The gigawatt programme

The programme holds a contracted envelope of 1,200 MW; at a power usage effectiveness of 1.20, 1,000 MW reaches the rack input. Its requirement is 1,100 MW of racks. At 140 kW nominal and 72 accelerators per rack, that is 7,857 racks and 565,704 accelerators. Without the runtime the envelope holds 7,142 racks and 514,224 accelerators. On its own arithmetic the programme is 715 racks and 51,480 accelerators short, and needs about a 9% reduction at the rack input.

Our coefficient is at the accelerator die and the programme's constraint is at the rack input. Between them sit the accelerators' share of rack draw, entered as 72%, and an upstream cascade factor defaulted to one. Two readings follow, and both travel together. The conservative reading, 21% of 72%, is about 15% at the rack input. The stated basis treats the die figure as a floor on wall watts and gives 21%. Both are modelled. A managed rack then draws about 119 kW or about 111 kW against 140 kW nominal. The 7,857 racks draw about 934 MW or about 869 MW, inside 1,000 MW on either reading with about 6 to 12 percentage points to spare.

Inside the same 1,000 MW the envelope holds 8,415 racks on the conservative reading and 9,041 on the stated basis, against 7,142 unmanaged. That is 605,880 to 650,952 accelerators, modelled. It meets the requirement of 565,704 with 40,176 to 85,248 accelerators beyond it, hardware the programme would still buy. It is about 18% to 27% more racks in the same contracted envelope. The headroom is about 66 MW to about 131 MW, modelled: a mid-sized data centre rather than slack.

Per accelerator, the same run gave up to +22% tokens per watt in Balanced mode, derived. On a fixed fleet that is the same tokens for less energy. More output comes from the capacity released, about 18% to 27% more racks in the same envelope, modelled.

The mechanism is duty-cycle intelligence on bursty fleets. Most of the saving is idle-floor reduction and most of the rest is clock-down at mid utilisation, so a week of each hall's telemetry shows the hours it works on; a fully saturated hall is a grid-flexibility conversation instead. In this model the busy-and-changing 55% of the declared day carries the coefficient. On a mean draw of 95 kW per rack the fleet averages about 746 MW. The continuous demand avoided is about 62 MW to about 86 MW: about 540,000 to 760,000 MWh a year, modelled.

At your own illustrative 20 pence a kilowatt-hour, that energy is worth about GBP 110,000,000 to about GBP 150,000,000 a year, modelled on the measured H100 NVL run until a paired measurement per hall type makes it the programme's own. The freed watts have three uses, and they exclude each other: fill them with accelerators, bank them as energy, or hold them as resilience margin. Choose one and plan against it; adding two counts the same watts twice.

The programme is about to request 1,100 MW. The runtime removes about 166 MW from that request on the conservative reading and 231 MW on the stated basis, modelled. Capital not spent is megawatts not requested times the cost per megawatt of new capacity. The second term belongs to your own capital plan, so multiply it yourself and you will know which assumption you have made. Between capacity needed on 1 June 2029 and an own build energising on 1 December 2031 sit 30 months. It is a deferral rather than an avoidance, because an alternative expansion exists.

With 900 MW committed today and 90 MW a year of growth, the headroom is used up in 37 months on the conservative reading and 48 on the stated basis, modelled. Your capacity plan proceeds as it would; the envelope you already hold yields more while it does.

Racks that fit only because the runtime reduces 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. The mitigation therefore lives with the programme's engineers rather than in the software. On the conservative reading, 8,415 energised racks draw about 1,178 MW unmanaged against a 1,000 MW envelope. That is about 178 MW of unplanned draw on a total failure, modelled. At that size the programme's engineers and the system operator set the cover per hall: an agreed curtailment right, reserved headroom, or a request sized below the modelled figure. A curtailment right keeps the capacity figures above; reserve held inside the envelope spends them, and the safety case sets the balance.

A capacity planner's first question decides how this applies. Is the binding limit a contractual demand or energy figure measured over an interval, or the rating of installed plant? A sustained reduction relieves the first, and the arithmetic holds as written. It does not relieve a transformer, a breaker or a busway, which is still sized for the peak the racks can draw. Here the limit is contractual over a 30-minute settlement interval, and pinned runs last at most 6 minutes. The envelope is exceeded in those minutes without setting the interval, so the fit stands. At 45 minutes the verdict turns conditional.

A gigawatt programme is a set of halls, filled on a schedule, running different workloads with different provisioning bases and limits. The model takes one of each. Run it once per hall with that hall's own inputs, then add the rack counts, which are integers and add cleanly. A percentage of one hall is not a percentage of a programme. The coefficient is the one thing that aggregates, because it belongs to accelerators and a workload rather than to a building.

One paired measurement per hall type settles the modelled half, and it belongs before a capacity request is sized rather than after.