you do it.

Who is DGX Spark for?

Solo ML researchers who want a workstation that can hold a 70B-parameter model in memory.
AI startup founders doing local fine-tuning before they pay cloud bills.
Enterprise developers who need an air-gapped dev box for sensitive data.
Educators and labs teaching modern LLM workflows on real hardware.

It is not for gamers, not for video editing, and not a replacement for a multi-H100 cluster if you're training a foundation model from scratch.

Software: what comes pre-installed

DGX Spark ships with DGX OS, NVIDIA's Ubuntu derivative, plus the full NVIDIA AI Enterprise stack: CUDA 13, cuDNN, TensorRT-LLM, NIM microservices, NeMo for fine-tuning, and Blueprints for common agent and RAG patterns. Anything you build on DGX Spark moves bit-for-bit to a cloud DGX node — same drivers, same runtime, same containers.

Chaining two DGX Sparks

The built-in ConnectX-7 SmartNIC lets you link two DGX Spark units at 200 Gb/s, giving you 256 GB of unified memory and enough headroom to fine-tune ~405B-parameter class models. For most solo developers a single unit is the right answer; two is the upgrade path before you hit cloud GPUs.

Bottom line

DGX Spark is the first time "personal AI supercomputer" stops being a marketing phrase. If your job involves running, fine-tuning, or serving large open-source LLMs locally and you don't want to manage a cloud bill, $3,999 for 128 GB of GPU-addressable memory and the full NVIDIA stack is hard to beat in 2026.

NVIDIA DGX Spark Review 2026: The RTX-Powered Personal AI Supercomputer