Clusters
Multi-Machine Builds
| Build | Combined VRAM | Combined Power | Combined Price | Real-life power |
|---|---|---|---|---|
| 2x DGX H100 | 1,280GB | 20,400W | $400,000 | 17.00 homes · 0.2267 EV charges/hr |
| 4x DGX H100 | 2,560GB | 40,800W | $800,000 | 34.00 homes · 0.4533 EV charges/hr |
| 2x GB200 NVL72 racks (superpod) | 27,648GB | 240,000W | $6,000,000 | 200.00 homes · 2.6667 EV charges/hr |
Even a Single Rack Is Already a Cluster
The GB200 NVL72 rack isn't one GPU — it's 72 Blackwell GPUs networked together with NVLink inside one rack, already acting as a cluster before you even connect a second rack.
Datacenter Clustering vs. Desktop Cards Piled Up — the Honest Difference
Datacenter clustering
DGX/HGX servers and NVL72 racks use NVLink and NVSwitch to connect GPUs with hundreds of GB/s of bandwidth, so their memory pools together into one giant logical GPU. Multiple servers then connect over InfiniBand networking built specifically for this. It's engineered, tested, and supported as a single system — with real cooling and power design to match.
Desktop cards piled up
Four RTX 4090s in one PC are still four separate 24GB cards — the 4090 has no NVLink, so there's no fast memory pooling. Software can split a model across them over the much slower PCIe bus, but it's a workaround, not a true 96GB GPU. You'll also hit consumer power supply, motherboard slot, and cooling limits long before you hit datacenter scale.