Skip to content
Who it is forPublic bodies running funded compute

For public bodies running funded compute

In your words

"The compute is funded, the envelope is fixed, and the connection date is not ours to move."

For public bodies with funded compute: more research capacity inside the fixed envelope, without a new connection, from a runtime measured at up to 21% less GPU die power on NVIDIA H100 NVL.

The situation

More compute inside the funded envelope, and a shorter wait for the next one.

The envelope was set when the programme was funded. A capital case for more power goes to a committee, and the connection date at the end of it sits outside the body's control.

Most of that is contracted capacity, or capacity requested, rather than a bill. So drawing less shows up as a smaller request, made sooner, with a shorter wait at the end of it.

ADAPT is a per-GPU software runtime on the accelerators the programme already runs. It changes how much compute the funded envelope holds, without a new connection.

What you are measured on

The lines this decision is judged against.

Research capacity delivered
Accelerator-hours allocated to the user community inside the funded period.
The capacity request
A smaller request made sooner, which shortens the wait at the end of it.
Capital preserved
About GBP 57m to GBP 80m of build not needed, one-off, modelled, for your finance function to place.
The supplier check
What installs, what it touches, what privilege it needs and how it comes off, in writing for the security and change control reviewer.
Duty cycle
Serving estates with a daily shape recover most. A queue kept full recovers less, and a baseline under your own queue measures how much.

Released capacity

How it shows up, in megawatts and money.

The instrument's 40 MW campus quick start: 5,000 NVIDIA HGX B200 8-GPU nodes at 8 kW, inside a 40 MW envelope.

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 funded envelope

    modelled

    About 6,050 to 8,400 kW

    40 MW of draw, from the conservative reading to the stated one.

  • Nodes that now fit beyond the estate

    modelled

    890 to 1,329 nodes

    The power left once managed, in whole nodes of 8 kW.

  • Accelerators that now fit

    modelled

    7,120 to 10,632 accelerators

    Those nodes at eight accelerators each.

  • Accelerator-hours a year, once added

    modelled

    About 43,660,000 to 65,200,000

    Worked from the quick start's accelerators that now fit, once they are added: 8,760 hours a year at its working utilisation of 70 per cent.

  • Capital not spent on the next build

    modelled

    About GBP 57m to GBP 80m, 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.

Capital not spent and capacity released 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.

The security review

ADAPT installs as one service per accelerator under a restricted systemd unit, beside the existing driver, and sits under whatever allocates the work. Stopping the service releases the clocks, and a written procedure removes it.

The security and change control reviewer receives the unit template, the removal procedure and a software bill of materials before anything installs. The runtime is designed to fail open to full performance on a component fault.

A three-week baseline on the programme's own accelerators turns the modelled figures here into measured ones, under the programme's own name.

Senior responsible owner
The programme scope, and the baseline.
Estates or facilities lead
The funded envelope, and which limit binds.
Security and change control reviewer
What installs, what it touches and how it comes off.
Commercial and finance leads
The supplier record, and how capital not spent is treated.

For the finance lead

  • Released capacity is room for 7,120 to 10,632 more accelerators inside the same funded envelope, modelled, so the next capital bid buys accelerators rather than a new connection.
  • Every modelled figure here recomputes in the quick start below, on the programme's own hardware and envelope.

Case studies that fit

Situations like yours, worked through.

All case studies

Your own numbers

Run it on your site, then talk to us.

The instrument quick start

40 MW campus

Opens on 5,000 B200 nodes in 40 MW. Put in the programme's own hardware and envelope.

Open this quick start
A conversation

Bring the envelope, the funded scope and the date. We start with how the estate runs across a week.

Talk about a programme