How big models actually run

Clusters, explained without the hand-waving

No single GPU can hold a frontier model. So we combine many — but how you combine them is the whole game.

What a cluster is

A cluster, in plain words: multiple GPUs wired together with a fast interconnect so they behave like one much larger GPU, instead of several small separate ones. When a model is too big to fit on one card, you split it across the cluster — and the cluster's job is to make that splitting fast enough that it still feels like one machine.

1 GPU

80 GB

One H100. Fine for small models; nowhere near a frontier model.

1 DGX H100

640 GB

8 GPUs in one box, wired with NVLink. A real cluster in a single chassis.

GB200 NVL72

13,824 GB

72 GPUs as one NVLink domain — a whole rack acting like a single giant GPU.

The honest difference

This is the part most spec sheets won't tell you. Adding up GPU memory on paper is easy. Making those GPUs actually work together on one model is not. There are two very different things people call a "cluster":

A pile of desktop cards

Several GeForce RTX 5090s in separate machines, or crammed into one.

  • No NVLink fabric — cards talk over slow PCIe / ordinary network
  • Splitting one big model across them is painfully slow
  • Memory adds up on a spreadsheet, not in practice
  • Great for many small jobs; wrong tool for one huge model

A real datacenter cluster

DGX systems and GB200 racks, built for this from the ground up.

  • Purpose-built NVLink fabric between every GPU
  • Any GPU reads any other's memory at hundreds of GB/s
  • One model spans all GPUs and still runs fast
  • The combined memory is usable, not just theoretical

That's why every recommended build in this store uses real datacenter hardware — never a stack of desktop cards.

How you scale up

One DGX H100 gives you 640 GB. Need more? You network several DGX systems together. Need far more, as one seamless pool? You step up to a GB200 NVL72 rack, where all 72 GPUs already live in a single NVLink domain — no networking seams at all.

Every build, combined

When a build combines machines, here's the combined memory, power, and price.

Startup Build — DeepSeek V4 Pro

for DeepSeek V4 Pro

3 DGX H100 systems, networked together, give 1,920GB of pooled GPU memory — exactly enough headroom for DeepSeek V4 Pro's 1,920GB minimum.

3× DGX H100 Machines
1,920 GB Combined GPU memory
30,600 W Combined power draw
$1,125,000 Total price

Startup Build — GLM 5.2 Max

for GLM 5.2 Max

2 DGX H100 systems give 1,280GB of pooled memory, comfortably above GLM 5.2 Max's 903.6GB minimum, at the lowest real unit count that clears the bar.

2× DGX H100 Machines
1,280 GB Combined GPU memory
20,400 W Combined power draw
$750,000 Total price

Startup Build — Nemotron 3 Ultra

for Nemotron 3 Ultra

Nemotron 3 Ultra needs 660GB minimum — just over what a single DGX H100 (640GB) provides, so the cheapest safe build is 2 DGX H100 systems (1,280GB).

2× DGX H100 Machines
1,280 GB Combined GPU memory
20,400 W Combined power draw
$750,000 Total price

Mid-Size Growth Cluster — GB200 NVL72

One GB200 NVL72 rack provides 13,824GB of GPU memory in a single NVLink domain — enough to run all three models on this site simultaneously with room to spare for growth, higher traffic, or larger future models.

1× GB200 NVL72 Machines
13,824 GB Combined GPU memory
120,000 W Combined power draw
$2,750,000 Total price

Not sure how many machines you need?

Tell us your model and we'll size the cluster for you.

Help Me Choose