The cluster being quoted

An integrator, an OEM or an NVIDIA partner is quoting an inference cluster. The workload is understood. Racks, network, electrical design and the connection application are not yet fixed. Once the transformer is ordered and the application lodged, the envelope is set for the site's life. This is the one point where a software runtime changes a specification rather than a bill. At design stage the binding constraint is contracted capacity, not consumption. Every regime we examined rations what a site asks for rather than the energy it uses. Here a reduction converts into value as a smaller capacity request.

The reference rack is a declared input: the NVIDIA GB300 NVL72, 72 accelerators and 36 host processors, liquid-cooled. OEM specifications put it at 132 to 142 kW. The model uses 140 kW, with the accelerator dies at about 101 kW, a 72 per cent share. The client needs 2,880 accelerators: 40 whole racks drawing 5,600 kW unreduced, against a 5 MW envelope of IT load. Fitting needs a reduction of about 11 per cent, modelled.

Our coefficient sits at the accelerator die and your specification at the rack input, so the conversion is one multiplication. It is 21% times your accelerator share times an upstream cascade factor we default to one. Two readings travel together. The conservative reading gives about 15 per cent at the rack, modelled. The stated basis, which treats the die figure as a floor on wall watts, gives 21 per cent, modelled.

Busway, transformer and switchgear stay sized on nameplate peak plus margin, about 7,440 kW here, as they are today. That is the recorded convention, 155 kW peak plus about 20 per cent headroom, against 5,600 kW nominal, modelled. The reduction enters through the connection application, a declared demand built on a utilisation assumption, and a sustained-mean measurement is the evidence for that assumption.

Ofgem said in June 2026 that it was considering mandatory disclosure of intended ramp rates and utilisation factors. The design authority who owns the application converts that evidence into a smaller specification. Everything priced below rests on that row.

Unreduced, a 5 MW envelope holds 35 racks, 2,520 accelerators, 5 racks short, modelled. With the runtime assumed, effective rack draw falls to about 119 kW on the conservative reading and about 111 kW on the stated basis. The 40 racks then draw about 4,750 to 4,420 kW and fit with about 4 to 10 points of margin, modelled. Our re-derivation delta at the die is under half a point; the conservative margin is roughly fifteen times that.

What gets smaller is the connection application. Contracted capacity not requested is about 850 to 1,180 kW, modelled. The Ofgem Curate consultation of 29 July 2026 proposed a connection commitment for data centres in Great Britain. Against it, that is about GBP 200,000 to GBP 840,000 not posted, modelled. At ten times the scale, 400 racks, the capacity not requested is about 8 to 12 MW, modelled, itself a small data centre.

Hold the envelope at 5,600 kW and 47 racks fit on the conservative reading, 50 on the stated basis. That is 7 to 10 more racks and 504 to 720 more accelerators, about 18 to 25 per cent, modelled. Those racks are hardware the client or integrator still buys. The runtime removes the power constraint, so direction B is a larger quote, not a cheaper one. Per accelerator the run gave up to +22% tokens per watt, derived, at 0% throughput change, derived. Per quote, the token uplift is the rack count. The two directions spend the same watts once: choose one, quote it, and plan against it.

The energy line is the smallest. Direction A avoids about 850 to 1,180 kW of continuous demand. At a 40 per cent duty factor that is about 3,000 to 4,100 MWh a year, worth about GBP 740,000 to GBP 1,030,000 at a United Kingdom commercial tariff, modelled. The figure uses a 40 per cent duty factor, because the saving comes from the idle and mid-utilisation hours that bursty load leaves; a baseline run on the client's own accelerators sets the real figure.

Heat follows the draw: about 21 to 29 kW per rack leaves the cooling loop, modelled. That feeds liquid-cooling sizing on your loop design. Capital is your field: megawatts not built times your own cost per megawatt. At an illustrative GBP 10m per MW, which is not a default, that is about GBP 8m to GBP 12m one-off, modelled. Where demand is growing it is deferred rather than avoided.

Three signatures decide this scenario. The partner's software-portfolio owner decides which third-party software the organisation carries, on a window of roughly two months. That owner asks for a support model and a qualification pack. The end client's design authority owns the diversity factor and the connection application. Whoever signs the capital authorisation fixes the envelope. Their first question is how the licence sits in the accounts, a question for their auditor, so bring the auditor in early.

The runtime installs driver-adjacent above NVML, one lightweight service per accelerator, with no hardware and no application code change: a software line in the quote. A cluster specified into an envelope on its strength keeps the runtime as a standing condition of the configuration. That condition belongs in the quote and the first conversation. A three-week validation on the client's own fleet, run before the application is lodged, gives the design authority a number they can put in writing.