For GPU cloud and neocloud operators
"Our book is sold ahead of our megawatts. The racks are on the dock and the power is not."
For operators who own or contract their accelerators: a runtime measured at up to 21% less GPU die power on NVIDIA H100 NVL over 48 hours, so more of the capacity you have sold fits the megawatts you hold, modelled.
The situation
Space is not what is missing. Power is.
Capacity is contracted with a date on it. The racks are ordered or already delivered, and the site that should take them is not yet energised.
Three routes are already priced. More power is a queue and a build. Fewer accelerators give up the revenue that justified buying them. A later date moves the problem to a customer who holds a contract with you.
ADAPT works on the constrained side instead: the same work on less power, inside the envelope you already hold. It is a per-GPU software runtime, installed driver-adjacent on the hosts you run, and it changes how each accelerator draws power while your work runs.
What you are measured on
The lines this decision is judged against.
- Revenue per installed megawatt
- Headroom you free is capacity you can sell, and it lands in the quarter it is freed. It is worth most when the book is sold out.
- The build you do not bring forward
- About GBP 14m to GBP 20m, one-off, modelled: the largest pool on this page, taken as a tranche connected sooner or a request you never have to make.
- Tokens per watt
- The number your engineers ask for first: up to +22%, derived from the same run. It is the same saving, read from the serving side.
- Gross margin
- Power sits in cost of revenue, so energy you no longer buy lands in margin with nothing offsetting it.
- Duty cycle
- On-demand and spot beside reserved capacity, serving traffic with a daily shape and gaps between jobs. That is the shape of the measured run, and a three-week baseline reads yours.
Released capacity
How it shows up, in megawatts and money.
The instrument's 10 MW neocloud quick start: 1,785 NVIDIA HGX H100 8-GPU nodes at 5.6 kW each, inside a 10 MW envelope, with the capacity sold.
From the 48-hour run on NVIDIA H100 NVL
- measured
Up to 21%
Less GPU die power
NVIDIA H100 NVL, one continuous 48-hour run, managed and baseline arms under an equal power cap, read as NVML GPU die power. Die power is a lower bound on wall power.
- derived
Up to +22%
Tokens per watt
From die power and serving throughput on the same run: the same work for less energy.
- derived
0%
Throughput change in the run
The same 48-hour run.
- observed
10 to 15 degrees C
Cooler at the die
Observed in the same testing.
Modelled on the quick start, whole numbers
Power freed inside the envelope
modelledAbout 1,510 to 2,100 kW
10 MW of draw, from the conservative reading to the stated one.
Nodes that now fit beyond the fleet
modelled318 to 475 nodes
The power left once managed, in whole nodes of 5.6 kW.
Accelerators that now fit
modelled2,544 to 3,800 accelerators
Those nodes at eight accelerators each.
Revenue from released capacity
modelledAbout GBP 2m to GBP 3m each year
The quick start's working figure of GBP 1.6m per megawatt year, at 70 per cent utilisation. Your own figure replaces it.
Capital not spent on the next build
modelledAbout GBP 14m to GBP 20m, one-off
Power freed times GBP 9.5m per megawatt, the average project capital cost of a data centre in the Ofgem consultation. Ofgem, Curate: Demand Connections Reform, consultation of 29 July 2026, paragraph 4.12.
| Figure | Value | Kind | Basis |
|---|---|---|---|
| Power freed inside the envelope | About 1,510 to 2,100 kW | modelled | 10 MW of draw, from the conservative reading to the stated one. |
| Nodes that now fit beyond the fleet | 318 to 475 nodes | modelled | The power left once managed, in whole nodes of 5.6 kW. |
| Accelerators that now fit | 2,544 to 3,800 accelerators | modelled | Those nodes at eight accelerators each. |
| Revenue from released capacity | About GBP 2m to GBP 3m each year | modelled | The quick start's working figure of GBP 1.6m per megawatt year, at 70 per cent utilisation. Your own figure replaces it. |
| Capital not spent on the next build | About GBP 14m to GBP 20m, one-off | modelled | Power freed times GBP 9.5m per megawatt, the average project capital cost of a data centre in the Ofgem consultation. Ofgem, Curate: Demand Connections Reform, consultation of 29 July 2026, paragraph 4.12. |
Selling released capacity and not funding the next build are two readings of the same megawatt: count one, not both. Each range runs from the conservative reading to the stated one. The conservative reading applies the die figure to the accelerators' share of node draw. The stated reading treats die power as a floor on wall power. Hardware other than NVIDIA H100 NVL stays modelled until a baseline run on it.
Installing it
What installing it involves, and who signs.
For the reviewer: what installs, what it listens on, what it reads and keeps, and how it comes off.
ADAPT installs alongside what already runs: one service per accelerator under a restricted systemd unit, beside your existing driver. Workloads, scheduler and hardware stay as they are.
Stopping the service releases the clocks, and a written procedure takes it off. The runtime is designed to fail open to full performance on a component fault, so where a site is filled by virtue of the reduction, keep the margin at the site: reserved headroom, or a commitment set below the modelled figure.
A three-week baseline on your own racks turns the modelled figures on this page into measured ones for your fleet.
- Chief technology officer or platform lead
- The install on the hosts, and the operating mode.
- Head of capacity or infrastructure
- Which limit binds, and how released capacity is sold.
- Security and change control reviewer
- What installs, what it touches and how it comes off.
- Finance lead
- The build date, and which reading of the megawatt to count.
For the finance lead
- The largest pool is the build you do not bring forward: capital not spent, one-off, modelled on inputs you can change.
- Energy not bought recurs each year; capital not spent happens once. Take each released megawatt where it is worth most to you: sold as capacity, kept as energy not bought, or held as a build you do not fund.
- Every modelled figure here recomputes in the quick start below, on your own hardware, envelope and working figures.
Case studies that fit
Situations like yours, worked through.
- Case study, modelledNine megawatts into eightA 64-rack requirement fits an 8 MW site that holds 57 racks today, with 3 to 8 racks to spare, modelled.
- Case study, modelledSilicon on the dockRacks delivered ahead of their power, and the weeks between dock and earning.
- Case study, modelledThe fleet this is forA 7,200 kW envelope that holds 51 racks today holds 60 to 65 with the runtime, modelled, and seven signs your fleet is the same kind.
Your own numbers
Run it on your site, then talk to us.
10 MW neocloud cluster
Opens on 1,785 H100 nodes in 10 MW with the capacity sold. Change the hardware, the envelope and the assumptions to your own.
Bring the envelope, the racks and the date. The first question we ask is how your fleet runs across a day.