Case studies
Find your situation. Take the study.
Each case study starts from a situation in an operator's own words and works it through to racks, megawatts and months. Every one is modelled on inputs you can change. The one measured figure is up to 21% less GPU die power on NVIDIA H100 NVL over 48 hours, managed and baseline arms under an equal cap.
GPU cloud and neocloud
The requirement does not fit the site we have.
Nine megawatts into eight
A 64-rack, 4,608-accelerator requirement fits an 8 MW site that holds 57 racks today, with 3 to 8 racks to spare, modelled.
Who it is for
A neocloud or GPU cloud operator that owns or contracts its accelerators and controls the workload.
The fleet is paid for. When do we retire it?
The fleet that is already paid for
About 14 to 16 months of deferred retirement on a 2,400-accelerator fleet, inside a range of 8 months to a little under 3 years set by the frontier, modelled.
Who it is for
A neocloud or GPU cloud operator running a fully depreciated inference fleet inside a subscribed envelope.
The racks are here. The power is not.
Silicon on the dock
Three of four racks waiting for power start earning inside the same 2,800 kW interim envelope, recovering 48 rack-weeks at a 16-week gap, modelled.
Who it is for
A neocloud or GPU cloud operator with a part-energised site.
We are full, and there is silicon in the yard.
The site that is full
272 installed racks release about 1,279 kW inside a 12 MW envelope, room for 32 racks against 24 waiting in the yard. The site then draws less than before while carrying about 9 per cent more accelerators, modelled.
Who it is for
A neocloud or GPU cloud operator with a fully allocated site.
We are on a committed power contract. We pay for it whether we draw it or not.
The power that is paid for either way
A 12 MW envelope already paid for holds 6,048 to 6,480 accelerators where it held 5,112, with 432 to 864 beyond the requirement, modelled.
Who it is for
A compute operator that owns its accelerators, sells the hours and holds its own electricity account on a committed quantity, short of capacity against its requirement.
We would rather grow than save.
Headroom put back to work
48 racks become 54 inside the same 6,800 kW: 432 more accelerators selling compute, and each unit of work takes about 8 to 12 per cent less energy, modelled.
Who it is for
An operator that owns its accelerators, sells compute from an envelope it cannot enlarge and has demand it cannot serve.
Our limit is cooling, not supply.
The heat that cannot leave
47 to 50 racks under a design-day ceiling that holds 40 today, wherever pinned runs are shorter than the window the limit is measured over, modelled.
Who it is for
A site that owns its accelerators and controls what runs on them, limited by heat rejection on its design day.
Are we a good candidate?
The fleet this is for
A 7,200 kW envelope that holds 51 racks today holds 60 to 65 with the runtime, 3 to 8 racks beyond a 57-rack requirement, modelled.
Who it is for
A GPU cloud operator that owns its accelerators, holds its own site and electricity account, and is sold out.
New builds and campuses
We have demand now and a connection date years out.
The connection that is years away
A 500-rack, 36,000-accelerator deployment applies for 81 to 86 MW rather than 100 MW: 14 to 19 MW of contracted capacity never requested, modelled.
Who it is for
A greenfield compute developer or operator holding its own electrical envelope, with a network operator as counterparty.
We are planning at gigawatt scale.
The gigawatt programme
A 1,100 MW requirement fits inside a 1,000 MW envelope with about 66 to 131 MW of headroom to spare, modelled.
Who it is for
A programme that owns or will own its accelerators, controls what runs on them, and holds or is requesting its own electrical envelope above a gigawatt.
Your own numbers
The same arithmetic, run on your site.
Bring your envelope, your racks and your date. The instrument gives a first reading now, and a three-week validation on your own fleet makes the figure yours.