The Spark cluster

Three DGX Sparks, 384GB pooled over ConnectX: the compact rack that punches up.

TECHNICAL DIFFICULTY: 2 OF 4 · SEAT AND GO

How hard is this to assemble? three boxes to cable together, but each is plug in and go

NVIDIA's answer to the Mac Studio: 128GB of coherent memory per box, and the interconnect to pool them. Three Sparks cost less than one 512GB Ultra and bring CUDA-native inference, which matters for the newest model drops that land on NVIDIA stacks first. The trade: noisier, hotter, and per-dollar memory is worse than Apple. Assembly is cabling, not carpentry.

WHO THIS BUILD IS FOR

Teams and developers who batch jobs, run agent fleets, or test against NVIDIA-first releases the day they ship. Anyone whose work needs 200GB+ of model weights and a machine room that fits under a desk.

ESTIMATED TOTAL

$19,497

prices refreshed weekly
last check: Oct 3, 2026

The parts

NVIDIA DGX Spark 128GB x3 128GB 273GB/s nvidia

128GB each, pooled with the built-in ConnectX networking. 384GB total, NVFP4 native.

street price via Newegg, checked Oct 3

$19,497

What it runs

GLM-5.3-Flash

4-bit is 160GB of weights in 384GB pooled: a real fit with context room. Its home-turf CUDA stack.

DeepSeek V4.1 Flash

3-bit is about 207GB in 384GB pooled: fits with room to spare.

Qwen3.8-Omni-Flash

4-bit is about 62GB of weights in 384GB pooled: fits easily. Voice, vision, and text in one model.

WHAT IT WILL NOT RUN

The 2.8T class: Kimi K3 needs 700GB at 2-bit. Tensor parallelism across three boxes adds latency in single-user chat; batch jobs hide it, conversation does not.

POWER, NOISE, AND THE ROOM IT LIVES IN

About 720W for the trio at load, datacenter-adjacent loud. A closet with a door or a garage shelf is the right home. The install is two Ethernet cables and power; the tuning is software.

NEXT: THE $30k TIER

The question that keeps showing up on X: I have 20 to 30 thousand dollars, what do I buy? One answer is a single 512GB box that runs everything. The other is a rack of NVIDIA that trades silence for CUDA-native speed. This tier is where those two answers live, priced and compared.

See the $30k tier →