Colocation, Power Contracts and Site Selection

TL;DR: Site selection is decided by three facts and none of them is about compute. Power density per rack, because a hall built for 10 to 15 kW cabinets cannot host a 120 kW rack no matter how much total capacity the building has. Liquid cooling capability, because above roughly 40 kW there is no air-cooled answer and retrofitting a facility loop into an operating hall is a construction project. And contracted capacity, because colocation bills reserved kilowatts whether or not they are drawn, and redundant A and B feeds mean reserving roughly twice the actual load. On top of those sits the lead time asymmetry that shapes every schedule: new utility supply takes years, facility fit-out takes months, and GPUs take weeks to months. Anyone planning backward from a hardware delivery date has the dependency inverted.

The three filters, applied in order

filter 1: power density per rack
  the question is not "how many megawatts does the building have"
  it is "how many kilowatts can one cabinet take, and how many such cabinets are in a row"
  a hall rated at 10 kW per rack with 5 MW total cannot host GB300 NVL72 at 120 kW+
  it can host, at most, 41 racks' worth of total power spread as 500 low-density cabinets
sanity: total capacity and per-rack capacity are different numbers and only the second one
        decides whether your hardware fits

filter 2: liquid cooling
  above roughly 40 kW per rack, air is not viable in a normal cabinet
  the site needs a facility water loop reaching the rows, or space and permission to add one
  retrofit into a live hall is construction: months, disruption, and often not permitted
  a rear-door heat exchanger is the intermediate step, taking a rack into the 40 to 70 kW
    range without a full direct-to-chip loop

filter 3: contracted capacity and its price
  colocation prices reserved kW, not consumed kWh, plus a separate energy charge
  redundancy means reserving both feeds: a 132 kW rack with A and B reserves about 264 kW
  so the contract line is roughly twice the number an engineer quotes from the spec sheet
sanity: the reserved-versus-drawn gap is the single most common surprise in a first colo
        contract, and it doubles the largest recurring line item

The lead-time asymmetry, which sets the schedule

what the asymmetry means for planning
  the utility conversation starts roughly two years before the first job runs
  the GPU order is placed late, deliberately, because holding hardware is expensive and the
    parts improve
  a team that orders compute first ends up storing it, and stored GPUs earn nothing
  at $2.5 per GPU-hour, 512 GPUs waiting three months is
    512 x 24 x 90 x 2.5 = about $2.76M of unrealized capacity
sanity: the cost of being early on compute is larger than almost any discount for ordering
        early, which is why the schedule is built backward from power availability

Build, colocate or rent

TCO: Buy vs Rent has the break-even arithmetic. The site-selection point is narrower: renting is the only option that does not require answering the three filters at all, 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? 30 to 40 degrees is the target for warm-water designs, and a hall that can only supply chilled water at 12 degrees is spending money on cooling you did not need.
• 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

Whether you know that the physical layer gates everything. A candidate who talks only about GPUs has never been near a deployment. The strong answer names per-rack density rather than building total, mentions liquid cooling as a filter rather than a feature, and states the lead-time asymmetry with the consequence that the power conversation starts years before the hardware order. If cost comes up, the reserved-versus-drawn distinction and the doubling from redundant feeds is the detail that shows real exposure to a contract.

Key takeaways

• Per-rack power density eliminates more sites than total building capacity, and they are different numbers.
• Above roughly 40 kW per rack liquid cooling is required; a rear-door heat exchanger covers the 40 to 70 kW middle ground.
• Colocation prices reserved capacity, and redundant A and B feeds reserve roughly twice the drawn load.
• Utility supply takes years, facility fit-out months, GPUs weeks to months, so schedules are built backward from power.
• Three months of 512 idle GPUs is about $2.76M at $2.5 per GPU-hour, which is why compute is ordered last rather than first.