Data centre electricity demand and the UK grid — how much power will AI need by 2030, and can generation keep up?

Abstract

1. Introduction

Electricity demand from data centres has moved in five years from a specialist topic to a question asked by energy ministers, grid operators and local councils. The immediate cause is generative AI: the servers that train and run large models draw far more power per rack than the web and email servers of the previous decade, and the companies deploying them announce sites measured in hundreds of megawatts rather than tens. In the UK the question has a sharper edge because it coincides with a national plan to run the electricity system on at least 95 per cent clean power by 2030 (DESNZ, 2024), a connection queue that had grown to more than 700 GW (NESO, 2025b), and a government strategy that wants large AI data centres built here, in AI Growth Zones each served by at least 500 MW (DSIT, 2025d).

Public debate on this subject suffers from two failure modes. The first is alarm: treating every announced project as certain and every megawatt of connection request as demand that will materialise. The second is complacency: the observation, correct for the 2010s, that efficiency gains kept data centre energy use nearly flat despite enormous growth in computing (Masanet et al., 2020), extended into an assumption that the same will happen again. Neither position survives the evidence now available, which includes, for the first time, metered UK consumption data (DESNZ, 2026b) and a post-reform picture of the connection pipeline (NESO, 2025b).

This paper addresses four questions. What is the credible range of data centre electricity demand growth to 2030, worldwide and in the UK? How does it compare with generation and network capacity plans? What does the evidence say about efficiency, siting, grid connection queues and emissions? And what follows for UK policy and for UK SMEs choosing cloud services? Its contribution is not new primary data but a reconciliation of sources that are often quoted against each other without noting that they measure different things, together with a worked UK projection whose every assumption is stated and can be changed by the reader.

Section 2 reviews the literature by theme. Section 3 sets out the method and the model. Section 4 presents findings with figures and tables. Section 5 discusses the implications, including a sub-section for SMEs. Section 6 states the limitations and Section 7 concludes. A companion article covers the per-query figures for individual AI prompts in plainer terms (How much electricity does AI use?); this paper concentrates on the system level.

2. Literature review

2.1 Global estimates and the efficiency debate

The modern baseline for global data centre energy use is the work of Masanet et al. (2020), who argued in Science that widely circulated extrapolations of runaway growth were wrong, and that "growth in energy use has slowed owing to efficiency gains that smart policies can help maintain in the near term". Their integrated bottom-up model found that massive efficiency gains had kept energy use roughly flat over the preceding decade even as demand for data grew rapidly (Northwestern University, 2020). The paper was influential precisely because it corrected a tendency to over-forecast, and any new projection has to explain why the next decade should differ from the last.

The IEA's Energy and AI report is the most comprehensive attempt to do so. It estimates that data centres consumed around 415 TWh in 2024, about 1.5 per cent of world electricity, having grown at roughly 12 per cent a year since 2017, more than four times faster than total electricity consumption (IEA, 2025a). The United States accounted for 45 per cent of that, China for 25 per cent and Europe for 15 per cent. In the IEA Base Case consumption more than doubles to around 945 TWh by 2030, "just under 3% of total global electricity consumption", growing at about 15 per cent a year (IEA, 2025b). The difference from the 2010s is attributed to accelerated servers, the graphics processing unit (GPU) and similar hardware used for AI, whose consumption grows at 30 per cent a year in the Base Case against 9 per cent for conventional servers; accelerated servers account for almost half of the net increase (IEA, 2025b). The report also publishes sensitivity cases for 2035: a Lift-Off Case exceeding 1,700 TWh (4.4 per cent of global demand), a High Efficiency Case of around 970 TWh (2.6 per cent) and a Headwinds Case in which demand plateaus at about 700 TWh (under 2 per cent), around a Base Case of about 1,200 TWh (IEA, 2025a; 2025b).

The IEA's shorter-horizon market reports are consistent with this. Electricity 2025 noted that data centres in China alone consumed over 100 TWh in 2024 and could double that by 2027, while stressing that "projections indicate a wide range of uncertainties" (IEA, 2025d). Electricity 2026 forecasts global demand growing 3.6 per cent a year over 2026 to 2030, with roughly half of the United States' increase driven by data centres and EU demand rising about 2 per cent a year (IEA, 2026). Ember's review of 2025 found global demand up 2.8 per cent, or 849 TWh, with solar alone meeting three-quarters of the increase and power-sector carbon intensity falling to 458 gCO2e/kWh (Ember, 2026). Data centre growth therefore matters less for the global system than for the specific grids where it concentrates.

2.2 The United States as the leading indicator

The United States is where AI demand has arrived first and where the best bottom-up data exist. LBNL's 2024 report for the Department of Energy found that US data centre consumption was stable at about 60 TWh between 2014 and 2016, reached 76 TWh (1.9 per cent of US electricity) in 2018 and 176 TWh (4.4 per cent) in 2023, and projects 325 to 580 TWh in 2028, equal to 6.7 to 12 per cent of national consumption (Shehabi et al., 2024; LBNL, 2025). The authors are explicit that the rapid growth in accelerated servers caused total demand to more than double between 2017 and 2023, which is the empirical break from the flat 2010s that Masanet et al. (2020) could not have foreseen.

EPRI's updated scenarios go further. Taking 2024 consumption as 177 to 192 TWh (4 to 5 per cent of US electricity), they project 380 to 790 TWh by 2030, or 9 to 17 per cent, with nominal IT capacity of 56 to 132 GW (EPRI, 2026). EPRI notes that these figures are about 60 per cent higher than its own 2024 projections, but adds a caution that runs through this paper: announced capacity "should be treated as a pipeline indicator rather than a near-term peak forecast", because non-IT loads, ramp-up, load shapes, on-site generation and flexibility all affect what the grid actually sees (EPRI, 2026).

2.3 Europe, Ireland and the data gap

European evidence is thinner. The European Commission's Joint Research Centre estimated EU data centre consumption at 45 to 65 TWh in 2022, 1.8 to 2.6 per cent of EU electricity, noting "a lack of official statistics on the energy use of digital infrastructure" (Kamiya and Bertoldi, 2024). The IEA expects European consumption to grow by more than 45 TWh, or 70 per cent, between 2024 and 2030 (IEA, 2025b).

