Headroom instrument
Three questions. Your capacity, in your own numbers.
Answer three questions and see what the power you hold is worth: electricity you no longer buy each year, or more accelerators inside the same power and the build you do not fund.
Your answers
What you run, how much, and the power you hold.
Quick start
Set by the quick start. Change any answer and the dashboard follows.
On these answers: 7,120 to 10,632 more accelerators the freed power carries, modelled. Or read it as capital avoided, the build you do not fund: £57m to £80m, one-off. The working is in the dashboard.
Question 1
Systems in this facility
Pick from 96 systems in the hardware register, and add each kind you run. Where the vendor publishes a draw it fills in; where it does not, type what yours pulls.
Dell PowerEdge XE9680L (HGX B200 8-GPU, liquid-cooled)
5,000 nodes, 8 accelerators each, 8 kW each, vendor figure
All 96 systems. Scroll the list for the rest.
Question 2
How much of it?
Racks or nodes, accelerators or megawatts: whichever you know. Change the unit and the same fleet is shown in it.
5,000 nodes, 40,000 accelerators, 40 MW of draw. Your own figures.
Question 3
What power do you hold?
Site IT power at the rack input, the PUE of the hall, and what you pay for electricity. The PUE sets the cooling line; the price sets the energy line.
£209.73 per MWh
Default: DESNZ non-domestic electricity prices, table QEP 3.4.1, Extra Large band (above 150,000 MWh a year), 2026 Q1 (provisional), Climate Change Levy excluded, published 30 June 2026. Your contract price is the one that counts.
Your dashboard
What each freed megawatt is worth, then whether it fits.
Measured: up to 21% less GPU die power on NVIDIA H100 NVL, over 48 hours, with paired arms under an equal power cap. Every figure below is modelled from it and your answers.
Energy you no longer buy each year shows here once you say how the fleet runs.
How does your fleet run most of the day?
7,120 to 10,632 accelerators
More accelerators the freed power carries, modelled.
Capital avoided, the build you do not fund: £57m to £80m, one-off, modelled.
Each freed megawatt counts once: as energy you no longer buy, or as more work.
In balanced mode, tokens per watt run up to 22% higher, derived: the same work for less energy.
It reads the same energy you no longer buy, in another unit, once that line is in.
Derived from GPU die power and serving throughput on the same 48 hour run on NVIDIA H100 NVL, balanced mode.
An engineer reads your answers and this dashboard and writes back with the working.
Send this readingHow the three week validation runs- Fleet today
- 40 MW
- Managed, stated to conservative
- 32 to 34 MW
- Power you hold
- 40 MW
Fit
Whether the fleet fits the power you hold, and by how much.
- Spare today, unmanaged
0 kW
your own figureYour power less your fleet's draw. Managed, 6,000 to 8,400 kW is left: see Capacity.
Tokens per watt
In balanced mode, tokens per watt run up to 22% higher, derived: the same work for less energy.
It reads the same energy you no longer buy, in another unit, once that line is in.
- Reduction we model
15% to 21%
modelledBalanced, from the conservative reading to the stated one.
- Margin against what you need
+15 to +21 points
modelledThe reduction we model, less the reduction you need.
Your fleet fits the power you hold today. Managed, the same envelope carries the extra capacity below, with no new connection and no build.
Capacity
What now fits inside the same power.
- Power this frees
6,000 to 8,400 kW
modelled40 MW of draw x 15% to 21%.
- More accelerators the freed power carries
7,120 to 10,632 accelerators
modelledPower this frees over the managed draw of one node, about 6.3 to 6.8 kW, in whole nodes x 8 accelerators each.
- Nodes that now fit beyond your fleet, including any spare power today
890 to 1,329 nodes
modelledThe power left once managed, in whole nodes of 8 kW, each drawing about 6.3 to 6.8 kW once managed.
- Accelerators that now fit beyond your fleet, including any spare power today
7,120 to 10,632 accelerators
modelledThose nodes x 8 accelerators each.
- Power left once managed
6,000 to 8,400 kW
modelledThe power you hold less what the managed fleet draws.
Revenue
What released capacity earns where you sell it, and the build you do not fund.
- Capital avoided
£57m to £80m
modelledone-offPower this frees x £9,500,000 per MW to build.
Renting capacity rather than holding the power? The saving lands as less power drawn by the accelerators you run, inside the capacity you already contract.
Energy
Energy you no longer buy, and heat the hall no longer has to reject.
Energy waits for how your fleet runs across a day, asked below.
- Step 1
Run it here
Your inputs and one measured coefficient, in your browser.
- Step 2
Send yourself the working
Ask below, then download the document built from your figures.
- Step 3
Run a three week validation on your fleet
Your own measurements replace the modelled figures.
