For colocation operators and their tenants
"Our halls were designed for a lower density than the kit arriving in them."
For the operator that sells contracted power and space, and the tenant inside it: more cabinets energised in halls already built, modelled from a runtime measured at up to 21% less GPU die power on NVIDIA H100 NVL.
The situation
One demarcation line, and value on both sides of it.
If you operate the building, the constraint is density. The kit arriving draws more per cabinet than the halls were designed for, so there is space you cannot energise and signed business waiting on power.
If you are the tenant, you hold the accelerators and decide what runs on them. The megawatts, and the plan for them, belong to the building.
ADAPT is a per-GPU software runtime. On dedicated bare metal the tenant installs it on the accelerators it runs, and the operator makes the introduction. The value lands on both sides of the line: sellable density for the building, room for more work for the tenant.
What you are measured on
The lines this decision is judged against.
- Operator: cabinets you can energise
- Freed headroom inside a building already built is space you can sell without re-engineering a hall.
- Operator: the tenant you keep
- A requirement that now fits an existing hall is a tenant retained and a capital plan that holds.
- Tenant: work inside the contract
- Freed headroom is room for more of your own accelerators, and their work, inside the same contracted envelope.
- Tenant: authority and duty cycle
- Bare metal with root gives the runtime what it needs. Serving traffic with a daily shape and gaps between jobs is the measured case.
- Both: which limit binds
- A sustained reduction relieves a limit counted over an interval. A hall often has several limits, so the arithmetic starts with the one that binds.
Released capacity
How it shows up, in megawatts and money.
The instrument's custom fleet quick start, read as a hall-scale tenancy: 4,285 NVIDIA HGX H100 8-GPU nodes at 5.6 kW, inside a 24 MW contracted 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 contracted envelope
modelledAbout 3,630 to 5,040 kW
24 MW of draw, from the conservative reading to the stated one.
Nodes that now fit beyond the fleet
modelled764 to 1,139 nodes
The power left once managed, in whole nodes of 5.6 kW.
Accelerators that now fit
modelled6,112 to 9,112 accelerators
Those nodes at eight accelerators each.
Capital not spent building a hall
modelledAbout GBP 34m to GBP 48m, 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 contracted envelope | About 3,630 to 5,040 kW | modelled | 24 MW of draw, from the conservative reading to the stated one. |
| Nodes that now fit beyond the fleet | 764 to 1,139 nodes | modelled | The power left once managed, in whole nodes of 5.6 kW. |
| Accelerators that now fit | 6,112 to 9,112 accelerators | modelled | Those nodes at eight accelerators each. |
| Capital not spent building a hall | About GBP 34m to GBP 48m, 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. |
The avoided build sits with the party that signs the megawatts, usually the operator. The work that now fits inside the contract sits with the tenant. 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 tenant installs ADAPT on the accelerators it controls: one service per accelerator, beside the existing driver. Nothing is added to the building, its cooling or its distribution, and stopping the service releases the clocks.
The operator holds the envelope and the introduction. Introduce the tenant, or send them this page: the arrangement works best with both parties in the room.
The runtime is designed to fail open to full performance on a component fault. Where space is occupied by virtue of the reduction, agree the margin at the site: reserved headroom, or a commitment set below the modelled figure.
- Tenant platform owner
- The install on its own accelerators, and the operating mode.
- Operator capacity or facilities lead
- The contracted envelope, and which limit in the hall binds.
- Tenant security and change control reviewer
- What installs, what it touches and how it comes off.
- Commercial leads on both sides
- How freed headroom and any avoided build are recorded in the lease.
For the finance lead
- For the operator, released headroom is space you sell in a hall already built, or a hall you do not build: about GBP 34m to GBP 48m, one-off, modelled, with whoever signs the megawatts. Where energy passes through to the tenant at cost, the space is the line to count.
- For the tenant, freed headroom is room for more of its own accelerators inside the contracted envelope. Left unfilled, it is energy not bought each year where the tenant carries the tariff.
- Every modelled figure here recomputes in the quick start below, with the tenant's hardware, the contracted envelope and the build cost.
Case studies that fit
Situations like yours, worked through.
- Case study, modelledThe tenant and the landlordA 32-rack tenant requirement fits a 4 MW contracted critical load that holds 28 racks today, with 1 to 4 racks to spare, modelled.
- Case study, modelledThe power that is paid for either way432 to 864 more accelerators inside a 12 MW envelope already paid for, modelled.
- Case study, modelledThe site that is fullA full hall that energises the racks waiting in its yard.
Your own numbers
Run it on your site, then talk to us.
Custom fleet
Opens on 24 MW of H100 nodes. Set the tenant's hardware, the contracted envelope and the build cost, and read what fits.
Operators and tenants use the same door. Say which side of the line you are on, and bring the envelope.