Ireland is the exception, and the cautionary tale. The Central Statistics Office publishes metered consumption: data centres took 5 per cent of Ireland's metered electricity in 2015, 11 per cent in 2020, 22 per cent in 2024 and 23 per cent in 2025, when consumption rose 10 per cent in a year to 7,663 GWh while all other users grew 2 per cent (CSO, 2026). Ireland shows what an unconstrained cluster does to a small grid, and it is the reference point against which UK policy is implicitly being designed.

2.4 The UK evidence base

Until June 2026 the UK had no official measurement of data centre electricity use. The Electricity System Operator's 2022 briefing put consumption at about 3.6 TWh in 2020, expected demand to reach just under 6 per cent of national consumption by 2030 through a near doubling of colocation power and a more than tenfold increase in hyperscale capacity, and modelled a range rising to as much as 35 TWh by 2050, with growth concentrated in London (National Grid ESO, 2022).

DESNZ's special feature in Energy Trends changed the picture by matching a list of 2,423 colocation and hyperscale sites to meter-level data. It found consumption of 4.5 TWh in 2024, 2 per cent of the 249.2 TWh drawn from the grid in Great Britain, up from 4.1 TWh in 2023 and 1.3 TWh (41 per cent) higher than in 2020 (DESNZ, 2026b). The geography is extraordinarily concentrated: the South East took 1.8 TWh (40 per cent), of which Slough alone accounted for 1.3 TWh, and London 1.7 TWh (37 per cent); in Slough, data centres consumed 65 per cent of all grid electricity. The article notes that its estimate is well below NESO's 7.6 TWh figure for 2023, because it measures electricity actually consumed rather than capacity-based assumptions, and that enterprise data centres inside ordinary businesses are excluded (DESNZ, 2026b).

NESO's written evidence to Parliament provides the forward view. It expects around 5.2 GW of connected data centre capacity and "just over 20 TWh of demand by 2030", within a modelled range of 3.7 to 6.3 GW, and reports 72.8 GW of connection requests through to 2039, with nearly 59 GW awaiting connection at transmission level and 40 of 173 queued projects sized between 500 and 1,500 MW (NESO, 2025a). It assumes an average ramp of seven years from connection to full demand, notes that the average British data centre in 2025 was 12 MW, and stresses that "growth projections for data centres remain highly uncertain" (NESO, 2025a).

2.5 Efficiency: PUE, chips and workloads

Power usage effectiveness (PUE) is the ratio of total facility energy to the energy delivered to IT equipment; a PUE of 2.0 means a watt of overhead for every watt of computing (Google, 2025). The Uptime Institute's surveys show an industry average of 2.50 in 2007, 1.55 by 2014 and 1.56 in 2024 (Uptime Institute, 2024), then 1.54 in 2025, "the sixth consecutive year that this headline figure has virtually stood still" (Uptime Institute, 2025). Newer facilities do better: sites commissioned within five years average 1.48, and many recent builds achieve 1.3 or lower (Uptime Institute, 2024; 2025). The hyperscalers are in a different league: Google reports a fleet-wide PUE of 1.09 for 2025 and over three times more computing performance per unit of energy than five years ago (Google, 2025); Microsoft reports 1.17 (Data Center Dynamics, 2026). The IEA gives cooling a share of about 7 per cent in efficient hyperscale sites but over 30 per cent in less efficient enterprise facilities (IEA, 2025b).

The literature on AI workloads themselves is younger and more contested. De Vries (2023) argued in Joule that inference, not training, would dominate, and estimated that AI-related electricity consumption could reach 85 to 134 TWh a year by 2027, with a hypothetical scenario in which every Google search used generative AI requiring about 29.2 TWh (de Vries, 2023; ScienceDaily, 2023). Luccioni, Jernite and Strubell (2024) measured 88 models across 10 tasks and found energy per 1,000 inferences ranging from about 0.002 kWh for text classification to about 2.9 kWh for image generation, and that "using multi-purpose models for discriminative tasks is more energy-intensive compared to task-specific models". Patterson et al. (2021), writing from Google, showed that choices of model architecture, hardware, data centre efficiency and location can together cut the energy and carbon of training by a factor of 100 to 1,000. These strands are not contradictory: per-task energy can fall steeply while total energy rises, if usage grows faster still.

2.6 Grid connections, siting and policy

The UK connection queue is the most studied bottleneck. Ofgem introduced queue-management milestones in November 2023 when the queue was nearly 400 GW (Ofgem, 2023); by its April 2025 decision on the reform package, new applications had reached 444 GW in 2023/24 alone, connection rates were "2–3 times slower than the rate needed" for the 2030 target, and the regulator adopted a "first ready and needed, first connected" principle in which demand-only projects are automatically deemed needed (Ofgem, 2025). The Clean Power 2030 Action Plan put the queue at 739 GW and estimated that around twice as much transmission infrastructure would be needed by 2030 as had been built in the previous decade, at about £40 billion a year (DESNZ, 2024). Its connections annex states that demand connections will not be constrained by the clean power pathways and calls transmission-connected demand projects such as data centres "by nature strategically important" (DESNZ, 2025b). NESO's implementation reduced a queue of over 700 GW to a prioritised pipeline of 283 GW, of which 132 GW aligns with 2030 and 151 GW with 2035, alongside around 100 GW of new or expanded transmission demand connections, against existing connected generation of about 111 GW (NESO, 2025b; 2025c).

