Nemotron 3 Ultra
NVIDIA's flagship open-weight model, tuned for enterprise agent workloads, released under NVIDIA's Open Model Development Warranty license.
The memory math
550 billion parameters
× 1 GB per billion
× 1.2 (20% working room for activations / KV cache)
= 660.0 GB minimum GPU memory
Any hardware combination below this number can't even load
Nemotron 3 Ultra, let alone answer a question — the model's weights literally do not fit.
Cheapest working build
9x NVIDIA H100 SXM Cluster of 9
| Combined GPU memory | 720 GB (needed 660.0 GB) |
|---|---|
| Combined power draw | 6,300 W |
| Total price | $288,000 |
5.2x
an average home's power draw
6.3 kWh
used per hour of running
0.07
EV battery charges per hour (90 kWh each)
Why this is a cluster, not one box: a single
NVIDIA H100 SXM only has 80 GB of memory — not
enough on its own. We wire 9 of them together over NVLink so they share the
load like one giant GPU. See how clusters work.
Turnkey alternative
Prefer fewer boxes to manage? 2x NVIDIA DGX H100 (8x H100) gives you 1,280 GB for $700,000 — more expensive per GB because you're also paying for the integrated CPUs, networking, and support that come with a full system, not just raw GPUs.
| Combined power draw | 20,400 W (17.0x an average home) |
|---|
Why not a pile of RTX 4090s? On paper,
28 desktop cards would add up to enough memory. In
practice, RTX 4090s only talk to each other over ordinary PCIe/Ethernet — 10–50x slower
than the NVLink used in datacenter hardware. The GPUs spend most of their time waiting on each
other instead of computing, so a model this size becomes impractically slow. That's why every
build above uses datacenter-grade hardware.