The tenant and the landlord

One company owns the accelerators and decides what runs on them. Another owns the building, holds the electricity account, sells contracted kilowatts and signs the capital for the next megawatt. Reducing what the accelerators draw creates value in that hall, and the lease decides which company keeps it. We read seven colocation agreements. In each, the line between the two runs through the output breaker of the power distribution unit. The meter sits on its supply side, and everything an accelerator host does is on the tenant's side.

None of the seven restricts software installation, and none carries a minimum-draw covenant. In the operators' own terms, the customer's equipment, with the software on it, is the customer's property. Commitment clauses act on the excess: exceeding a commitment brings rebilling and a right to power down, and every kilowatt kept below it stays the tenant's to fill. Three of the seven are filed with a securities regulator, four are undated published terms, and all are drafted in the United States. Your own agreement settles it. Search it for a software restriction, a minimum-draw covenant and a headroom-sharing clause.

The tenant holds a contracted critical load of 4 MW, measured as IT load at the rack input. It needs 32 racks of 72 accelerators at a nominal 140 kW each: 2,304 accelerators drawing 4,480 kW. Fitting needs a reduction of about 11%. Without the runtime, 4 MW holds 28 whole racks and 2,016 accelerators, so the tenant is 288 accelerators and 4 racks short. That shortfall is the largest figure on the page, and it contains no number of ours.

Our coefficient is at the accelerator die; the tenant is billed at the supply side of the power distribution unit. Between them sits the accelerators' share of rack draw, 72% here, and a cascade factor set to one. Two readings travel at equal weight. The conservative reading applies 21% to the accelerator share alone and returns about 15% at the metered plane. The stated basis treats die power as a floor on wall watts and returns 21%, a physics argument, modelled.

On the conservative reading a managed rack draws about 119 kW, and 4 MW holds 33 racks and 2,376 accelerators. The requirement is met with 1 rack, 72 accelerators and 197 kW of headroom to spare. On the stated basis a rack draws about 111 kW and 36 racks fit, 2,592 accelerators. Then 4 racks, 288 accelerators and 461 kW sit beyond the requirement. The fit holds at any accelerator share of 51% or more; a rack with more host and network overhead enters its own share and the arithmetic reruns.

The coefficient applies to the busy-and-changing hours only, 55% of the day. At an illustrative mean draw of 95 kW a rack the fleet averages 3,040 kW. The demand avoided is about 253 to 351 kW: about 2,200 to 3,100 MWh a year, modelled. At an illustrative 20 pence a kilowatt-hour, your own tariff, that is about GBP 440,000 to GBP 620,000 a year, modelled. On a colocation bill where power is about a quarter, from one operator's disclosed segment, it is about 4% to 5% of the bill.

The rack count and the bill percentage are alternative uses of the same watts and are never added. Per space, 33 racks inside the same 4 MW against 28 is about 18% more accelerators on the conservative reading, and about 29% on the stated basis. Per accelerator, the derived ratio is up to +22% tokens per watt in Balanced mode. That is the same tokens for less energy, which is this energy line restated.

Wholesale, with power recharged at cost: the tenant installs on its own hardware inside its own space. Its base rent on committed kilowatts does not move, its metered power payment falls by the energy line, and the freed capacity is its own. Both halves reach the party that deployed. Retail, with power bundled into the licence: the capacity still reaches the tenant, but the tenant's bill does not move. The energy saving lands in the landlord's accounts. Two operators' audited filings from February 2026 describe those positions, one passing utilities straight through and one bundling power and carrying the exposure.

Now the landlord who wants the reduction across its hall. It holds the meter, the contracted kilowatts and the capital for the next megawatt, and the accelerators sit on the tenant's side of its breaker. The measured figure travels with whoever owns the accelerators and controls the workload, so the install is the tenant's to make. The landlord gains without an install of its own: it sees first the requirements that do not fit, keeps them in its hall, and receives the avoided capital for the next megawatt. It brings the tenant into the room. No published lease we read shares headroom released inside contracted kilowatts. Agreeing one is a landlord-and-tenant conversation, and we will sit in it.

Racks that fit only because the runtime is reducing their draw make continued operation a standing condition. On the conservative reading 33 racks are energised, 5 of them only under the reduction. Unmanaged they draw 4,620 kW against a 4,000 kW contract. A total runtime failure puts 620 kW of unplanned draw on the commitment, or 1,040 kW at 36 racks on the stated basis. The runtime fails open to full performance by design, right at the accelerator, so the mitigation lives in the site plan.

In a shared estate the excess is also contractual. It triggers rebilling and, at enough of an excess, a right to power down that the tenant has already given its landlord. So the validation comes before the rack order. One question then belongs to the landlord. The distribution plant feeding the space has a rating, and a sustained reduction relieves a contractual figure over a settlement interval without relieving a transformer or breaker.