The connection that is years away

You have demand now and a connection date years out, and the wait is published. The Ofgem Curate consultation of 29 July 2026 counts about 73 GW of data-centre demand in the Great Britain queue at June 2025. Half of it sits in the 100 to 500 MW band. Priority in the reformed queue goes by strategic designation and readiness; capital alone does not move you. Every connection regime rations contracted capacity, and none rewards consumption. A deployment that draws less asks for something different. A smaller application is cheaper to hold, smaller to underwrite and easier to satisfy at a milestone review.

A connection application is denominated in maximum import, and a sustained reduction is a reduction in the mean. One question about your site decides whether the arithmetic runs. Is your contracted figure set by a saturated coincident maximum, or by a diversified expected maximum below it? On the second, the chain runs on your inputs. On the first, the conversation starts with your day profile. The mechanism is duty-cycle intelligence on bursty fleets: most of the saving is idle-floor reduction, and most of the rest is clock-down at mid utilisation. A fleet bursty enough for the reduction to exist has a contracted figure below its saturated maximum.

Our coefficient is at the accelerator die, and your application is denominated at the grid connection. Four links sit between them: your accelerator share of rack draw, your non-rack load, your power usage effectiveness and your design margin. We supply only the first link's default, from a published rack specification. We set the upstream cascade factor to one, which gives our own physics argument no credit.

Two readings travel together everywhere on this page. The conservative reading applies the die figure to the accelerator share only: about 15% at the rack input, modelled. The stated basis carries the die figure across unchanged: 21% at the rack input, modelled. Power usage effectiveness scales both cases alike. Non-rack load sits unchanged in both and dilutes the ratio.

The rack is the NVIDIA GB300 NVL72: 72 Blackwell Ultra accelerators and 36 Grace processors, liquid-cooled, at 140 kW nominal from OEM system specifications. The accelerators are about 72 per cent of that draw. Five hundred racks is 36,000 accelerators and 70 MW at the rack input. Add 5,500 kW of non-rack load, apply a power usage effectiveness of 1.20 and a 10 per cent design margin. The application rounds up to 100 MW, the regulator's own hypothetical data centre.

Run the same chain with the reduction and the application is 86 MW on the conservative reading and 81 MW on the stated basis, modelled. That is 14 to 19 MW of contracted capacity never requested. The gap between the readings is 5 MW. The measurement uncertainty on the coefficient moves the application by under half a megawatt, so it barely touches the 14 to 19 MW; your rack draw basis and accelerator share are the inputs to pin down. Held the other way, the same 100 MW carries 591 to 635 racks against 501, modelled. Take one use of the watts, never both.

In Great Britain, the Ofgem Curate consultation of 29 July 2026 proposed a connection commitment, returned at energisation and forfeited on early exit. Every figure here is worked on that proposal as published. Against that proposal, 14 MW never requested is about GBP 3m to GBP 10m of commitment not posted, modelled. At 19 MW it is about GBP 5m to GBP 14m. Because the commitment returns, its cost is the carry: about GBP 1m to GBP 13m across the regulator's 2029 to 2035 connection dates, modelled. Scaled to the megawatts not requested on the conservative reading, the Ofgem estimate of 20 to 170 basis points of return on a hypothetical 100 MW data centre is about 3 to 24 basis points, modelled.

In Ireland, the EirGrid connection policy of 20 May 2026 requires new dispatchable generation or storage matched to a data centre's maximum import, about 1.16 MW per MVA. At unity power factor, a lower bound, the smaller ask is about 16 to 22 MW of generation never built, modelled. You supply the installed cost per megawatt.

The channel that outweighs both is the connection date: whatever your compute plausibly sells for, energising earlier is worth one to two orders of magnitude more than the commitment. No jurisdiction publishes a link from application size to date, so the smaller application stands on what it is: cheaper to hold, smaller to underwrite and easier to satisfy at a milestone review.

ADAPT is per-GPU software, driver-adjacent. A site connected at 86 MW on its strength has made continued operation a standing condition of the configuration. The runtime is designed to fail open to full performance on any component fault, which protects the workload and returns the site to full draw. On the conservative reading a total software failure exposes about 14 MW against the 86 MW contracted, modelled.

The reserve lives at the site, in three shapes, and never in the 5 MW gap between the two readings. The first is a facility-level limit the runtime cannot fail out of, the one shape that supports the full modelled reduction. The second is a firm connection at the reduced figure with the difference carried as non-firm, a shape whose setpoint the network issues. The third is an application sized partway, where value and exposure grow together.

Alongside the smaller application, the conservative reduction avoids 37,086 MWh a year at a 40 per cent duty factor and 92,716 MWh at full duty: about GBP 8m to GBP 19m at the DESNZ published non-domestic Extra Large band for Q1 2026, modelled. Site token output is unchanged, because the rack count is fixed. Per accelerator the runtime holds throughput while cutting power: up to +22% tokens per watt, derived, with 0% throughput change in the run. The tokens stay the same and the power you procure falls.