The power that is paid for either way

The phrase covers two contracts pointing in opposite directions. On the sell side an operator is paid for committed compute whether the customer uses the hours or not. A reduction in drawn power lands in gross margin. On the buy side it pays for a committed quantity of electricity or network capacity whether it draws it or not. Below that floor the bill is fixed, so the watts the runtime frees become room to sell more compute inside power already bought. This page works the buy side.

The operator holds a 12 MW facility envelope and enters a power usage effectiveness of 1.2, which resolves it to 10,000 kW at the rack input. Its racks draw 140 kW nominal and carry 72 accelerators, which take 72% of rack draw. The requirement is 11 MW: on whole racks, 78 racks, 5,616 accelerators and 10,920 kW. The envelope holds 71 racks today, 5,112 accelerators. The operator is 504 accelerators short, seven racks, and needs about 8% off at the rack input. None of that contains a number of ours.

Our coefficient sits at the accelerator die and the constraint at the rack input, so both readings travel together. Applied to the accelerator share alone, 21% at the die is about 15% at the rack input. Carried whole, on the stated basis, it stays at 21%. A managed rack then draws about 119 kW or about 111 kW, and the same 10,000 kW holds 84 or 90 racks. That is 6,048 to 6,480 accelerators, modelled, against a requirement of 5,616. The requirement fits on both readings, with 432 to 864 accelerators beyond it and 731 to 1,373 kW of headroom.

The margin is about 7 points on the conservative reading and about 13 on the stated basis. That is more than twenty times the measurement uncertainty carried to the rack input, and the fit survives at any accelerator share above about 40%. So the plane conversion holds the answer. The three-week validation on your own racks then measures the coefficient on this rack and on this fleet's day. The extra accelerators are hardware you still buy; the runtime creates the room.

On a committed contract the megawatts are already bought, so the energy line runs second. Credit goes to the busy-and-changing 45% of the day only; the idle and pinned bands get nothing. At a mean draw of 95 kW a rack the fleet averages 7,410 kW, and the reduction removes about 4,400 to 6,100 MWh a year, modelled. At the marginal cost this operator avoids below its commitment, entered as zero, that is GBP 0 a year. At an illustrative blended 20 pence a kilowatt-hour, the invoice average, it reads about GBP 880,000 to GBP 1,200,000 a year, modelled.

The gap between those columns is your contract, and only one of them is money. The capacity line and the energy line are alternative uses of the same watts and are never added. On this contract the choice is close to made, because the energy use of those watts is worth close to nothing. The value lands on the revenue line, as more work sold inside 12 MW already paid for. That is about 18% to 27% more accelerators, modelled: enough to close the operator's 504-accelerator shortfall, with room for 432 to 864 more as its sales grow.

Real charging structures put the marginal cost at or near zero across part or all of the range. The Georgia Power large-load schedule floors billing demand at the greatest of the contract minimum, 50% of contract capacity or 500 kW. It carries a ratchet over the preceding eleven months. In Great Britain the distribution capacity charge is levied on the agreed maximum import capacity. It falls only when that agreed figure is reduced, which is a contractual act.

The German network charge regulation rewards high full-load hours, energy divided by peak: sites above ten gigawatt-hours a year earn a reduced charge at 7,000, 7,500 or 8,000 hours. The runtime lowers energy and the model holds the peak, so full-load hours fall. This site sits at 5,944, below every threshold, and keeps its position. A site just above a threshold can cross it: one at 8,100 hours falls below 8,000, and its minimum share of the published charge moves from 10% to 15%. That site runs its own hours through the model, and its energy adviser confirms the position before the rack order.

A Great Britain reader will ask how the reduction reaches the capacity charge, and the answer is a trade. Route A banks it: reduce the agreed import capacity by the headroom, and the charge falls in proportion. You keep 5,616 accelerators with no room left. Route B fills it: keep the agreed capacity and put 432 more accelerators inside it on the conservative reading. The charge does not move. Route C holds it as reserve against the standing condition below. Choose one. Route A makes the runtime a condition of the connection agreement, so it comes after the mitigation.

Some accelerators fit only because the runtime reduces what they draw. If they are energised, removing the runtime puts the site over its commitment. The runtime fails open to full performance on a component fault, so the accelerators keep working. The site plans for it with one of three measures: an agreed curtailment right, reserved headroom, or a commitment set far enough below the modelled figure to absorb a total failure. Here 84 racks unmanaged draw 11,760 kW against 10,000 kW, an unplanned 1,760 kW, or 2,600 kW on the stated basis, and the plan protects the bill as well as the commitment.