AI-specific policy is recent. The AI Opportunities Action Plan committed to expanding sovereign compute twentyfold by 2030 and to AI Growth Zones with streamlined planning and accelerated clean power, piloted at Culham (DSIT, 2025a). The AI Energy Council followed in April 2025 (DSIT, 2025b), and by June 2025 ministers were pairing the twentyfold compute target with connection reforms said to be capable of freeing more than 400 GW from the queue (DSIT, 2025c). The Compute Roadmap in July 2025 set the target of "at least 6GW of AI-capable data centre capacity by 2030" with zones each serving at least 500 MW and at least one exceeding 1 GW (DSIT, 2025d), and the Delivering AI Growth Zones paper of November 2025 introduced connection reallocation and reservation mechanisms, developer-built high-voltage infrastructure, a Connections Accelerator Service and electricity discounts of up to £24/MWh in Scotland, £16/MWh in Cumbria and £14/MWh in the North East from April 2027 (DSIT, 2025f). The accompanying ministerial statement claimed the programme could unlock up to £100 billion of private investment (UK Parliament, 2025). Five zones had been designated by January 2026 (DSIT, 2026a; 2026b). Internationally, the IEA warns that "grid constraints could delay around 20% of global data centre capacity planned for construction by 2030" and that several jurisdictions have imposed moratoriums while operators process backlogs (IEA, 2025c).

2.7 Gaps

Three gaps stand out. First, the UK still lacks measured data on enterprise data centres and on the load factor of connected capacity. Second, no published UK projection reconciles the metered baseline with the connection pipeline and a stated attrition rate. Third, the emissions consequence of siting AI capacity in Britain rather than elsewhere is rarely quantified with current grid intensities. The method below addresses the second and third.

3. Method

3.1 Design and data

This is desk research: a secondary analysis of published sources, all retrieved on 24 and 25 August 2026 and listed in the references. Numerical claims are taken only from primary or official sources (IEA, DESNZ, NESO, Ofgem, gov.uk, CSO, LBNL, EPRI, peer-reviewed papers, corporate reports and the Uptime Institute survey), with trade press used only to confirm corporate figures not on a primary page at the time of writing. Sources were checked for definitional consistency before any figure was placed alongside another: metered consumption (DESNZ), modelled demand (NESO), connection capacity (GW) and IT capacity are four different quantities and are labelled as such throughout.

3.2 The basic energy identity

The annual electricity of a data centre follows from its IT load, its overhead and how hard it runs:

Here E is annual energy in megawatt-hours, P (subscript IT) is the installed load of the computing equipment in megawatts, PUE is power usage effectiveness (facility energy divided by IT energy), LF is the load factor (average draw as a fraction of installed load) and 8,760 is the hours in a year. A 500 MW IT site at a PUE of 1.2 and a load factor of 0.8 uses 500 × 1.2 × 0.8 × 8,760 = 4,204,800 MWh, or 4.2 TWh, a year: about 1.3 per cent of UK electricity demand in 2025 from a single site (DESNZ, 2026a).

3.3 Growth with efficiency

To separate demand for computing from demand for electricity, the paper uses a compound-growth identity:

E with subscript t is energy in year t, E with subscript 0 the base-year energy, g the annual growth rate of computing work performed, and ε the annual improvement in computing work per unit of energy. Electricity grows only when g exceeds ε. The identity is used in Section 4.3 to infer what growth in underlying computing demand is implied by the IEA's server projections and Google's reported efficiency gains.

3.4 The UK projection model

The UK projection is built from connected capacity, not announced projects, because NESO gives an expected connected capacity for 2030 with a range, whereas announcements are only a pipeline indicator (EPRI, 2026). Annual energy is:

E is annual energy in 2030 in terawatt-hours, C the grid connection capacity of each fleet segment i in gigawatts, LF the load factor of that segment relative to its connection capacity, the sum runs over the n segments, and 8.76 converts gigawatts at full load to terawatt-hours a year. The load factor absorbs PUE, because connection capacity is facility capacity, and ramp-up, because NESO assumes seven years from connection to full demand (NESO, 2025a). Three scenarios and one trend extrapolation are defined in Section 4.4, with every input in Table 2. The load factors are the paper's own assumptions, chosen so that the central scenario reproduces NESO's published figure and the others bracket it.

3.5 Emissions

Emissions are calculated as energy multiplied by grid carbon intensity, using 126 gCO2/kWh for the UK in 2025 (Carbon Brief, 2026) and 458 gCO2e/kWh as the 2025 global power-sector average (Ember, 2026). No allowance is made for corporate renewable power purchase agreements, which change accounting but not the physical grid in the hour of consumption.

4. Findings

4.1 The global range to 2030 and 2035

Figure 1 plots the IEA scenarios and the LBNL range for the United States. The central global path is a doubling in six years, and the width of the 2035 fan, 700 to 1,700 TWh, is the honest statement of uncertainty. The High Efficiency Case at 970 TWh shows that even aggressive efficiency only flattens the curve rather than reversing it (IEA, 2025b).

Table 1 sets the principal estimates side by side, with the quantity each measures. The numbers are not rivals: the LBNL and EPRI ranges for the United States are consistent with the IEA's statement that the United States accounts for nearly half of global growth to 2030 (IEA, 2025a), and the Irish series shows what the top of the range looks like on a small system.

4.2 The UK: demand against generation and connections

Figure 2 places the UK data centre numbers next to the system they sit in. In 2025 Great Britain generated 293.6 TWh against demand of 323 TWh; renewables supplied 153.0 TWh (52.1 per cent), fossil fuels 94.9 TWh (almost all gas, 31.8 per cent), nuclear 35.9 TWh, the lowest since the 1980s, and net imports 29.7 TWh (DESNZ, 2026a). Demand rose 1.2 per cent, the first time in two decades it had risen two years running, but remained below 2019 (DESNZ, 2026a; Carbon Brief, 2026). The year before, demand had been 319.0 TWh and renewables had exceeded half of generation for the first time, at 50.4 per cent, while net imports reached a record 33.4 TWh (DESNZ, 2025a). The trend has continued into 2026: in the three months to May, renewables provided 53.1 per cent of major power producers' generation and the low-carbon share reached 69.8 per cent (DESNZ, 2026c). Against that background, 4.5 TWh of metered data centre consumption in 2024 is small, and even the central 2030 scenario of about 20 TWh is comparable with the year-on-year swings in gas generation.

