The corporate estate

A large organisation runs accelerators for its own purposes. It sells no compute, so there is no revenue per accelerator-hour. The hall is a corporate data hall, air-cooled, with a worse power usage effectiveness than a purpose-built facility. The capital for the accelerators is already spent, and the programme wants more of them than the hall can power. The envelope cannot grow on the programme's timescale. Changing the building's electrical infrastructure is a capital project with an outage, in a building that runs the corporation's other systems. The hardware is bought; the room is what is missing.

The hall's electrical rating is 3 MW at the facility plane. Its power usage effectiveness is 1.45, entered for an older air-cooled hall, which resolves the envelope to about 2,069 kW at the node input. The node is an NVIDIA DGX H100: eight H100 SXM accelerators at a published maximum of 10.2 kW. The accelerators account for about 55 per cent of that draw. The funded scope is 2,200 kW: 215 nodes, 1,720 accelerators. The day shape is 35 per cent busy and changing, 55 per cent idle and 10 per cent pinned, illustrative until yours replaces it.

Our coefficient is at the accelerator die and your constraint is at the node input. Between them is your accelerator share of node draw. The reduction at the node input is our die figure times your share and an upstream cascade factor we default to one. The conservative reading applies 21% to the 55 per cent share and gives about 12 per cent, assuming nothing about the rest of the node. The stated basis treats die power as a floor on wall watts and gives 21 per cent. Both are modelled.

Today the hall holds 202 nodes and 1,616 accelerators, against a programme of 215 nodes and 1,720. It is 104 accelerators short, 13 nodes, and that figure contains no coefficient of ours. Closing the gap needs about 6 per cent off the draw at the node input. The conservative reading supplies about 12 per cent, so the requirement fits with about 6 points to spare, the widest conservative margin in this set. The stated basis fits with about 15 points.

Inside the same envelope the managed hall holds 229 to 256 nodes and 1,832 to 2,048 accelerators, modelled. Every accelerator the programme bought is energised, and 112 to 328 more stand beyond the requirement. That is an uplift of about 13 to 27 per cent on today, with 129 to 336 kW of headroom left at the node input.

For an estate that runs its own work, the count is the answer: every accelerator already bought is energised, and the same building holds about 13 to 27 per cent more compute, modelled. That is a growth statement for a programme measured on delivery. At an illustrative availability of 90 per cent, the 216 to 432 additional accelerators give about 1,700,000 to 3,400,000 accelerator-hours a year, modelled. The organisation values each hour on its own terms, and on a platform with a queue every one of them goes to work.

Energy is computed from your mean operating draw, 7.14 kW per node, and only across the busy-and-changing 35 per cent of the day. The reduction avoids about 540 to 990 MWh a year, modelled. At an illustrative 20 pence a kilowatt-hour that is about GBP 110,000 to GBP 200,000 a year, modelled. If the room is banked instead, that is its value; for a programme with bought accelerators waiting, energising them comes first. The freed envelope has three exclusive uses: energise the waiting accelerators, bank the energy, or hold it as resilience margin. This page never adds them.

The licence lands on an operating line. The value arrives as capital not spent on electrical works in an occupied building. Inside a corporation those are two directors and often two organisations. Which line is hard to move is a question for the finance function in the first conversation rather than the fourth, because the answer decides who has to be in the room.

The accounting treatment is your adviser's to settle. The documents they will reach for are IAS 38 on intangible assets, the interpretations committee agenda decisions of March 2019 and March 2021 on cloud arrangements, FRS 102 paragraphs 18.3B and 18.4, and ASC 350-40. The question those documents turn on is whether the customer controls the software. A runtime installed on hardware you own is a different object from software reached over a network.

Three questions decide more than any coefficient. Is the binding limit a contractual demand figure measured over a settlement interval, or the rating of a transformer, breaker or busway? A sustained reduction relieves the first and not the second. Is the platform oversubscribed? And can a service be given the authority to set accelerator clocks on the estate?

A corporate buyer also runs a supplier review of any agent taking privilege on production infrastructure. It runs in parallel with the technical evaluation and can add a month on its own. The reviewer will ask for install requirements in writing, a privilege model and a data-protection position for the telemetry. Next come a removal procedure that has been run and timed, and an update and rollback policy with a bounded blast radius per accelerator. Last come a named accountable person and a certification position. That pack belongs in the first conversation. The order of work is the supplier position, then the validation, then the numbers.