Colocation, Power Contracts and Site Selection
A suitable site must deliver the required rack power, cooling and service capacity. Contract terms and facility lead times shape both cost and deployment dates.
TL;DR: Check per-rack capacity as well as the building total. Confirm that cooling, floor loading and electrical redundancy support the chosen system. Then read how the contract charges for reserved capacity and energy. A and B feeds provide redundancy; they do not automatically double consumption or billed capacity.
The three filters, applied in order
First check power at the rack and row. A building with 5 MW in total may still be unable to supply a cabinet that requires more than 120 kW. Dividing the building total by rack power gives only an aggregate capacity bound; it does not establish a usable layout.
Next check the cooling system the hardware requires. Direct-to-chip cooling may need facility water, compatible temperatures and fluids, sufficient flow and a supported CDU arrangement. Retrofitting these into an occupied hall can affect the schedule. There is no single rack-power threshold that determines every air or liquid design.
Finally, read the capacity and energy terms in the contract. In a full A/B design, each rack gets two independent power paths so that either can carry the whole load if the other fails, which means each path must support the failover load. For a 132 kW rack, that means two capable paths, not 264 kW of normal consumption. Whether redundant capacity is billed separately is a contractual question.
- Power at the rack and row
Can one cabinet position feed the rack, not just the building total?
- Cooling the hardware requires
Facility water, temperatures, flow and a supported CDU arrangement
- Contract terms
Redundant paths sized for failover, and how reserved versus used power is charged
The lead-time asymmetry, which sets the schedule
Ask for verified dates for utility supply, electrical work, cooling, networking, hardware delivery and commissioning. Any of them can become the critical path. Coordinate purchase dates with readiness instead of assuming that GPUs always arrive faster than facility work.
For a hypothetical delay-cost scenario:
512 GPUs x 24 hours/day x 90 days = 1,105,920 GPU-hours
at an assumed $2.50 per GPU-hour: $2,764,800
sanity: this is a valuation at the assumed rate, not automatically a rental invoice or lost revenue.
Use the cost model appropriate to the purchase: rental charges, financing and depreciation, or work postponed. Then compare that cost with the risk of hardware arriving too late.
Build, colocate or rent
| Option | Control | Time to first job | Cost shape | Where it fits |
|---|---|---|---|---|
| Own facility | Total | Years | Capital heavy, lowest marginal cost | Very large, long-lived, predictable demand |
| Colocation | Hardware yours, building theirs | Months, if a suitable hall exists | Reserved capacity plus energy plus hardware capital | Most organizations buying their own GPUs |
| Rent from a GPU cloud | None over the physical layer | Days | Per GPU-hour, no capital | Uncertain demand, bursty work, or a first deployment |
TCO: Buy vs Rent has the break-even arithmetic. The site-selection point is narrower: renting shifts physical facility operation to the provider, but you still need to evaluate service capacity and reliability, and that is often the honest reason to start there. An organization that has never run a hall should not discover the difference between total and per-rack capacity while GPUs are on a truck.
What to ask a colocation provider, in order
- What is the maximum kilowatts per rack, contiguous, in one row? Not the building total. Ask for a row, because a cluster wants adjacent racks.
- Is there facility water to the row, and at what supply temperature? Confirm the allowed supply range and water specifications with the equipment vendor, then assess cooling cost for the site’s climate.
- What is the floor loading limit? Dense racks are heavy, and this eliminates raised-floor halls more often than people expect.
- What is reserved versus billed, and what happens if we exceed? The answer determines the true recurring cost.
- What is the commissioning schedule and who does the acceptance? Schedules slip here without anyone announcing it, so ask who signs off and against what.
- What is the path to more capacity in the same hall? Growing in place is far cheaper than a second site, and availability of adjacent space is worth contracting for now.
What interviewers are listening for
Connect the hardware requirements to a site the organization can actually use. State per-rack density, cooling requirements and commissioning dependencies. For cost, distinguish installed redundant capacity, contracted capacity and metered energy instead of assuming that two feeds imply twice the bill.
Key takeaways
- Building capacity and per-rack capacity answer different questions.
- Match cooling to the system and facility specifications; a single power threshold does not determine every design.
- Check how the contract prices reserved capacity, energy and redundancy.
- Build the schedule around verified lead times and commissioning dependencies.
- Label delay-cost calculations with the assumed rate and whether they represent rental charges, opportunity cost or another measure.