The right-hand panel is where the problem lives. The 72.8 GW of data centre connection requests through to 2039 is almost the same size as the entire de-rated generation fleet of 74.5 GW (NESO, 2025a; DESNZ, 2026a). Nearly 59 GW of that is at transmission level, almost all in projects of 100 MW or more (NESO, 2025a). NESO's expected connected capacity of 5.2 GW by 2030 implies that around 93 per cent of the requested capacity will not be connected by then: some will connect later, much will never be built, and some is the same demand applied for at several sites. The post-reform generation pipeline of 283 GW, with about 100 GW of transmission demand connections alongside it (NESO, 2025b), makes the same point: the system's ability to build substations, transformers and lines, not its ability to generate energy over a year, limits how fast AI capacity can arrive.

Two further observations follow from the metered data. Existing capacity is where the network is already tightest: Slough and Hillingdon absorbed 1.1 TWh of the 1.3 TWh growth between 2020 and 2024, and data centres now take almost two-thirds of Slough's electricity (DESNZ, 2026b). And NESO's analysis suggests that at most about 20 per cent of future demand could be located in Scotland (NESO, 2025a), which bounds how far moving capacity to where the power is can relieve the South East.

4.3 Efficiency: what the numbers actually show

Figure 3 shows why PUE has stopped being the main efficiency story. The industry average fell from 2.50 in 2007 to the mid-1.5s by 2014 and has barely moved since (Uptime Institute, 2024; 2025). Remaining gains come from replacing old sites with new ones, because in legacy facilities "further upgrades often present a much weaker business case" (Uptime Institute, 2025). The hyperscale figures of 1.09 and 1.17 (Google, 2025; Data Center Dynamics, 2026) show what is possible when a building is designed around the computing.

The larger efficiency lever is inside the IT load. Google's threefold gain in computing performance per unit of energy over five years (Google, 2025) is, by Equation 2, an efficiency improvement ε of about 24.6 per cent a year. Yet the IEA projects accelerated-server electricity growing at 30 per cent a year (IEA, 2025b). Substituting into Equation 2, the implied growth in computing work is (1.30 × 1.246) − 1, or roughly 62 per cent a year. That arithmetic, which is the paper's own, is the quantitative form of the rebound argument made by de Vries (2023): efficiency at Google's pace is being outrun by demand growing more than twice as fast. Luccioni, Jernite and Strubell's (2024) finding that general-purpose models are far more energy-intensive than task-specific ones for the same job suggests that much of that demand is discretionary rather than inherent.

Figure 4 summarises where the energy goes in a facility and marks the two efficiency frontiers, overhead and IT.

4.4 The worked UK projection

Table 2 gives the results. Scenario 0 is the counterfactual in which AI adds nothing beyond the 2020 to 2024 trend: 4.5 TWh growing at 9 per cent a year reaches 7.5 TWh in 2030. Scenario A takes NESO's low capacity of 3.7 GW at a load factor of 0.35, reflecting sites connected but not yet filled, and gives 11.3 TWh. Scenario B takes NESO's expected 5.2 GW at 0.45 and gives 20.5 TWh, matching NESO's own figure, which implies a load factor of 20 ÷ (5.2 × 8.76) = 0.44. Scenario C assumes the government's 6 GW AI-capable target is met on time and energised, adds an assumed 1.5 GW for the existing colocation estate, and applies 0.55; it gives 36.1 TWh, about 11 per cent of present demand.

Three points about these results. First, the range is wide but bounded: even Scenario C leaves Britain at half of Ireland's present share (CSO, 2026), and the ESO's 2022 expectation of just under 6 per cent by 2030 (National Grid ESO, 2022) sits almost exactly on Scenario B. Second, the difference between scenarios is mostly about connection and ramp, not about how much computing people want; the load factor matters as much as the capacity, and the load factor is the number nobody publishes. Third, on annual energy alone all three scenarios are absorbable: renewable generation rose about 6 per cent in 2025 to a record 152 TWh on Carbon Brief's measure (Carbon Brief, 2026), an increase of roughly 9 TWh in a single year, so the whole of Scenario B's increment above today's consumption is equivalent to about two years of recent renewable growth, and solar output alone rose 34 per cent to 20.1 TWh (DESNZ, 2026a). What is not automatically absorbable is 5 GW of new, flat, round-the-clock demand appearing in particular places.

4.5 Emissions

Data centre emissions are modest at the global level. The IEA puts them at about 180 million tonnes of CO2 today, rising to 300 million tonnes in its 2035 Base Case and 500 million tonnes in the Lift-Off Case, in each case below 1.5 per cent of energy-sector emissions (IEA, 2025a). The location of demand matters more than its size. Scenario B's 20.5 TWh at Britain's 2025 intensity of 126 gCO2/kWh (Carbon Brief, 2026) is about 2.6 million tonnes a year; the same energy at the 2025 global average of 458 gCO2e/kWh (Ember, 2026) would be about 9.4 million tonnes. Britain's grid, which ran at a record 98.8 per cent zero-carbon for a half-hour on 22 April 2026 and, on NESO's transmission-system measure, took 44 per cent of its electricity from renewables in 2025 (NESO, 2026), is therefore among the better places to put computing, provided the demand does not itself hold gas plant on the system. That proviso is real: gas generation rose 5 per cent in 2025 and grid intensity edged up from 124 to 126 gCO2/kWh, because demand grew while nuclear output fell (Carbon Brief, 2026; DESNZ, 2026a).

The corporate reports illustrate both sides. Google's data centre electricity rose 27 per cent in 2024 and its total electricity demand 37 per cent in 2025, yet it reported data centre energy emissions down 12 per cent in 2024 and operational emissions down 2 per cent in 2025 on the back of 8 GW and then 12 GW of new clean-energy contracts, while supply-chain emissions rose 22 and then 25 per cent (Data Center Dynamics, 2025; Google, 2026). Microsoft's electricity consumption rose from 23.6 TWh in its 2023 financial year to 29.8 TWh in 2024 (Microsoft, 2025), and its total reported emissions rose 25 per cent to 20.29 million tonnes in the following year, partly because it stopped counting unbundled renewable certificates (Data Center Dynamics, 2026). Readers of such reports should distinguish operational grid emissions, which clean-power contracts can offset in annual accounting, from embodied emissions in chips and buildings, which they cannot.

4.6 Policy instruments in place

Table 3 lists the UK measures that bear directly on the findings above. Their common feature is that they act on connection and location rather than on annual energy, which is the correct target.

