A B300 server is 8 GPUs, so 4,096 GPUs is 512 machines. Every figure below is built from that, using prices listed on our own inventory today.
The hardware
A new B300 HGX node lists at $563,000, so 512 of them is $288,256,000. Refurbished lines have sat at $530,000, which would take the same cluster to $271,360,000, a $16.9m difference for hardware that does the same work. That spread is why condition is a commercial question and not a technical one.
Power and space
At roughly 14 kW per node, 512 nodes draw about 7.2 MW before any cooling overhead. At the North American wholesale average of about $196 per kW per month, that is $1,405,404 a month, or $50,594,544 across 3 years. In a constrained market at $400 per kW the same footprint costs $103,254,000, which is a $52m swing driven entirely by where you put it.
At 14 kW per node you are fitting perhaps four nodes to a rack, so this is a 128-rack deployment and a genuine multi-megawatt facility conversation, not a colocation order form.
The fabric
Every node needs transceivers at both ends of every link, and 800G class optics run about $1,000 each. Budget roughly $16,000 per node for optics and its share of switching, which is $8,192,000 across the cluster. It is the line most commonly missing from a first-pass budget, and it is bigger than most people guess by an order of magnitude.
Getting it there
At 512 crated nodes this is freight on the critical path, not a shipping afterthought. Duties, import VAT and handling vary by destination and are frequently a seven-figure line on a cluster this size. If you have no registered entity in the destination country you also need an Importer of Record, and arranging one late is a common reason hardware waits at a border while the facility bill runs.
The shape of it
Hardware $288,256,000. Colocation and power $50,594,544 over 3 years in North America. Optics and fabric $8,192,000. Freight, duties and commissioning on top, varying by route. The accelerators are about 83% of the bill, which is why they get the attention, and the other 17% is where the schedule risk lives.
For comparison, renting the same 4,096 GPUs at $4.00 per GPU-hour for 3 years is $430,571,520, with no asset at the end.
We build this table against a real destination, a real density and a real utilisation before anyone commits. Price the hardware on the live inventory, the facility side on colocation and power, or run your own numbers with the cluster cost calculator.