The fleet this is for

The fit is strongest when seven things hold, and six can be checked without us. You own or contract the accelerators and decide what runs on them. You hold your own site or electrical envelope. You are commercially sold out. Your fleet is bursty rather than saturated. You hold the electricity account, or a tariff that charges for the marginal kilowatt-hour. You can give a service the authority to set accelerator clocks on your estate. And your binding limit is a contractual figure measured over a settlement interval, not the rating of installed plant. Four have their own studies where they read differently: a tenant in someone else's hall (The tenant and the landlord), an estate that sells no compute (The corporate estate), a committed power contract (The power that is paid for either way) and a limit set by cooling plant (The heat that cannot leave).

Four are public: your commercial model, site tenure, published availability and billing basis. One needs a week of your own telemetry, read only, as three day-shape bands: busy and changing, idle, and pinned. One is a question for your facility engineer, and one for your change-control and information-security functions. The mechanism is duty-cycle intelligence on bursty fleets. The majority of the saving is idle-floor reduction and most of the rest is clock-down at mid utilisation, so a week of read-only telemetry shows how much of your day it works on; a fully saturated fleet is a grid-flexibility conversation instead.

The operator holds a 7,200 kW envelope, IT load at the rack input, and needs 8 MW of racks. Each rack draws 140 kW nominal and carries 72 accelerators. The requirement rounds down to 57 whole racks, 4,104 accelerators, drawing 7,980 kW. Fitting that needs about a 10 per cent reduction at the rack input, modelled. Today the envelope holds 51 racks, 3,672 accelerators, so the operator is 6 racks and 432 accelerators short. That shortfall contains nothing of ours: it is your own requirement against your own envelope.

Our coefficient is at the accelerator die; the constraint is at the rack input. Between them sit the accelerators' share of rack draw, declared at 72 per cent, and an upstream cascade factor defaulted to one. Both readings travel together. The conservative reading applies the die figure to the accelerator share alone: about 15 per cent, modelled. The stated basis treats a watt not drawn at the die as a watt not converted, distributed or cooled: 21 per cent, modelled. This page plans against the conservative reading.

On the conservative reading a managed rack draws about 119 kW. The 57-rack requirement draws about 6,770 kW and sits inside 7,200 kW with about 5 percentage points to spare, modelled. The envelope now holds 60 racks, 4,320 accelerators. The shortfall closes, and 3 racks and 216 accelerators sit beyond the requirement, with 427 kW of headroom. On the stated basis it holds 65 racks and 4,680 accelerators: 8 racks and 576 accelerators beyond the requirement, 896 kW of headroom and about 11 points to spare, modelled.

Two things travel with those figures. ADAPT, a per-GPU software runtime, creates the room, and the operator fills it with accelerators on its own order schedule. And the margin holds. About 5 points is about eighteen times our own measurement uncertainty at that plane, and the fit survives at any accelerator share of about 47 per cent or more. The arithmetic settles the plane conversion. The operator's own racks then measure the die-level figure on its own load shape.

The freed watts have three uses that exclude each other: fill them with accelerators, bank them as energy, or hold them as margin against the standing condition below. The fleet's mean draw is 5,415 kW at a declared 95 kW per rack. Only the busy-and-changing share of the day, 65 per cent, carries the coefficient; the idle and pinned bands are credited at zero. Continuous demand avoided is about 530 kW conservative and about 740 kW on the stated basis, modelled.

Over 8,760 hours that is about 4,700 to 6,500 MWh a year, modelled. At an illustrative 20 pence a kilowatt-hour, your own tariff, it is about GBP 930,000 to GBP 1,300,000 a year, modelled. The energy line is conservative by construction: the model credits only busy-and-changing hours, while the measured run put the majority of recovered energy in the idle floor, so idle hours are upside the validation measures.

A tenancy that fits because the runtime is reducing rack draw keeps the runtime as a standing condition of the configuration, and the site plans its cover from the start. The runtime fails open on a component fault by design, so the accelerators keep serving. On the conservative reading, modelled, 60 energised racks draw 8,400 kW unmanaged against a 7,200 kW envelope: 1,200 kW to cover on a total runtime failure. The site covers it with an agreed curtailment right, reserved headroom, or fewer racks energised, and planning on 57 rather than 60 brings the draw to cover down to 780 kW.

Plan on 57 to 59 racks on the conservative reading, with margin in hand, modelled. At 60 racks that reading still fits, by under a point; from 61 to 65 the room rests on the stated basis, and the validation on your own racks measures the reduction they deliver. At 66 racks, which need about 22 per cent, the route is a larger envelope.

Duty cycle sets the saving, so the three-week validation on your own racks, at your own plane, turns this arithmetic into your own rack count.