Power Envelopes

Compute is an electrical load. Everything else about its expansion is downstream of that fact.

The constraint on building computing capacity used to be capital, then it was silicon. In every market we examined, it is now electricity — specifically, the regulated process by which a large industrial load is granted a connection to the transmission network. This process was designed for a world in which large new loads arrived rarely and predictably. It is now absorbing requests at a rate several times its design assumption.

Four quantities, routinely confused

Careful analysis of this area requires distinguishing four things that are usually described with the same word.

1. 01Requested capacity: a position in an interconnection queue. Costs little to obtain and carries no commitment to build.
2. 02Contracted capacity: a signed agreement for delivery, typically with financial commitment and milestone obligations.
3. 03Energised capacity: electrical supply physically delivered to the site and available for use.
4. 04Consumed capacity: power actually drawn, which for datacentres runs well below energised capacity at almost all times.

Public reporting moves between these four without marking the transition. A figure describing requested capacity will be reported as though it described consumed capacity, producing estimates of electricity demand that are wrong by more than a factor of two.

Queue duration

Median duration from request to energisation for loads above 300 MW is 51 months in our dataset. The distribution has a long right tail: the ninetieth percentile exceeds eight years. Duration is only weakly related to project size and strongly related to whether new transmission construction is required. Projects that can be served by existing transmission clear in a fraction of the time.

This has produced a predictable behavioural response. Developers file interconnection requests early, at multiple sites, for capacity well above their expected need, in order to hold queue positions. The consequence is queue inflation: the requested figure rises, study resources are consumed on projects that will never be built, and genuine projects are delayed behind speculative ones.

Behind the meter

A growing share of new large-load supply does not appear in interconnection records at all, because it does not use the transmission network. On-site generation — gas turbines predominantly, with a smaller contribution from fuel cells and co-located renewables with storage — allows a facility to begin operating years before its grid connection is delivered.

We estimate this at approximately 11 per cent of new large-load supply, with wide uncertainty. It is difficult to measure because it falls between regulatory regimes: too small for generation licensing in many jurisdictions, not part of the load record because it never crosses the meter. Analysis that relies on grid statistics alone will progressively understate compute capacity as this share grows.

The policy instruments available here are well understood and largely unused: queue reform with financial commitment requirements at entry, cluster-based rather than serial study processes, flexible-load tariffs that price curtailable connections differently, and advance transmission build in anticipated corridors. Each has been implemented somewhere. None has been implemented at scale in the markets facing the largest queues.