5. Discussion

5.1 How much power will AI need?

The credible answer for 2030 is a global data centre total of about 945 TWh, within a range that widens to 700 to 1,700 TWh by 2035, with AI accelerators responsible for almost half of the increase (IEA, 2025a; 2025b). For Great Britain it is 11 to 36 TWh of annual consumption by 2030, most likely around 20 TWh, from 4.5 TWh today: a move from 2 per cent to roughly 6 per cent of demand, which is what the system operator expected in 2022 before the AI boom (National Grid ESO, 2022) and still expects now (NESO, 2025a). The AI boom has not so much raised the UK's expected 2030 demand as changed what it consists of: fewer, larger, denser sites, seeking power outside London and the Thames Valley.

This finding sits between the two positions described in the introduction. The complacent reading, that efficiency will flatten demand as it did in the 2010s, fails because the United States data show that the flat decade ended in 2017 (Shehabi et al., 2024) and because the implied growth in computing demand now exceeds even hyperscale efficiency gains by a wide margin (Section 4.3). The alarmed reading fails because it confuses the 72.8 GW of applications with demand: NESO's expected 5.2 GW is a fourteenth of it, EPRI's advice to treat announcements as a pipeline indicator applies here (EPRI, 2026), and even the government's 6 GW target, fully delivered, produces a share of demand half of Ireland's today.

5.2 Can generation keep up?

On energy, yes. Renewable output in Great Britain reached 153.0 TWh in 2025 and the Clean Power 2030 ranges of 43 to 50 GW of offshore wind, 27 to 29 GW of onshore wind and 45 to 47 GW of solar (DESNZ, 2024) imply annual additions that dwarf 20 TWh. On capacity and location, the answer is conditional on three things. The first is transmission build: the Action Plan's requirement for twice the previous decade's transmission construction by 2030 (DESNZ, 2024) was set before AI Growth Zones added 500 MW to 1 GW loads in Blyth, Culham, Wales and Lanarkshire, and the North East alone is described as needing a 1.1 GW increase in energy capacity within six years (DSIT, 2025e). The second is the shape of the demand. Data centres run flat; the Clean Power target is defined on annual energy and a 95 per cent generation share (House of Commons Library, 2026), and flat demand on winter evenings is served at the margin by gas, which is why intensity rose in 2025 despite record renewables (Carbon Brief, 2026). NESO's suggestion that non-firm connections "could be acceptable if it allows data centres to connect and become operational sooner" (NESO, 2025a) and the FES observation that cooling load can shift in time (NESO, 2025d) point to flexibility as part of the answer, but AI training loads are not easily interrupted and evidence that operators will accept curtailment is thin. The third is that the reformed queue works as intended. Ofgem's finding that connection rates were two to three times too slow (Ofgem, 2025) and NESO's report of up to ten-year waits (NESO, 2025c) describe a system that was failing before the AI applications arrived; the December 2025 pipeline is a plan, not a delivery record, and the next application window does not open until the second half of 2026 (NESO, 2025b).

5.3 Comparison with the literature

The paper's UK projection is more conservative than the impression given by connection-request figures and more aggressive than trend extrapolation, and it is consistent with the two official projections that exist (National Grid ESO, 2022; NESO, 2025a). Its global framing follows the IEA. It departs from parts of the literature in treating PUE as largely exhausted as a lever for the existing estate, following the Uptime Institute's six years of flat averages (Uptime Institute, 2025), and in giving weight to the rebound argument of de Vries (2023) on the strength of the arithmetic in Section 4.3. The strongest counter-argument to the paper's own conclusion is the one made by Masanet et al. (2020) for the previous decade and embodied in the IEA's High Efficiency and Headwinds Cases: that projections made during a build-out systematically overshoot, because they extrapolate from a period of catch-up investment; that chip and model efficiency will improve faster than assumed; and that much announced capacity will be cancelled when the economics of AI services are tested. EPRI's 60 per cent upward revision in two years shows how fast such projections move, and they can move down as well as up. If that view is right, Scenario 0 or A is the outcome, UK data centre demand stays below 4 per cent of the total, and the policy risk becomes stranded grid investment rather than shortage. The paper's response is that even then the connection problem is real for the sites that are built, and that the reservation and reallocation mechanisms in the growth-zone policy (DSIT, 2025f) are the right kind of instrument because they can be scaled back without loss if demand disappoints.

5.4 Implications for UK policy

Five implications follow. First, measure. The DESNZ metered series is the most important UK contribution to this field in years; extending it to enterprise sites and publishing load factors for connected capacity would remove the two largest uncertainties in Table 2. Second, price location, not just energy. The zone discounts of £14 to £24/MWh (DSIT, 2025f) are fixed subsidies; a locational signal reflecting the real cost of serving flat demand in the South East against surplus wind in the North would do the same job more cheaply. Third, be candid that annual clean-power accounting and hourly physics differ: a data centre with a 100 per cent renewable contract still raises gas burn on a still winter evening, and the hourly carbon-free share that Google now reports, 66 per cent in 2024 (Data Center Dynamics, 2025), is the more honest metric to require of operators. Fourth, protect the 95 per cent target from the zones rather than the reverse: the Compute Roadmap's interest in behind-the-meter and advanced nuclear supply (DSIT, 2025d) makes sense only if it adds capacity rather than diverting it. Fifth, plan for attrition. If NESO is right that 93 per cent of requested capacity will not be connected by 2030, the reservation mechanism needs expiry and clawback, or it becomes a new form of queue-blocking.

5.5 Implications for UK small and medium-sized businesses

Most UK SMEs will never see the inside of a data centre, but almost every one now rents part of one. The findings bear on that purchase in six practical ways.

Where your data lives is a carbon decision. Computing on the British grid in 2025 carried about 126 gCO2/kWh (Carbon Brief, 2026) against a world average of 458 (Ember, 2026). A firm that reports its emissions, or is asked to by a larger customer, will get a materially lower figure from a UK or northern-European cloud region than from a global default, and should ask its provider for the hourly carbon-free share rather than an annual matching claim.

