Every colocation contract has a number on it. Somewhere in the schedule there is a figure in kilowatts, and that figure is what you are paying for whether you draw it or not. The question nobody asks until it is urgent is how long that number will last, and the answer is almost always known months before anybody acts on it.
Running out of contracted capacity is not a technical failure. It is a forecasting failure, and the lead time on the remedy is measured in months while the lead time on the problem is measured in weeks.
Contracted, installed, and drawn are three different numbers
The first source of confusion is vocabulary, and it is worth settling before any forecast is attempted.
Contracted capacity is what you have committed to pay for. Installed capacity is what the distribution actually supports at your racks, which may be less if the build was staged. Drawn capacity is what your equipment is pulling right now, and billed capacity may be none of those three, depending on whether your agreement bills on commitment, on measured peak, or on some blend of the two.
Forecasting against the wrong one of these produces a confident answer to a question nobody asked. The distinction between what you reserved and what you use is covered in detail in our guide to contracted capacity versus actual draw, and the billing consequences depend on which pricing model your contract uses.
For capacity planning, the number that matters is drawn against installed. That ratio is your real headroom.
Measure at the right place, and at the right resolution
A monthly figure from an invoice is not a forecast input. It is a lagging summary that hides exactly the behaviour you need to see.
What you want is per-circuit or per-rack data at intervals short enough to show peaks: fifteen minutes is generally sufficient, one minute is better if you can get it. The reason is that capacity is constrained by peak, not by average, and a hall averaging sixty per cent can still trip a branch circuit during a synchronised workload start.
Three figures are worth tracking per circuit. The rolling peak over the last thirty days, which is what your protection actually sees. The 95th percentile, which strips out one-off anomalies and is a fairer basis for trend. And the ratio of peak to average, which tells you how spiky the load is and therefore how much headroom the peak demands.
A site with a peak-to-average ratio of 1.1 can be run much closer to its limit than one at 1.6, and the two look identical on a monthly bill.
AI load does not grow the way enterprise load grows
Traditional enterprise capacity planning assumes reasonably smooth growth: servers get added, utilisation creeps up, and a straight line through twelve months of history is a defensible projection.
GPU workloads do not behave like that. Power draw is a function of what is running, and the difference between an idle cluster and the same cluster under a sustained training run is very large. A fleet that has not changed physically can move substantially in drawn power because the job mix changed, which means hardware count is not a proxy for power in the way it used to be.
The consequence for forecasting is that you need two curves rather than one. A fleet curve, projecting installed hardware, which is relatively predictable because you control procurement. And a utilisation curve, projecting how hard that hardware will be driven, which is a business question rather than a technical one.
Multiply them and you get a range, not a line. The range is the honest answer. Our note on power density trends covers why the per-rack figure has been moving as fast as the fleet count.
The headroom you cannot use
Nameplate headroom overstates what is available, usually by more than people expect, and for several compounding reasons.
Redundancy consumes it first. In an N+1 or 2N arrangement, the whole point is that a path can be lost without dropping load, which means neither path may be loaded to a point where the survivor cannot carry the total. Our comparison of N+1 versus 2N redundancy sets out what that costs in usable capacity.
Breaker derating takes more. Continuous loads are conventionally limited to eighty per cent of breaker rating, so a 32 A circuit is a 25.6 A circuit for planning purposes.
Phase imbalance takes more again. Three-phase distribution is only as good as its worst phase, and a hall that is balanced on paper drifts as equipment is added one unit at a time.
And the cooling plant has its own ceiling, which may bind before the electrical one does. A hall with electrical headroom and no cooling headroom has no headroom.
Realistic usable capacity is frequently sixty to seventy per cent of the nameplate figure. Planning against nameplate is how sites run out while the dashboard still shows room.
Working backwards from lead times
The trigger point for action is not a utilisation percentage. It is a date, derived from how long the remedy takes.
Adding capacity in an existing hall where the distribution is already installed may take weeks. Adding capacity that requires new distribution, or a new room, or an upstream change at the substation, takes considerably longer, and transformer and switchgear lead times have been the binding constraint on a great many projects rather than money or will.
So the calculation runs the other way round. Establish the lead time for the remedy you would actually need. Add contingency, because lead times quoted at enquiry are not lead times delivered. Then find the date on your forecast where projected draw crosses usable capacity, and subtract. That is your decision date, and it is usually a great deal sooner than people expect.
What good practice looks like in a month
The discipline is modest and it is monthly rather than annual.
Pull per-circuit peak and 95th percentile for the period. Update the fleet curve with anything procured or decommissioned. Revisit the utilisation assumption with whoever owns the workload roadmap, because that is the input most likely to have changed without anyone telling the facilities side. Recompute usable capacity rather than assuming it is static, since redundancy posture and phase balance both move. Then restate the crossover date and compare it to the decision date.
If the two have converged, the conversation is now, not next quarter.
Signing the next increment
When the decision date arrives, the useful question is not simply how many more megawatts. It is how the increment is structured.
Staged capacity with defined step dates is usually better than a single large commitment taken early, because it matches payment to use. Expansion rights over adjacent space are worth more than they appear, since capacity you can reach without a migration is categorically cheaper than capacity you cannot. And a right of first refusal on neighbouring capacity costs little to negotiate at signature and is unobtainable later.
What you are buying is optionality, and optionality is cheap to acquire in advance and expensive to acquire under pressure. That is the whole argument for forecasting at all.
Staged Capacity Against a Real Timeline
Rax operates where transmission investment has run ahead of load. Talk to our team about sizing a deployment against your forecast rather than against a brochure.
Contact Rax Data & Energy