Skip to content
Who it is forEnterprise accelerator estates

For enterprises running their own accelerator estates

In your words

"We run our own accelerators, and the building has become the limit."

For organisations that run their own accelerators: more of the work in the queue through the same feed, from a runtime measured at up to 21% less GPU die power on NVIDIA H100 NVL.

The situation

The floor, the feed and the cooling set the pace.

Demand from inside the business has grown faster than the room for accelerators. The budget for the accelerators is rarely the constraint. The floor, the feed and the cooling are.

So the platform's growth has become a facilities question. A capital case for more power goes to a committee, and the work that waits is a model trained later and a roadmap that slips.

An operator sells the headroom it frees. An enterprise puts it to work. ADAPT, a per-GPU software runtime, releases that headroom from the accelerators already installed.

What you are measured on

The lines this decision is judged against.

Work done
Freed headroom, filled with the accelerators that now fit, means more jobs through the same feed for the same teams.
The capacity build not brought forward
About GBP 9m to GBP 12m of one-off capital, modelled, and the largest pool on this page: a smaller capital case for more power, or none at all.
Cost per unit of work
Tokens per watt, read as cost per unit of work: the number your engineers ask for first.
Who holds the electricity account
Energy not bought lands in whichever budget pays the bill, and the work that now fits lands with your platform team either way.
The supplier process
What installs, what it touches, what privilege it needs, what it records and how it comes off, in writing before anything is approved.

Released capacity

How it shows up, in megawatts and money.

The instrument's installed fleet quick start: 1,071 NVIDIA HGX H100 8-GPU nodes at 5.6 kW, already paid for, inside a 6 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 same feed

    modelled

    About 910 to 1,260 kW

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

  • Nodes that now fit beyond the fleet

    modelled

    191 to 285 nodes

    The power left once managed, in whole nodes of 5.6 kW.

  • Accelerators that now fit

    modelled

    1,528 to 2,280 accelerators

    Those nodes at eight accelerators each.

  • Capital not spent building capacity

    modelled

    About GBP 9m to GBP 12m, 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 counts where your budgets keep it. Ask early where it would land, because the work that now fits counts either way. 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 sits under whatever schedules your work and changes what a watt buys, not what runs next. It installs as one service per accelerator under a restricted systemd unit, beside your existing driver.

Your platform, security and change control teams receive the unit template, the removal procedure and a software bill of materials before anything installs. Stopping the service releases the clocks.

A three-week baseline on your own estate turns the modelled figures on this page into measured ones for your fleet.

Platform owner
The install, the operating mode and the baseline.
Facilities lead
Which limit binds at the room, and where the electricity account sits.
Security and change control reviewer
Privilege, network posture, data kept on the host and removal.
Finance and procurement
Where capital not spent would land, and the supplier record.

For the finance lead

  • On our model the largest pool is the capacity build not brought forward: one-off capital, modelled on inputs you can change.
  • Released headroom is room for 1,528 to 2,280 more accelerators through the same feed, modelled, so the next growth needs accelerators rather than a bigger feed. Until they arrive, it is energy not bought each year.
  • Every modelled figure here recomputes in the quick start below, on your own estate.

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

Installed fleet, already paid for

Opens on 6 MW of H100 nodes already installed. Put in your own estate and read the accelerators that fit.

Open this quick start
A conversation

Bring the envelope and the queue. The first question we ask is how your estate runs across a working day.

Talk about your estate