Guided Path: Mid-Size Company

Goal: serve more users and more traffic than a single-model startup box can handle, with a clear path to add capacity later without starting over.

1. The model

2. The build

A single DGX H100 node has 640GB of NVLink-connected memory — 2.3x what this model needs at minimum. That headroom is the point: it's what lets many concurrent users hit the model at once without queueing, and leaves room to run a second model alongside it.

NVIDIA press image

1x NVIDIA DGX H100 node

GPU / systemNVIDIA DGX H100 (8x H100 node) × 1

Combined memory640 GB

Total price$350,000

⚡ 8,500 W sustained draw

🏠 = 7.08x an average home (~1,200W continuous)

🔋 = drains a 90 kWh EV battery in 10.59 hrs

I want this build →

3. Room to grow

When traffic outgrows one node, the honest next step is a second DGX H100 wired to the first over NVLink/InfiniBand — a real cluster (see the Clusters page), not more desktop cards crammed into a case.

NVIDIA press image

2x NVIDIA DGX H100 nodes (networked cluster)

GPU / systemNVIDIA DGX H100 (8x H100 node) × 2

Combined memory1,280 GB

Total price$700,000

⚡ 17,000 W sustained draw

🏠 = 14.17x an average home (~1,200W continuous)

🔋 = drains a 90 kWh EV battery in 5.29 hrs

I want this build →