Silicon on the dock

Accelerators arrive on a date the buyer does not set. Power arrives later still. When the first lands ahead of the second, the operator holds a paid-for asset consuming its own earning window. Depreciation waits for the in-service date, so the loss shows on nobody's report. The earning window is fixed at the far end, so the gap costs days from the highest-earning part of the asset's life. Operators count that gap as dock-to-live, and this page counts in the same unit.

Twenty-four rack-scale, liquid-cooled racks are on site, 1,728 accelerators in all. Each is NVIDIA's published GB300 NVL72 configuration: 72 accelerators and 36 Grace CPUs in one NVLink domain. Nominal draw at the rack input is 140 kW, from published OEM specifications of 132 to 142 kW, with a peak of about 155 kW. The accelerators account for about 101 kW of that, a 72% share. That share is your field to set, and the largest lever in the model.

The envelope energised today is 2,800 kW of IT load at the rack input; the full envelope, when it lands, is 3,500 kW. Unmanaged, the fleet draws 3,360 kW, so the shortfall is 560 kW. The gap to full energisation is yours to supply and is illustrated at 8, 16 and 26 weeks. Our coefficient was measured on H100 NVL, so the GB300 rack stays modelled until a baseline run on it: one engine, calibrated per architecture.

Our coefficient sits at the accelerator die and your constraint sits at the rack input. The reduction at your rack is our die figure multiplied by your accelerator share. The upstream cascade factor stays at one, giving our own physics argument no credit. The conservative reading applies 21% to the 72% share: about 15% at the rack, a managed rack drawing about 119 kW instead of 140 kW. The stated basis reads the die figure as a floor on wall watts: 21% at the rack, about 111 kW. Both are modelled and neither travels alone.

Without the runtime the 2,800 kW envelope holds 20 racks, 1,440 accelerators, exactly full with no headroom. On the conservative reading it holds 23 racks, 1,656 accelerators. On the stated basis it holds 25, capped at the 24 delivered. Three of the four waiting racks stop waiting on the conservative reading, 216 accelerators, and the fourth does not. On the stated basis all four do, 288 accelerators. The share of the fleet earning now moves from about 83% to about 96%, or to 100%, modelled. Plan against three racks; the fourth is what the validation is for.

Every rack that stops waiting does so for the whole gap, so the recovered quantity is rack-weeks. What falls is fleet-average dock-to-live. At a 16-week gap, four racks wait without the runtime and one waits with it. Two weeks of dock-to-live come off every delivered rack and 48 rack-weeks come back, modelled. At 8 weeks the figures are one week and 24 rack-weeks; at 26 weeks, about three weeks and 78.

The connection date, the equipment delivery and the commissioning schedule stay exactly where they were. What moves is the date each admitted rack starts earning. It moves by software installation rather than by construction.

Every money input here is a third party's published figure, read at source on 20 August 2026. Across four published lists and utilisations of 50% to 90%, a waiting rack forgoes about GBP 2,500 to GBP 20,000 of contribution a day. At one published rack-scale figure and 70% utilisation, the anchor is about GBP 9,300. Three recovered racks over a 16-week gap is 336 rack-days: about GBP 3 million at the anchor, inside about GBP 850,000 to GBP 7 million, modelled.

The band is gross value at published list figures. To make it your figure, replace list with your realised revenue per accelerator-hour, take off the running cost of the newly lit racks (close to zero where the envelope is already contracted), and add any contractual remedies avoided on pre-sold capacity. Leave the depreciation charge out: the revenue is what would have paid it.

Adding 560 kW of envelope is measured in months at best. The Ofgem July 2026 connections consultation assumes GBP 9,500,000 per MW of project capital. On that basis the work is about GBP 5 million one-off, modelled. The connection commitment proposed in the same consultation would hold about GBP 133,000 to GBP 399,000 of capital until energisation. None of it acts inside the gap. Temporary or behind-the-meter generation adds an envelope rather than changing what the work needs, so it composes with a runtime. Splitting across sites forks the clock, since each fragment arrives with its own energisation date. Slipping the date onto a customer moves the loss and can carry contractual remedies.

The energy bill does not fall. Before, 20 racks draw 2,800 kW inside a 2,800 kW envelope. After, 23 racks draw about 2,733 kW, so the site draws about the same and does about 15% more work, modelled. The run gave up to +22% tokens per watt, derived, with throughput unchanged. That ratio is the same tokens for less energy; the extra output here is rack-count arithmetic. At full energisation the 3,500 kW envelope holds 29 to 31 managed racks against 25 unmanaged, 5 to 7 beyond the 24 delivered, modelled, counted once.

Racks energised by virtue of the reduction depend on it. Remove the software and the site is back over its limit, with a reserve of 67 kW against three racks of unmanaged draw. The runtime fails open to full performance on a component fault, which protects the workload and returns the racks to full draw, so the site plans for that event. The mitigation lives at the site: reserved headroom, an agreed curtailment right, or a commitment set far enough below the modelled figure to absorb a total failure.