Headroom put back to work

The site is at its envelope, and there is demand the operator cannot serve. A smaller bill is not what it wants. It wants more work out of the same building, in the currency NVIDIA has named: compute per megawatt inside a fixed power budget. Reducing what accelerators draw creates room, and this operator fills it with racks rather than banking it. That is a growth decision: consumption rises, work rises by more, and energy per unit of work falls. The scope is an installed fleet with no cooling change and no rack change.

The inputs are the operator's: a 6,800 kW envelope at the rack input, 48 racks of 72 accelerators at 140 kW, and a target of 54. Our coefficient is at the die and the constraint is at the rack input, so the conversion is written out. It is the die figure, times the accelerator share, times an upstream cascade factor defaulted to one. Both readings travel together. The conservative reading, 21 per cent times a 72 per cent share, is about 15 per cent at the rack input, modelled. The stated basis takes the die figure as a floor on wall watts, 21 per cent, our physics position.

Fifty-four racks draw 7,560 kW unmanaged, so the target needs a reduction of about 10 per cent to sit inside 6,800 kW. The conservative reading clears that by about 5 points and the stated basis by about 11, modelled. The target survives at any accelerator share of about 48 per cent or more.

Within the installed fleet, with no cooling change and no rack change, the envelope holds 57 racks on the conservative reading and 61 on the stated basis, modelled. That is about 19 to 27 per cent more than the installed 48. At 57 racks the conservative margin is a third of a point, so the plan sizes to 54 and keeps three racks of margin in hand. The 432 accelerators added at the target are hardware the operator buys.

Everything here is modelled. The day shape puts 55 per cent of hours in the busy-and-changing band the coefficient is credited in. The per-rack reduction on the energy line is therefore about 8 per cent conservative and about 12 per cent on the stated basis. Before absorption, 48 unmanaged racks consume about 40,000 MWh a year. After it, 54 managed racks consume about 41,200 MWh on the conservative reading and about 39,700 MWh on the stated basis. That is about 1,300 MWh a year more, or roughly flat.

At the illustrative tariff that is about GBP 250,000 a year more on the conservative reading, and roughly flat on the stated basis, modelled, because 432 more accelerators draw power. Energy per unit of work falls about 8 to 12 per cent, the per-rack reduction seen from the other end. The freed watts have three mutually exclusive uses. Fill the envelope with racks. Bank it as energy: about 3,300 MWh and about GBP 660,000 a year at 48 racks, conservative. Or hold it as resilience margin. Choose one and plan against it.

Racks that fit only because the runtime is reducing what they draw make continued operation a standing condition of the configuration. The runtime is designed to fail open to full performance on a component fault, so the accelerators keep serving, and the site agrees its cover for a total failure. On a total runtime failure the target of 54 racks draws 760 kW over the envelope, modelled. At 57 racks it is 1,180 kW over, and at 61 racks 1,740 kW over.

Absorption turns headroom into racks, so the cover for a total runtime failure is agreed before the first order. Three covers exist at the site: an agreed curtailment right, reserved headroom, or fewer racks energised. Sizing to 54 rather than 57 keeps about 5 points of margin over the reduction the target needs.

Every step spends margin, modelled, on the conservative reading. Fifty racks need about 3 per cent and fit by about 12 points. Fifty-four need about 10 and fit by about 5. Fifty-seven fit by a third of a point, and 58 do not fit. On the stated basis 61 racks fit by under a point and 62 do not. The ratios hold at ten times the scale and the whole-rack counts do not. A 68 MW envelope holds 485 racks unmanaged and absorbs to 540. That adds 3,960 accelerators, modelled, for about 8,400 MWh a year more on the conservative reading, with 7,600 kW to cover on a total runtime failure, a site engineering decision.

Per accelerator, the efficiency ratio is up to +22% tokens per watt in Balanced mode. It is derived from NVML die power and vLLM serving throughput on the same 48-hour H100 NVL run. On a fixed fleet the runtime holds throughput, 0% change in the run, derived, and cuts power. More output comes through the capacity released, and that is rack arithmetic, modelled: one rack in eight more at the target, about 19 to 27 per cent more at the ceiling. Size the target on work rather than racks.

The value is the gross margin on 432 more accelerators of work, net of the electricity above, in an envelope that could not otherwise hold them. Your margin per accelerator-hour and your delivered cost complete it, and we run that arithmetic with you before the order.

Under absorption the cost of being wrong is racks in a hall the site cannot power, so the sequence matters. Run the model on today's day shape and on the one you expect afterwards. Validate for three weeks on your own racks, with the rack plane measured alongside NVML. Then order in stages and measure between them.