The constraint on British compute is electricity

Between May 2024 and August 2025, at least 9 gigawatts of projects sitting in the electricity transmission queue changed their declared technology from battery storage to data centre.

They did not move. They did not get built. Somebody amended a form, because the value of the queue position was higher under the new label. That single detail, published by Ofgem in a consultation on 29 July 2026, explains more about British compute capacity than any investment announcement of the last two years.

The queue is bigger than the country

The numbers are worth setting down carefully because they are routinely reported as demand forecasts, which they are not.

Ofgem’s demand connections update of 6 November 2025 recorded contracted offers in the demand queue rising from 41 GW in November 2024, split 17 GW transmission and 24 GW distribution, to 125 GW in June 2025, split 97 GW transmission and 29 GW distribution. A tripling in seven months, almost entirely at transmission level.

The Curate consultation of 29 July 2026 attributes roughly 73 GW of that to data centres, across about 315 projects ranging from 1 MW to 1,500 MW apiece. The concentration is severe: 166 projects between 100 and 500 MW account for 36,632 MW, and 40 projects above 500 MW account for a further 31,408 MW, so 93% of the queued data centre capacity sits in schemes of 100 MW or more. Peak electricity demand across the whole of Great Britain in 2025/26, a single winter, was 45 GW.

Then the arithmetic that ought to end the argument. At the average capital cost Ofgem uses of £9.5m per megawatt, the queued pipeline implies £693bn of capital expenditure, which the regulator notes is around 23% of United Kingdom gross domestic product in 2025. Bloomberg’s estimate of global artificial intelligence infrastructure spending by Alphabet, Amazon, Meta and Microsoft combined, for 2026, is about $650bn.

Nobody is going to build £693bn of data centres in Britain. Ofgem’s position, stated plainly, is that the volume of projects now far exceeds even the most ambitious demand forecasts.

Why the queue stopped being information

The reason is a design decision, not a conspiracy, and Ofgem describes it without embarrassment. Under the reformed connections process, generation and storage projects must show they are both ready and needed. Demand projects only have to show they are ready, because all demand is treated as needed. The regulator’s own assessment is that the absence of a needed filter has unintentionally led to excessive growth in the demand queue.

Underneath that sits an incentive structure the Curate document sets out with unusual candour. Entering the queue is cheap, historically costing little more than the studies and administration involved. Land carrying both planning permission and a grid connection is worth a great deal more than land carrying neither, so securing a connection is itself a value-creating act regardless of whether anything is ever built. Developers holding portfolios hedge across land, planning and connection risk by making more applications than they intend to deliver, which increases scarcity, which raises the value of a connection, which encourages more applications.

Ofgem identifies five kinds of party in the queue, and only two of them are the ones most readers picture. Hyperscalers and co-location providers, yes. But also energy and infrastructure developers packaging combined power and compute propositions, powered land developers whose entire business is assembling land, planning and connection into an investment product for somebody else to complete, and independent landowners doing a smaller version of the same thing. For a developer who never intends to build, the regulator notes, only one project in a portfolio needs to sell for the strategy to pay.

Set against that, the 9 GW of batteries that became data centres reads less like an anomaly and more like the system working exactly as its incentives instruct.

What Ofgem is proposing to do about it

The Curate package, out for consultation until 16 September 2026, has two components and both are aimed squarely at the economics above.

A data centre commitment fee, applying to projects above 40 MW, secured from the point a connection offer is accepted until the site energises. Ofgem proposes setting it between 2.5% and 7.5% of average data centre capital expenditure, which works out at roughly £237,500 to £712,500 per megawatt of requested capacity. It is returned if the project proceeds and forfeited if it terminates or breaches the rules. On a 200 MW scheme the low end is £47.5m of committed security and the high end is £142.5m.

Alongside it, data centre queue management milestones requiring evidence of progression at defined stages: a credible compute customer or end user, procurement of long-lead electrical equipment, financial capability and technical readiness, with separate pathways for facilities the developer will operate itself and those intended for lease or sale. Projects that cannot produce the evidence lose their position.

Read the first milestone again, because it is the interesting one. Demonstrating a credible end user makes speculative capacity reservation structurally difficult in a way no fee alone would, since it requires the existence of a counterparty. Ofgem’s own analysis is that there are many parties developing data centre projects across the value chain and a comparatively small number of compute offtakers. Milestones of that kind do not slow the queue down. They reveal how much of it was ever attached to demand.

The price problem the reforms do not touch

Connection is one constraint. Cost per unit is the other, and it points the same way.

The House of Commons Library’s briefing on data centres, published on 27 May 2026, assembles the comparison from official statistics. On the Department for Energy Security and Net Zero’s figures, average United Kingdom industrial electricity prices including tax in 2024 were 26.6 pence per kilowatt hour, 63% above the median across International Energy Agency member countries and higher than any other member with data. Pre-tax prices were also the highest, at 94% above the IEA average. In the first half of 2025, British industrial prices including taxes and levies exceeded those anywhere in the European Union at every consumer size, and for the largest consumers were more than double the EU median.

The briefing also records that data centres are excluded from the British Industrial Competitiveness Scheme, which the government expects to cut electricity costs for eligible sectors by £35 to £40 per megawatt hour, and from the existing energy intensive industries schemes, despite being among the largest single consumers in the country.

What data centres get instead is geographically conditional. The Delivering AI Growth Zones paper of 13 November 2025 offers a discount for facilities that locate where generation is constrained and reduce system costs. For a 500 MW site that is worth up to £24 per megawatt hour in Scotland, up to £16 in Cumbria and up to £14 in the north east of England. Nothing elsewhere.

Britain has therefore built a policy that says compute is nationally important, prices its electricity above every comparable economy, withholds the industrial relief given to other heavy users, and offers a discount only in three regions where the grid happens to need the load. That is a coherent set of choices. It is not an accident. But an operator weighing Slough against Stockholm is weighing the whole of it, and the same paper concedes the primary point: it calls timely grid connections the single biggest blocker for establishing AI Growth Zones.

What follows for anyone buying capacity

The practical consequence is that capacity planning in Britain now has a dependency most procurement functions have never modelled.

If a colocation contract or a cloud region commitment assumes British capacity at a stated price in 2029, that assumption rests on a specific site holding a specific queue position, and the rules governing that position are being rewritten this year. A supplier whose pipeline consists largely of sites without a compute offtaker is about to find its milestones harder to meet and its committed security materially more expensive. Some of that cost will appear in renewal pricing. Some of it will appear as a delivery date moving.

The questions worth asking of a supplier are therefore not the ones on the standard due diligence template. Which of your sites hold accepted connection offers rather than applications. What is the contracted energisation date, and has it moved since the offer was made. Are the sites above 40 MW, which brings them into scope of the proposed fee. And is the capacity you are selling me inside an AI Growth Zone, because that determines both the electricity price and the priority.

None of those is a technology question, and none of them can be answered by the account team. That is the shift. Where British compute gets built, and what it costs to run, is now decided in an energy regulator’s consultation timetable rather than in a datacentre design review. Whether the workloads meant to fill this capacity belong on rented infrastructure at all is a separate question, taken apart in hybrid stopped being the thing you settled for, and this desk’s other work sits under cloud and infrastructure.