Beyond the mid-size build

For DeepSeek V4 Pro (1,600B parameters, needs 1,920 GB of GPU memory) — bigger than a single DGX B200 (1,440 GB) can hold. Two options.

What a "cluster" means here

Definition: a cluster is two or more separate machines connected together and used as a single combined system.

Whenever a build below combines machines, we show the combined GPU memory, combined power, and combined price.

Option 1: 2x DGX B200

Combined GPU memory
2,880 GB
Total price
$1,030,000
Total power
28,600 W

2,880 GB combined vs. 1,920 GB needed. Two complete DGX B200 systems, each already NVLink-connected internally, joined together — a cluster of two integrated machines rather than 16 loose cards.

28,600 W ÷ 1,200 W ≈ 23.8 average homes' continuous draw. At full load for 24 hours (one day): 28,600 W × 24 ÷ 1,000 = 686.4 kWh ÷ 90 kWh (one EV battery) = 7.63x an EV battery's worth of energy per day.

Request a quote for 2x DGX B200

Option 2: 1x GB300 NVL72 rack

Combined GPU memory
20,736 GB
Total price
$3,900,000
Total power
132,000 W

20,736 GB combined vs. 1,920 GB needed — far more than required for this model alone. In a GB300 NVL72, all 72 GPUs are wired by an NVLink switch fabric and software sees one huge GPU with shared memory: a real cluster, not a pile of separately-addressed cards. Choose this if you expect to grow well past DeepSeek V4 Pro's requirement, run multiple models at once, or serve many simultaneous users.

132,000 W ÷ 1,200 W = 110 average homes' continuous draw. At full load for 24 hours (one day): 132,000 W × 24 ÷ 1,000 = 3,168 kWh ÷ 90 kWh (one EV battery) = 35.2x an EV battery's worth of energy per day.

Request a quote for the GB300 NVL72 rack
Compare the two: 2x DGX B200 costs $1,030,000 for 2,880 GB (358 $/GB). The GB300 NVL72 rack costs $3,900,000 for 20,736 GB (188 $/GB) — a far higher entry price, but a much lower cost per GB if you'll actually use the extra capacity.
Back to the mid-size build