The fleet that is already paid for

A fleet of 2,400 prior-generation inference accelerators was bought three or four years ago and is fully depreciated. It runs inside a 1,200 kW envelope that is contracted, subscribed and full. Thirty-two accelerators sit in each of 75 air-cooled racks drawing 16 kW at the rack input. The accelerators account for 60 per cent of that draw. Every hour the fleet earns is close to free of capital cost. It also produces less output per watt than current silicon would in the same envelope. Somebody has to decide when it stops being worth running.

Where power is spare, a depreciated accelerator runs while it earns more than it costs. A machine earning a penny an hour passes that test forever. At a subscribed site the watt is the scarce input, so the test sharpens. The incumbent is retired when its margin per watt falls below what the replacement would earn in the same watt. Depreciation does not enter, because it is sunk. A fully depreciated fleet is the easier one to retire, which makes it the right subject for extending the window.

The runtime moves two terms and holds the third. Throughput is held, so revenue does not move. Energy cost and power fall together, because they are one physical quantity priced two ways. With energy at 30 per cent of operating margin, margin per watt rises by about 30 per cent on the derived route, modelled. That route carries the run's throughput behaviour inside it. The arithmetic route assumes throughput exactly held and gives about 35 per cent. We plan against the derived route and show the arithmetic route as the upper bound. They are one effect expressed twice and are never multiplied together.

Margin per watt converts into time. At an illustrative frontier of 25 per cent a year, that lift buys about 14 months on the derived route and 16 on the arithmetic route, modelled. Across a frontier of 10 to 50 per cent a year the deferral runs from about 8 months to a little under 3 years, modelled. The refresh still happens, about 14 to 16 months later at the illustrative 25 per cent frontier, and the runtime moves onto the new silicon calibrated for it. One precondition is hard: the retirement must be economic. Where a tenant names the part, the contract sets the retirement date, and the same reduction still releases 151 kW to 252 kW inside the envelope, modelled.

Inverted, the arithmetic says what a given deferral needs at the die. Twelve months at a 25 per cent frontier needs about 16 per cent. We measured 21, a margin of about 5 points and more than twelve times our re-derivation uncertainty, modelled. Twelve months at a 35 per cent frontier needs about 21 per cent, inside that uncertainty. At that frontier the reduction we measured buys about 10 months on the derived route we plan against and about 12 on the arithmetic route, modelled. One input separates those rows: the pace of the silicon you would refresh to. It is yours to look up in an afternoon.

Age does not set the saving; load shape does. Across our corpus the loaded power spreads show no trend by generation, and the saving rises with transition density rather than with a taller idle floor. Three things in your own telemetry decide it, in this order: the loaded power spread, the transition density, and the idle fraction with how far its floor can fall. A bursty fleet of any generation can carry all three.

The majority of the saving is idle-floor reduction, and most of the rest is clock-down at mid utilisation. The mid-utilisation channel returns more than twice as much per joule available. The load-bearing word in duty-cycle intelligence on bursty fleets is bursty.

Our coefficient sits at the accelerator die and your constraint sits at the rack input. Between them is your accelerator share, 60 per cent here. The conservative reading applies the 21 per cent to that share alone; the stated basis carries it through as a floor on wall watts. Both travel together. Fleet draw falls from 1,200 kW to about 1,049 kW or 948 kW, releasing 151 kW to 252 kW inside the same envelope. That is 85 to 94 racks, from 75, or 3 to 6 replacement racks at 40 kW each with the cooling work priced first, modelled. The capital argument needs none of this conversion.

The energy line is the smallest pool: about GBP 200,000 to GBP 330,000 a year at a 60 per cent mean draw and 25 pence a kilowatt-hour, modelled. Replacing 2,400 accelerators at a 3 to 1 throughput ratio takes 800 replacements, GBP 20m at an illustrative GBP 25,000 each. Deferring that spend by 14 to 16 months at the declared cost of capital is worth about GBP 2,500,000 to GBP 2,800,000 in present value, modelled. That is about 11 times the conservative energy line over the same window. Avoided capital moves spend from a capital line to an operating line, so ask your finance function early.

Seven days of read-only power and utilisation telemetry at 1 Hz answers the three deciding variables at once. The idle fraction shows at the left edge, transition density in the time series, and loaded spread in the height of the right-hand cloud. Thirty minutes on one accelerator settles which lever holds, a clock lock or a power limit. One column decides the trial: a throughput counter, because without it a saving cannot be told from a throughput loss. Bring that log, and the three-week validation puts your number beside ours.