A colocation invoice rarely matches the power a tenant actually used. That is not an error. In most contracts the tenant is billed against a committed capacity, and the gap between what was committed and what was drawn is one of the largest avoidable costs in hosted infrastructure.
For GPU and ASIC deployments, where nameplate figures are large and real-world draw is variable, that gap can be substantial.
The commitment is the bill
When a tenant takes space, they generally contract for an amount of power expressed in kilowatts. The facility then reserves that capacity: electrical distribution, cooling, and a share of the redundancy infrastructure are all allocated against it. Those reservations are real, and they cost the operator money whether or not the tenant draws the power.
So the commitment is typically billed in full. Consumption may be charged on top, or folded in, depending on the structure; our breakdown of colocation pricing models covers the common variants. In every variant, the committed figure is the floor.
This surprises tenants who assume that under-using capacity will reduce the bill. It will not. Under-use is simply capacity that was reserved and not consumed.
Nameplate is not draw
The first cause of over-commitment is sizing against nameplate ratings. A device's rated power is the maximum its power supply can deliver, not what it pulls in normal operation. Summing nameplate figures across a deployment produces a number that the equipment will essentially never reach simultaneously.
GPU servers make this especially pronounced. Draw varies enormously with workload: a training run that saturates every accelerator looks very different from an inference fleet with variable utilisation, which looks different again from an idle cluster awaiting scheduling. ASIC miners are steadier, but even there the figure that matters is measured draw under the firmware and tuning actually in use, not the datasheet.
The gap between the sum of nameplate ratings and realistic simultaneous draw is the diversity factor, and it is where most over-commitment originates.
Stranded capacity is paid-for air
Capacity that is contracted, reserved and never used is stranded. The tenant pays for it monthly. The operator cannot sell it to anyone else, because it is allocated.
A deployment that commits well above its genuine requirement pays that difference every month for the life of the term. Over a multi-year contract on a meaningful footprint, that is a significant sum transferred for nothing. It is also invisible on an invoice, because the invoice shows the committed figure and not what it could have been.
The opposite error is real too. Committing too tightly and then needing more power means renegotiating mid-term, from a weak position, in a facility that may have sold its spare capacity to someone else.
Ratchets, and why they outlive the spike
Many agreements include a ratchet: a clause that resets the billable minimum upward if actual draw exceeds the commitment, and holds it there for a defined period.
The logic is sound from the operator's side. A tenant who exceeds their reservation is consuming capacity that was not allocated to them, and the facility must now plan around the higher figure. But the practical effect for a tenant is that a short excursion can raise the floor for months afterwards.
This matters most during commissioning and burn-in, when a new deployment is being stress-tested and may briefly draw far more than its steady-state figure. A ratchet triggered during commissioning can set a billing floor based on a condition that never recurs in production. Where that risk exists, it is worth negotiating a commissioning exclusion explicitly rather than hoping the clause is applied generously.
Right-sizing the commitment
The goal is a commitment close to realistic peak draw, with deliberate rather than accidental headroom.
- Measure, do not sum. Where possible, base the figure on metered draw from equivalent equipment running a representative workload, not on datasheets.
- Separate steady state from commissioning. They are different numbers and should be handled differently in the contract.
- Price the headroom. Growth room has a monthly cost. That may be entirely worth paying for a roadmap you are confident in, and pure waste for one you are not.
- Ask how additional capacity is released. A contractual right to expand is worth far more than an informal assurance that there is space.
- Understand the measurement. Which meter, what interval, and is it per rack, per circuit or per room?
Our capacity planning guide for AI workloads goes further into forecasting the demand side, and the contract negotiation checklist covers where these terms sit in a wider agreement.
Density changes the constraint
Committed power is only half the picture. The other half is how that power is distributed, because a facility can have capacity available in aggregate and still be unable to deliver it where a tenant needs it.
High-density GPU racks concentrate load in a way older halls were never designed for. A room rated for a comfortable average per rack may not support a handful of racks drawing several times that figure, even if the total sits within the room's envelope. The limiting factor becomes the busway, the branch circuits, or the ability to remove heat from that specific row rather than the building's incoming supply.
This is why a commitment should be expressed with its distribution: how many kilowatts, across how many racks, at what per-rack ceiling. A commitment stated only as a total leaves the most important constraint unwritten.
Term length and the cost of being wrong
Commitments are usually made for multi-year terms, which means a sizing decision made with incomplete information is carried for the length of that term.
That argues for doing the measurement work before signing rather than after, and for structuring growth as a right to expand rather than as capacity reserved upfront. It also argues for honesty about the roadmap. A deployment plan that assumes a second phase which has not been funded is a plan that produces an over-commitment, and the monthly cost of that optimism is paid for years.
Questions worth asking before signing
- Is billing based on committed capacity, measured consumption, or both?
- Is there a ratchet, what triggers it, and how long does it persist?
- Does commissioning or burn-in count towards the ratchet?
- At what point in the rack or circuit is draw measured, and over what interval?
- What is the process and timeline for increasing committed capacity mid-term?
- Is unused committed capacity creditable or transferable in any circumstance?
The practical point
The headline rate gets negotiated hard and the committed figure often does not, even though the commitment multiplies the rate. Getting the commitment right is usually worth more than winning a small concession on price.
Rax Data and Energy sizes tenant capacity against measured draw rather than nameplate arithmetic, because over-committed customers are not good long-term customers. If you are planning a GPU or ASIC deployment and want help arriving at a realistic figure, get in touch.