Moving from the server cupboard to the cloud is usually an efficiency gain. The IEA's cooling shares of about 7 per cent for efficient hyperscale sites against over 30 per cent for enterprise facilities (IEA, 2025b), and the gap between an industry PUE of 1.54 and a hyperscale 1.09 (Uptime Institute, 2025; Google, 2025), mean that the same workload on a shared platform typically needs a fraction of the overhead energy it needs on a rack in a back office. The exception is a lightly used server that could simply be switched off; consolidation, not migration, is the first step. Our cloud services page describes how we approach that assessment.

Expect cloud and AI pricing to carry electricity risk. The policy papers reviewed here treat UK power cost as a barrier serious enough to subsidise (DSIT, 2025f). Providers will pass energy and connection costs through in region pricing and AI service tiers. Fixed prices for two to three years, and the ability to move workloads between regions, are worth more than they were.

Use the right size of model. Luccioni, Jernite and Strubell (2024) found that general-purpose generative models are orders of magnitude more energy-intensive than task-specific ones for routine jobs such as classification. An assistant that reads every incoming email through a large model to sort it costs more, in money and energy, than a small classifier, and the difference compounds. Our note on AI use policies for small firms covers the governance side.

Capacity in the South East is the tightest it has ever been. With data centres taking 65 per cent of Slough's grid electricity (DESNZ, 2026b), new capacity for UK-hosted services will increasingly be built in the North East, Wales and Scotland. For ordinary business applications the distance is irrelevant; for latency-sensitive workloads it is worth asking where a "UK" region physically is.

Resilience still matters. One in ten data centre outages still causes serious or severe disruption (Uptime Institute, 2025), and a grid under connection stress is not a grid with spare margin. Backups held outside the primary provider, described in our 3-2-1 backup guide for households and applied more formally in business IT support, are the SME's hedge against a problem it cannot otherwise influence.

6. Limitations

The paper is a secondary analysis and inherits the limits of its sources. The DESNZ metered series is designated Official Statistics in Development, excludes enterprise data centres, may miss smaller sites and may over-count where a data centre shares a building with offices (DESNZ, 2026b); its 2 per cent share uses grid-supplied consumption of 249.2 TWh as the denominator, whereas the shares in Table 2 use DUKES total demand of 323 TWh, which includes losses and other uses, so the two are not directly comparable. NESO's 20 TWh and 5.2 GW figures rest on a seven-year ramp assumption and on attrition rates that NESO itself calls highly uncertain (NESO, 2025a). The load factors in Table 2 are the author's assumptions, chosen to reproduce NESO's central figure and to bracket it; a reader who believes AI sites will run at 0.8 of their connection from year one will get materially higher numbers from Equation 3. The efficiency arithmetic in Section 4.3 combines a Google fleet figure with an IEA global server projection, which are not the same population. The IEA scenarios date from April 2025 and the EPRI revision shows how quickly such projections age. Corporate PUE and emissions figures are self-reported and use differing boundaries and fiscal years. The full text of Masanet et al. (2020) and de Vries (2023) was not accessible at the time of writing; the paper relies on their published summaries and on university and press summaries of their figures, and quotes no number from either that was not confirmed there. Finally, the paper does not model water use, embodied carbon or the economic value of the computing, all of which bear on whether the demand is worth serving.

7. Conclusion

Data centre electricity demand is growing faster than at any time since official measurement began, and AI accelerators are the reason. The credible global range is a doubling to about 945 TWh by 2030, and 700 to 1,700 TWh by 2035. In Great Britain, metered consumption of 4.5 TWh in 2024 is expected to become between 11 and 36 TWh by 2030, most probably about 20 TWh, or 6 per cent of demand, a share the system operator anticipated before the AI boom and half of what Ireland already carries. Britain can generate that energy; renewable output already exceeds it many times over. What Britain cannot yet do is connect 5 GW of flat, dense, location-specific load quickly, in the places where power is available rather than where fibre and customers are. Every UK policy instrument reviewed here, from queue reform to growth-zone reservations and locational discounts, is aimed at that constraint, which is correct, and none has yet been tested by delivery. Efficiency will continue to improve at the chip and model level, but the arithmetic of the projections implies computing demand growing at more than twice the pace of hyperscale efficiency, so efficiency will moderate the curve rather than flatten it. For small businesses the practical conclusions are undramatic: choose regions on carbon as well as price, consolidate before migrating, buy the smallest model that does the job, and keep a copy of everything somewhere the provider is not.

