What five megawatts of Blackwell actually asks of a data centre
2,048 GPUs across 256 nodes. The GPU order is the easy part. The facility conversation is where the deployment actually lives or dies.
Somebody sends you a requirement: 5.2 megawatts. You send it to four facilities. Three come back with a price. One of them is 28% cheaper. You are about to make an expensive mistake.
IT load is the power drawn by the equipment in the rack — the servers, the switches, the storage. It is the number the hardware vendor gives you, because it is the only number they have any visibility into.
Total facility power is IT load plus everything the building does to keep that equipment alive: cooling, UPS losses, power distribution losses, lighting, and the rest of the mechanical and electrical plant. The ratio between the two is Power Usage Effectiveness — PUE.
A good modern facility might run a PUE of 1.25 to 1.4. Which means a 5.2 MW IT load is roughly 6.5 to 7.3 MW of total facility power. Or, coming the other way, a facility quoting you on 5.2 MW of total power is only offering you about 3.7 to 4.2 MW of usable IT capacity.
That is not a rounding error. That is a fifth to a quarter of your deployment.
Partly because everyone assumes their own convention is the obvious one. A hardware person means IT load. A facilities person often means total power, because that is what the utility bills them for. Neither of them thinks they are being ambiguous.
And partly, in some cases, because ambiguity is commercially convenient. A quote that looks 25% cheaper than the competition wins the shortlist, and the conversation about what was actually being priced happens after you are emotionally committed.
Put it in the RFI as a question the operator has to answer in writing, before any pricing is discussed:
Those five questions, answered in writing before anyone quotes, will do more for your budget than any amount of negotiation afterwards.
Once you have settled the power basis, the same discipline applies to everything else the headline rate quietly excludes. Cross-connects. IP transit. Remote hands, which has a way of appearing at month three at a rate nobody mentioned at signature. Escalators. And the one that costs the most in the long run: contiguous expansion rights, which are cheap to secure on day one and close to impossible to retrofit when you need to double.
A rate of "$170 per kW" is not a price. It is an opening position with an unknown number of line items behind it.
2,048 GPUs across 256 nodes. The GPU order is the easy part. The facility conversation is where the deployment actually lives or dies.
If a supplier will not put the chain of custody in writing, there is a reason. Here is how to find out in one email rather than three months.
Node count, GPU class, kW per rack, target market, in-service date. That is enough for us to open the conversation with the right suppliers and the right facilities. Under NDA from first contact.