- Step 4
Book a call
On what the validation found, with the people who decide.
Assumptions
The figures the dashboard runs on.
Each one shows its default and where it came from. Type your own and the dashboard follows.
Show the five assumptions: four at their defaults, one yours to enter
- Default£ per MW
£9,500,000 per MW
Default £9,500,000 per MW. Ofgem, Curate: Demand Connections Reform, consultation of 29 July 2026, paragraph 4.12: the average project capital cost of building a megawatt of data centre, a benchmark at April 2026 market conditions.
- Electricity priceSet in question 3
£209.73 per MWh Change it
Default: DESNZ non-domestic electricity prices, table QEP 3.4.1, Extra Large band (above 150,000 MWh a year), 2026 Q1 (provisional), Climate Change Levy excluded, published 30 June 2026. Your contract price is the one that counts.
- Yours to enter£ per MW year
Yours to enter: what you sell a megawatt year of capacity for. Counted only where you sell capacity.
- Facility PUESet in question 3
1.25 Change it
Default 1.25. A working figure for a modern hall. Uptime Institute's July 2025 survey put respondents' weighted average annual PUE at 1.54, with newer and larger sites lower; enter your own.
- Default%
Default 70%. A working figure: set yours. It scales revenue and accelerator hours.
- Where these go
Every value here travels with your link and your request. The ones you change are listed beside their defaults.
Operating mode
Choose the mode. Balanced, the production default and the mode the measured run used, holds throughput.
Balanced shows the measured figure: up to 21% less GPU die power on NVIDIA H100 NVL over 48 hours, with managed and baseline arms under an equal cap, and a 0% throughput change in the run, derived. Performance and Green show preset targets, not measured, from the presets the runtime reads at boot, and the arithmetic uses the middle of each band.
Against every other route
The same capacity, priced in time and capital.
A new grid connection
- Time
- Up to seven years, Europe
- Capital
- Grid and civil works
- What you get
- More megawatts, eventually
New electrical equipment
- Time
- 18 to 54 weeks by category, North America
- Capital
- Capital project
- What you get
- More megawatts, eventually
More accelerators
- Time
- Lead time plus install
- Capital
- Hardware capital
- What you get
- Helps once power is freed inside the envelope
ADAPT, one software install
- Time
- Hours
- Capital
- None
- What you get
- Capacity released from hardware already racked
| Route to more compute | Time | Capital | What you get |
|---|---|---|---|
| A new grid connection | Up to seven years, Europe | Grid and civil works | More megawatts, eventually |
| New electrical equipment | 18 to 54 weeks by category, North America | Capital project | More megawatts, eventually |
| More accelerators | Lead time plus install | Hardware capital | Helps once power is freed inside the envelope |
| ADAPT, one software install | Hours | None | Capacity released from hardware already racked |
Stated, cited. Grid connection: Amazon Web Services, reported by Reuters, 3 February 2026. Electrical equipment lead times: McKinsey and Company, March 2026.
The assessment
What the assessment adds.
The dashboard reads your fleet against the power you hold. This reads how your day runs and what runs on it. Two attributions sit at equal weight, each with its basis.
How your fleet runs most of the day
Work arriving and finishing, power moving as it does: the pattern the measured run was taken on, and the hours this arithmetic counts. Idle hours are left out, so the figure errs low.
Nothing is assumed. The engine asks for the day shape instead of reading one in.
Two readings
Two attributions of the same coefficient, at equal weight.
One answer below completes the reading. Every field changes the result on screen.
- Tell us how your day splits across the three bands. All three are needed.
What would unlock a reading
Across a normal day, how does your fleet split between busy and changing, idle, and pinned at its limit?
This sets how much of the measured saving your day carries. It moves the answer more than anything else here.
Choose the shape of your day above.
What does one of your racks actually pull, measured, rather than its rated draw?
Your measured draw turns the saving into an electricity line, in megawatt hours a year and what they cost.
How many months until new power would actually energise for you?
On most estates in a connection queue this turns out to be the largest single line on the page, and it is entirely your own figure.
What does one accelerator produce, in whatever unit you think in?
States the extra capacity as output rather than as hardware.
Tell us more about the fleet (optional)
What runs on it
In your words. The split by class is optional: it describes the estate and reads each class against the mode, it does not enter the arithmetic.
Split the day by class, per cent (optional)
Who is reading
Optional, and none of it changes a number. It travels with your request, so the reply starts from your situation rather than from a form.
Your grid connection
Your name and email go in the request below. Everything set here goes with them.
The request
Ask for the assessment of this reading.
Your answers, the dashboard, the assumptions you set and everything below the questions go with it.
An engineer reads the figures and writes back with the working and what a baseline on your fleet would involve.