The request lands the same way in every facility: we’re out of power, time to talk about the next build. New switchgear, new cooling, maybe a new hall. What rarely gets asked first is the cheaper question — are we actually out of power, or have we just run out of the number on the design sheet?
Those are not the same thing. Most data centers said to be “at capacity” are not physically full. They are running well under what was provisioned, carrying power and cooling headroom that is built, cooled, and financed but never drawn. That gap is stranded capacity, and it is manufactured — layer by layer, out of conservative assumptions nobody ever goes back to measure.
Three cushions, stacked on one rack
Stranded capacity isn’t an accident. It’s the predictable sum of three margins, each reasonable on its own.
It starts with the nameplate. A device’s plate rating is the manufacturer’s worst-case ceiling, not its working draw — and real load rarely approaches it. Then comes the breaker: continuous loads are sized to no more than 80% of the circuit rating, so the usable number is already below the plate. Then comes design diversity — the assumption that every rack in the row could peak at the same moment, so each is provisioned as if it will. Stack those three cushions, repeat them down the row, and you exhaust the design capacity of the hall long before you reach its physical limit.
Every one of those margins was the safe call at design time. The problem is that none of them is ever reconciled against what the floor actually does in operation. Worst case gets provisioned once and then carried forever.
You can’t reclaim a cushion you can’t see
Reclaiming that headroom is usually pitched as a modeling exercise: feed better diversity numbers into the DCIM tool and the capacity reappears. But the numbers are the problem. If your finest power reading stops at the room or a floor PDU — or at a “dumb” rack PDU that reports nothing at all — then the rack-level diversity factor stays a guess, and the spreadsheet’s worst case survives by default.
This is where the measurement has to get granular. AKCP traces draw with a full Power Train model from the mainline down to the rack, reading voltage, current, power, power factor and accumulated kWh at the individual outlet. Now the diversity factor is observed, not assumed — you see what the row really pulls under real workloads, which is almost always well below the provisioned sum. You cannot release a cushion you were never able to measure.
Power and thermal headroom, measured together
A rack is bounded by whichever runs out first: amps or degrees. Power telemetry alone will hand back capacity the cooling can’t actually support. So the thermal side has to be measured at the same resolution — per-rack inlet/outlet ΔT and live heat maps show how much cooling headroom a cabinet genuinely has, rather than giving the whole row the margin you sized for its hottest position. Real-time PUE ties the two pictures together so efficiency tracks the load instead of lagging a quarter behind it.
With both sides measured, Capacity AI reads the live picture — power draw and thermal headroom together — and recommends where the next rack, or even the next server, actually fits. The recommendation is checked against real headroom, not nameplate math carried forward from design. That’s the line between an opinion about your floor and a fact about it: harness AI opinions with sensor facts.
The cheapest megawatt is the one you already paid for
This used to be an efficiency argument. It’s now a procurement one. New utility power is no longer something you simply order — it’s queued, and in a growing number of markets it’s newly taxed and politically contested. Against that backdrop, the megawatt you reclaim from your own stranded capacity is the fastest one you can add and the cheapest one you’ll ever own. It needs no interconnection, no permit, no new concrete.
AI density only sharpens the case. A GPU rack swings between near-idle and full draw far harder than the legacy load these provisioning rules were written for, so nameplate-based math strands even more of the floor than it used to — exactly where new power is most expensive to add.
Stranded capacity stays the most expensive real estate you own: fully built, fully cooled, fully financed, and producing nothing. Reclaiming even part of it defers the next build, the next cooling unit, the next lease — and now it sidesteps the queue and the tax as well. That’s the difference between optimizing against reality and managing a facility by worst-case assumption.
So before the next capacity meeting: what diversity factor are you actually provisioning against — and is it a number you measured this quarter, or one a spreadsheet picked at design and nobody has questioned since?