The heat that cannot leave

The site can import more power than it can reject the heat from. On a design-day ambient the plant derates, and the hall holds a lower load for a window of hours. ADAPT reduces the accelerator's own draw, and a watt not drawn at the rack is a watt of heat the plant does not reject. The arithmetic runs on the IT load your hall carries at the rack input on the day that binds you. Your facility engineer holds that number, from your plant's derate curve and commissioning data.

Declared for this site: a design-day ceiling of 5,600 kW of IT load at the rack input, a rack of 72 accelerators drawing 140 kW nominal, and a requirement of 6,300 kW. That requirement is 45 racks and 3,240 accelerators. The ceiling holds 40 racks and 2,880 accelerators today, so the site is 5 racks and 360 accelerators short on its design day, modelled. Closing the gap takes about an 11% reduction at the rack input, a figure with no coefficient of ours in it.

Is your binding limit the rating of installed plant, or a sustained condition measured over a window? A chiller, pump, heat exchanger or coolant distribution unit is sized for the peak heat the racks can make. A sustained reduction relieves a plant rating only where that rating is enforced on an averaged basis, and some are. Your facility engineer knows which yours is.

The model runs both answers. Where your plant limit is enforced over a sustained window, here 60 minutes, the full rack arithmetic runs. Where it is an instantaneous rating, the value is the energy line, with the cooling-side saving your facility engineer adds on top. The engineer confirms which applies, and the instrument supplies no default for this input.

Our coefficient is at the accelerator die and your ceiling is at the rack input. Between them sit the accelerator share of rack draw, declared at 72%, and an upstream cascade factor defaulted to one. Two readings travel together, both modelled. The conservative reading applies the die figure to the accelerator share alone: about 15% at the rack input. The stated basis gives 21%, treating a watt not drawn at the die as a watt not converted, distributed or cooled. That is a physics argument, and the validation measures it.

On the 60-minute window, a managed rack draws about 119 kW on the conservative reading and about 111 kW on the stated basis, modelled. The 5,600 kW ceiling then holds 47 to 50 racks, 3,384 to 3,600 accelerators, against 40 unmanaged. The requirement of 45 is met with 2 to 5 racks to spare, and 253 to 623 kW of headroom remains under the ceiling, modelled. The margin over the 11% needed is about 4 to 10 points, and the fit survives at any accelerator share of about 53% or above.

The declared design day is 50% busy and changing, 30% idle and 20% pinned, and the longest pinned run is 90 minutes. In pinned minutes the model takes the fleet at its full 6,300 kW against a 5,600 kW ceiling, with no reduction credited. The 47 to 50 racks hold across the busy-and-changing hours. On this declared design day the longest pinned run outlasts the 60-minute window, so your own longest run is the next figure to pull: under the window, the verdict clears. A longer window or a smaller pinned fraction clears it too.

The energy line runs under either answer. With a declared mean draw of 95 kW per rack, the 45-rack fleet averages about 4,275 kW at the rack input. The reduction applies to the busy-and-changing half of the day. The continuous demand avoided is about 323 to 449 kW, modelled: about 2,800 to 3,900 MWh a year over 8,760 hours. At an illustrative 20 pence a kilowatt-hour, yours to change, that is about GBP 570,000 to GBP 790,000 a year, modelled.

Those figures stop at the rack input. A watt not drawn at the rack is a watt of heat your plant does not have to reject, and the energy spent rejecting it adds to the line above. Your facility engineer can size that from your plant's coefficient of performance in about an hour. Same accelerators, same tokens, less energy: 0% throughput change in the run and up to +22% tokens per watt, both derived, in Balanced mode.

A site energising racks that fit under its ceiling on the strength of the reduction has made continued operation a condition of the configuration. The runtime is designed to fail open to full performance on a component fault. On the conservative reading, 47 racks unmanaged draw about 6,580 kW against a 5,600 kW ceiling. A total runtime failure therefore puts 980 kW above the ceiling, modelled, arriving over a time your hall's thermal mass sets. The site plan covers it with reserved thermal margin, an agreed curtailment right, or fewer racks energised.

Your facility engineer sets the balance: a curtailment right keeps the rack count, while reserved margin and fewer racks cover it by giving racks back, down to the 40 the ceiling holds today for full cover.

The mechanism is duty-cycle intelligence on bursty fleets, so a fleet average is the starting point for your own number. A three-week validation on your racks, measuring the rack plane alongside NVML on your load, settles the modelled half. Here it covers the design-day condition, or is extrapolated to it on terms agreed before it starts. Your facility engineer's hour on the cooling side settles the rest.