Skip to content
Who it is forSystems builders, integrators and OEMs

For systems builders, integrators and OEMs

In your words

"The customer wants the system. The site cannot power what we quoted."

For builders, integrators and resellers whose customers' sites come up short on power: more of the quoted configuration fits the site's envelope, modelled from a runtime measured at up to 21% less GPU die power on NVIDIA H100 NVL.

The situation

Demand is there. Sites that can power the quote are scarce.

A configuration is sized, and the site comes back short on megawatts rather than floor space or budget.

The order then shrinks, divides across two sites, or waits for a connection nobody can date. The constraint sits at the customer's site, and the order that does not close sits in your pipeline.

ADAPT is software on the system you ship: a per-GPU runtime, driver-adjacent, in the bill of materials only as software. It changes what a fixed envelope carries, so more of the configuration fits.

What you are measured on

The lines this decision is judged against.

Configurations that fit
A customer envelope that now carries more of the quoted system is an order that keeps its size and its date.
Compute per megawatt
The currency a site owner converts everything else into, inside the envelope they hold.
Your onboarding process
What your process requires before a software line can be carried. The partners page sets out who answers what.
Attribution
Headroom from cooling or in-rack technology belongs to the platform and the silicon vendor. ADAPT's claim is software on an installed fleet.
A result on the parts you ship
Measured on NVIDIA H100 NVL. A paired baseline on a part you ship, the same design as the published run, makes that part measured too.

Released capacity

How it shows up, in megawatts and money.

The instrument's GB300 quick start: 60 GB300 NVL72 racks at 132 kW, 7,920 kW of draw, quoted into a 6,500 kW 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 the order needs off to fit

    your own figure

    About 1,420 kW

    7,920 kW of draw against a 6,500 kW envelope. Nothing of ours in it.

  • Power freed

    modelled

    About 1,200 to 1,660 kW

    7,920 kW of draw, from the conservative reading to the stated one.

  • The order against the envelope

    modelled

    About 1,200 kW of the 1,420 kW gap closed, to the whole order with 2 racks spare

    On the conservative reading the runtime closes about 1,200 kW of the 1,420 kW gap, and a baseline on the site shows which reading your racks follow. On the stated reading the whole order fits, with room for two more racks.

  • Capital the customer does not spend

    modelled

    About GBP 11m to GBP 16m, one-off

    Power freed times GBP 9.5m per megawatt, the average project capital cost of a data centre in the Ofgem consultation. On the conservative reading the build the customer still needs falls from about 1,420 kW to about 220 kW. Ofgem, Curate: Demand Connections Reform, consultation of 29 July 2026, paragraph 4.12.

GB300 is modelled on the H100 NVL coefficient. 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 on the customer's hosts beside the driver, one service per accelerator. The thermal design, the cooling and the rack stay as shipped, and stopping the service releases the clocks.

The customer's own reviewer approves it from the security review. ADAPT is carried under its own name or introduced, the route is agreed with each partner in writing, and an account you introduce is recorded as yours.

Start with a paired baseline on a part you ship, and quote from your own measured figure.

Your solutions or product lead
Whether the line fits your catalogue and the quote.
Your customer's platform owner
The install on the fleet they own.
Your customer's security reviewer
What installs, what it touches and how it comes off.
Commercial leads
The route, carried or introduced, agreed in writing.

For the finance lead

  • The largest pool sits on your customer's books: the build they do not bring forward, which is why a configuration that did not fit now fits.
  • On your own books it shows up as an order that keeps its size and its date.
  • Every modelled figure here recomputes in the quick start below, on the configuration you last quoted.

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

60 GB300 racks into 6,500 kW

Opens on 60 GB300 NVL72 racks quoted into a tight envelope. Swap in the configuration you last quoted.

Open this quick start
A conversation

Bring the last configuration that did not fit. We start with the site's envelope and the customer's day.

Talk about a configuration