References

• Carbon Brief (2026) Analysis: UK renewables enjoy record year in 2025 – but gas power still rises. 2 January 2026. Available at: https://www.carbonbrief.org/analysis-uk-renewables-enjoy-record-year-in-2025-but-gas-power-still-rises (accessed 25 August 2026).
• Central Statistics Office (CSO) (2026) Data Centres Metered Electricity Consumption 2025: Key Findings. Dublin: CSO. Available at: https://www.cso.ie/en/releasesandpublications/ep/p-dcmec/datacentresmeteredelectricityconsumption2025/keyfindings/ (accessed 25 August 2026).
• Data Center Dynamics (2025) Google data center power use up 27%, emissions down 17% – report. 30 June 2025. Available at: https://www.datacenterdynamics.com/en/news/google-data-center-power-use-up-27-emissions-down-17-report/ (accessed 25 August 2026).
• Data Center Dynamics (2026) Microsoft reports 25 percent increase in CO2 emissions, on back of data center growth. 10 July 2026. Available at: https://www.datacenterdynamics.com/en/news/microsoft-reports-25-percent-increase-in-co2-emissions-on-back-on-data-center-growth/ (accessed 25 August 2026).
• de Vries, A. (2023) 'The growing energy footprint of artificial intelligence', Joule, 7(10), pp. 2191–2194. doi:10.1016/j.joule.2023.09.004. Record available at: https://research.vu.nl/en/publications/the-growing-energy-footprint-of-artificial-intelligence/ (accessed 25 August 2026).
• Department for Energy Security and Net Zero (DESNZ) (2024) Clean Power 2030 Action Plan: A new era of clean electricity – main report. London: DESNZ. Available at: https://www.gov.uk/government/publications/clean-power-2030-action-plan/clean-power-2030-action-plan-a-new-era-of-clean-electricity-main-report (accessed 25 August 2026).
• Department for Energy Security and Net Zero (DESNZ) (2025a) Digest of UK Energy Statistics 2025: Chapters 1–7 (Chapter 5, Electricity). London: DESNZ. Available at: https://assets.publishing.service.gov.uk/media/68dbe477ef1c2f72bc1e4c4d/DUKES_2025_Chapters_1-7.pdf (accessed 25 August 2026).
• Department for Energy Security and Net Zero (DESNZ) (2025b) Clean Power 2030 Action Plan: connections reform annex (updated April 2025). London: DESNZ. Available at: https://www.gov.uk/government/publications/clean-power-2030-action-plan/clean-power-2030-action-plan-a-new-era-of-clean-electricity-connections-reform-annex (accessed 25 August 2026).
• Department for Energy Security and Net Zero (DESNZ) (2026a) Digest of UK Energy Statistics 2026: Chapter 5, Electricity. 30 July 2026. London: DESNZ. Available at: https://assets.publishing.service.gov.uk/media/6a6a35c50ddb7e4831c629ed/DUKES_2026_Chapter_5.pdf (accessed 25 August 2026).
• Department for Energy Security and Net Zero (DESNZ) (2026b) Energy Trends: June 2026, special feature article – Data centre electricity consumption in Great Britain, 2020 to 2024. 30 June 2026. London: DESNZ. Available at: https://assets.publishing.service.gov.uk/media/6a3e9f25da47783d87723bf2/Data_centre_electricity_consumption_in_Great_Britain__2020_to_2024.pdf (accessed 25 August 2026).
• Department for Energy Security and Net Zero (DESNZ) (2026c) Energy Trends and Prices statistical release: 30 July 2026. London: DESNZ. Available at: https://www.gov.uk/government/statistics/energy-trends-and-prices-statistical-release-30-july-2026/energy-trends-and-prices-statistical-release-30-july-2026 (accessed 25 August 2026).
• Department for Science, Innovation and Technology (DSIT) (2024) Data centres to be given massive boost and protections from cyber criminals and IT blackouts. Press release, 12 September 2024. Available at: https://www.gov.uk/government/news/data-centres-to-be-given-massive-boost-and-protections-from-cyber-criminals-and-it-blackouts (accessed 25 August 2026).
• Department for Science, Innovation and Technology (DSIT) (2025a) AI Opportunities Action Plan. 13 January 2025. Available at: https://www.gov.uk/government/publications/ai-opportunities-action-plan/ai-opportunities-action-plan (accessed 25 August 2026).
• Department for Science, Innovation and Technology (DSIT) (2025b) AI Energy Council to ensure UK's energy infrastructure ready for AI revolution. Press release, 8 April 2025. Available at: https://www.gov.uk/government/news/ai-energy-council-to-ensure-uks-energy-infrastructure-ready-for-ai-revolution (accessed 25 August 2026).
• Department for Science, Innovation and Technology (DSIT) (2025c) Upgrading national grid to power AI future to be tackled at AI Energy Council. Press release, 30 June 2025. Available at: https://www.gov.uk/government/news/upgrading-national-grid-to-power-ai-future-to-be-tackled-at-ai-energy-council (accessed 25 August 2026).
• Department for Science, Innovation and Technology (DSIT) (2025d) UK Compute Roadmap. July 2025. Available at: https://www.gov.uk/government/publications/uk-compute-roadmap/uk-compute-roadmap (accessed 25 August 2026).
• Department for Science, Innovation and Technology (DSIT) (2025e) North East England set for billions in investment and thousands of jobs as UK and US ink tech partnership. Press release, 16 September 2025. Available at: https://www.gov.uk/government/news/north-east-england-set-for-billions-in-investment-and-thousands-of-jobs-as-uk-and-us-ink-tech-partnership (accessed 25 August 2026).
• Department for Science, Innovation and Technology (DSIT) (2025f) Delivering AI Growth Zones. Policy paper, November 2025. Available at: https://assets.publishing.service.gov.uk/media/6915a2609d50fc2fe81616fe/delivering-ai-growth-zones_web_accessible.pdf (accessed 25 August 2026).
• Department for Science, Innovation and Technology (DSIT) (2026a) AI Opportunities Action Plan: One Year On. Available at: https://www.gov.uk/government/publications/ai-opportunities-action-plan-one-year-on/ai-opportunities-action-plan-one-year-on (accessed 25 August 2026).
• Department for Science, Innovation and Technology (DSIT) (2026b) AI Opportunities Action Plan – 2026 Progress: Infrastructure. Available at: https://delivery.ai.gov.uk/4/ (accessed 25 August 2026).
• Electric Power Research Institute (EPRI) (2026) Powering Intelligence: Updated U.S. Data Center Scenarios. Palo Alto, CA: EPRI. Available at: https://restservice.epri.com/publicattachment/97025 (accessed 25 August 2026).
• Ember (2026) Global Electricity Review 2026: 2025 in review. Available at: https://ember-energy.org/latest-insights/global-electricity-review-2026/2025-in-review/ (accessed 25 August 2026).
• Google (2025) Power usage effectiveness – Google Data Centers. Available at: https://datacenters.google/efficiency/ (accessed 25 August 2026).
• Google (2026) Read Google's 2026 Environmental Report. Available at: https://blog.google/company-news/outreach-and-initiatives/sustainability/2026-environmental-report/ (accessed 25 August 2026).
