Clusters, explained
A cluster is several machines wired together and used as one combined pool of memory and compute. When a build uses more than one machine, we always show the combined memory, combined power, and combined price — but "combined" doesn't always mean "fast."
Real datacenter clustering
Inside an HGX B200 server, the 8 GPUs are connected with NVLink, a direct chip-to-chip link built for this job. Between servers or racks, datacenters use InfiniBand. Both move data at hundreds of gigabytes per second — fast enough that many GPUs can effectively act like one giant GPU with one giant memory pool. The GB300 NVL72 rack takes this to the extreme: all 72 GPUs are NVLink-connected across the whole rack.
Stacking desktop cards over PCIe
Our Small Startup Build uses 4 × RTX PRO 6000 Blackwell cards. Add up their memory and you get 384 GB, same as any other 384 GB figure on paper. But these cards have no NVLink. They only talk to each other over the PCIe slot they're plugged into — far slower than NVLink, and not built for GPUs to share memory as one pool. The memory total is honest arithmetic; the speed and flexibility of a true cluster is not there.
Same memory total, much slower connection, no NVLink at all on the consumer and workstation cards. That's the honest difference.
Combined totals, side by side
| Small Startup Build (PCIe-only) | Mid-Size Build (NVLink) | |
|---|---|---|
| Memory | 384 GB | 1,440 GB |
| Power | 2,400 W | 14,300 W |
| Price | $53,000 | $450,000 |
| NVLink? | No | Yes |