The site that is full

An operator runs one site, and the site is at its envelope. The sales team quotes lead times instead of capacity. In the yard sit accelerators the operator has already bought and cannot energise, because there are no watts left. The usual answers are slow or expensive. A new grid connection is measured in years, new plant in quarters, and a second site restarts procurement and splits the fabric. None produces a watt this quarter, and the yard stock depreciates whether or not it earns.

The envelope is 12 MW of IT load at the rack input. Racks draw 44 kW nominal and carry 32 NVIDIA H100 SXM accelerators at 700 W board power. Accelerators are therefore about 51% of rack draw: the largest lever in the model, set from your own two numbers. The cascade factor defaults to one, no credit to our own physics argument. The yard holds 24 racks, 768 accelerators. Revenue is an illustrative GBP 11,000 to GBP 14,000 per accelerator a year, at 85% of hours let.

Our coefficient is at the accelerator die and your constraint is at the rack input. Between them is the accelerator share, so the reduction at the rack is the die figure times that share and the cascade factor. The conservative reading applies the die figure to the accelerator share alone and returns about 11% at the rack. The stated basis treats the die figure as a floor on wall watts and returns 21%. Both travel together as a range, modelled.

Today 272 racks fit the envelope: 8,704 accelerators drawing 11,968 kW, with 32 kW to spare. Managed, the same 272 racks release about 1,279 kW on the conservative reading and about 2,513 kW on the stated basis. No hardware is added. At the managed rack draw those watts hold 32 racks and 1,024 accelerators, or 72 racks and 2,304 accelerators. The yard stock is 24 racks, so it fits on the conservative reading with 8 racks to spare, and with 48 on the stated basis, modelled.

After the retrofit, with all 24 yard racks energised, the site carries 296 racks and 9,472 accelerators, about 9% more than today. On the conservative reading it draws about 11,632 kW against 11,968 kW before. Less power, more accelerators, modelled. On the stated basis the draw is about 10,289 kW. The 32 kW the site already sat under its envelope is deliberately not spent: it is the model's only reserve, and the reviewer will ask why.

The 768 yard accelerators earn about GBP 8m to GBP 11m a year gross at the declared band, modelled. A second route with no shared inputs, filed operator revenue per megawatt of active power, puts the same watts at about GBP 5.7m to GBP 5.9m a year. Plan against that lower figure, with your own revenue per megawatt in it. The revenue is new: it does not exist without the released watts, because the accelerators that earn it cannot be energised without them.

There is no capital line: no hardware, no construction, no permits. The runtime installs in hours as a software deployment, driver-adjacent, and adds no draw of its own. Revenue begins when the racks are let, on your own installation and sales timeline, not when the software is installed.

ADAPT holds throughput while cutting power. The run gave up to +22% tokens per watt, derived from NVML die power and vLLM serving throughput, with 0% throughput change, derived. On a fixed fleet that gain is the energy line in another unit: the same tokens for less energy. Here the released watts go further and energise the 768 yard accelerators, a capacity effect counted separately and never multiplied with it.

In the case modelled here the released watts go to the yard stock and earn its revenue. Banked instead, about 1,279 kW at a 40% load factor is about 4,483,000 kWh a year. At the DESNZ published non-domestic Extra Large band that is about GBP 940,000, modelled. At nominal draw every hour it is about GBP 2,350,000, an upper bound. Where power is passed through to tenants, the lower bill is theirs and the released capacity to energise more accelerators stays with the operator.

Two properties of your site decide the result. First, the binding limit. Is it a demand or energy figure measured over a settlement interval, or the rating of installed plant? A sustained reduction relieves the first. Plant is sized for the peak the racks can draw, and a sustained reduction does not relieve a plant rating. We free mean watts, and whether mean watts are sellable is the first question we ask.

Second, the duty cycle. A site can be fully allocated on power while its accelerators average well under half utilisation. Power is allocated on provisioned draw, and utilisation is measured on time. The mechanism lives in that gap: 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. On a fully saturated fleet it does not transfer.

The runtime fails open to full performance on a component fault; the workload comes first. The retrofit alone carries no exposure, since 272 racks at nominal draw sit inside the envelope. Exposure appears only when released watts are spent on hardware. At 296 racks, nominal draw is 13,024 kW, 1,024 kW over the envelope, about 9%, modelled. About 79 racks, about 27% of the fleet, would have to release at once to reach it. So the commitment is set below the measured floor with the reserve held back. Ceiling enforcement at the site is part of the first conversation.