• House of Commons Library (2026) Clean power targets. Research briefing CBP-10182, 5 August 2026. Available at: https://commonslibrary.parliament.uk/research-briefings/cbp-10182/ (accessed 25 August 2026).
• International Energy Agency (IEA) (2025a) Energy and AI: Executive summary. Paris: IEA. Available at: https://www.iea.org/reports/energy-and-ai/executive-summary (accessed 25 August 2026).
• International Energy Agency (IEA) (2025b) Energy and AI: Energy demand from AI. Paris: IEA. Available at: https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai (accessed 25 August 2026).
• International Energy Agency (IEA) (2025c) Energy and AI: AI and energy security. Paris: IEA. Available at: https://www.iea.org/reports/energy-and-ai/ai-and-energy-security (accessed 25 August 2026).
• International Energy Agency (IEA) (2025d) Electricity 2025: Executive summary. Paris: IEA. Available at: https://www.iea.org/reports/electricity-2025/executive-summary (accessed 25 August 2026).
• International Energy Agency (IEA) (2026) Electricity 2026: Executive summary. Paris: IEA. Available at: https://www.iea.org/reports/electricity-2026/executive-summary (accessed 25 August 2026).
• Kamiya, G. and Bertoldi, P. (2024) Energy Consumption in Data Centres and Broadband Communication Networks in the EU. JRC135926. Luxembourg: Publications Office of the European Union. Available at: https://publications.jrc.ec.europa.eu/repository/handle/JRC135926 (accessed 25 August 2026).
• Lawrence Berkeley National Laboratory (LBNL) (2025) Berkeley Lab report evaluates increase in electricity demand from data centers. News release, 15 January 2025. Available at: https://newscenter.lbl.gov/2025/01/15/berkeley-lab-report-evaluates-increase-in-electricity-demand-from-data-centers/ (accessed 25 August 2026).
• Luccioni, A.S., Jernite, Y. and Strubell, E. (2024) 'Power Hungry Processing: Watts Driving the Cost of AI Deployment?', Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT '24), Rio de Janeiro, 3–6 June 2024. doi:10.1145/3630106.3658542. Preprint available at: https://arxiv.org/abs/2311.16863 (accessed 25 August 2026).
• Masanet, E., Shehabi, A., Lei, N., Smith, S. and Koomey, J. (2020) 'Recalibrating global data center energy-use estimates', Science, 367(6481), pp. 984–986. doi:10.1126/science.aba3758. Record available at: https://www.deeprogram.org/library-v2/recalibrating-global-data-center-energy-use-estimates/NCWHFELC (accessed 25 August 2026).
• Microsoft (2025) 2025 Environmental Sustainability Report: Environmental Data Fact Sheet. Available at: https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/msc/documents/presentations/CSR/2025-Microsoft-Environmental-Data-Fact-Sheet-PDF.pdf (accessed 25 August 2026).
• National Energy System Operator (NESO) (2025a) Written evidence submitted by the National Energy System Operator (DCU0081). UK Parliament committee evidence. Available at: https://committees.parliament.uk/writtenevidence/166090/html/ (accessed 25 August 2026).
• National Energy System Operator (NESO) (2025b) NESO implements electricity grid connection reforms to unlock investment in Great Britain. Press release, 8 December 2025. Available at: https://www.neso.energy/neso-implements-electricity-grid-connection-reforms-unlock-investment-great-britain (accessed 25 August 2026).
• National Energy System Operator (NESO) (2025c) Connections Reform Results. Available at: https://www.neso.energy/industry-information/connections-reform/connections-reform-results (accessed 25 August 2026).
• National Energy System Operator (NESO) (2025d) Future Energy Scenarios 2025: Pathways to Net Zero – Executive Summary. Available at: https://www.neso.energy/document/364546/download (accessed 25 August 2026).
• National Energy System Operator (NESO) (2026) Britain's electricity system breaks zero-carbon record, as gas reaches historic low and solar hits historic high. Press release, 24 April 2026. Available at: https://www.neso.energy/britains-electricity-system-breaks-zero-carbon-record-gas-reaches-historic-low-and-solar-hits-historic-high (accessed 25 August 2026).
• National Grid ESO (2022) Data Centres: What are data centres and how will they influence the future energy system? March 2022. Available at: https://www.neso.energy/document/246446/download (accessed 25 August 2026).
• Northwestern University (2020) Data centers use less energy than you think. News release, 27 February 2020. Available at: https://news.northwestern.edu/stories/2020/02/data-centers-use-less-energy-than-you-think/ (accessed 25 August 2026).
• Ofgem (2023) Ofgem announces tough new policy to clear 'zombie projects' and cut waiting time for energy grid connection. Press release, 13 November 2023. Available at: https://www.ofgem.gov.uk/publications/ofgem-announces-tough-new-policy-clear-zombie-projects-and-cut-waiting-time-energy-grid-connection (accessed 25 August 2026).
• Ofgem (2025) Connections Reform: Summary Decision Document (TMO4+ package). 15 April 2025. Available at: https://www.ofgem.gov.uk/sites/default/files/2025-04/Summary-Decision-Document-TMO4-package.pdf (accessed 25 August 2026).
• Patterson, D., Gonzalez, J., Le, Q., Liang, C., Munguia, L.-M., Rothchild, D., So, D., Texier, M. and Dean, J. (2021) Carbon Emissions and Large Neural Network Training. arXiv:2104.10350. Available at: https://arxiv.org/abs/2104.10350 (accessed 25 August 2026).
• ScienceDaily (2023) Powering AI could use as much electricity as a small country. 10 October 2023. Available at: https://www.sciencedaily.com/releases/2023/10/231010133607.htm (accessed 25 August 2026).
• Shehabi, A., Smith, S.J., Hubbard, A., Newkirk, A., Lei, N., Siddik, M.A.B., Holecek, B., Koomey, J., Masanet, E. and Sartor, D. (2024) 2024 United States Data Center Energy Usage Report. LBNL-2001637. Berkeley, CA: Lawrence Berkeley National Laboratory. Available at: https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf (accessed 25 August 2026).
• UK Parliament (2025) Delivering AI Growth Zones. Written ministerial statement HCWS1057, 13 November 2025. Available at: https://questions-statements.parliament.uk/written-statements/detail/2025-11-13/hcws1057 (accessed 25 August 2026).
• Uptime Institute (2024) Uptime Institute Global Data Center Survey 2024. Available at: https://datacenter.uptimeinstitute.com/rs/711-RIA-145/images/2024.GlobalDataCenterSurvey.Report.pdf (accessed 25 August 2026).
• Uptime Institute (2025) Uptime Institute Global Data Center Survey 2025. Available at: https://datacenter.uptimeinstitute.com/rs/711-RIA-145/images/2025.Annual.Survey.Report.pdf (accessed 25 August 2026).