Information dive: Energy grid knowledge exhibits delivery of AI in UK datacentres
Electrical energy grid “land grabs” to make sure capability forward of graphics processing unit (GPU) shipments. Additions to capability as ever-larger datacentres change on. And the arrival, deployment and “burning in” of Hopper and Blackwell GPUs.
These are a few of the issues we are able to see in knowledge from electrical energy grid supplier UK Energy Networks (UKPN), which supplies electrical energy utilisation charges taken half hourly for 96 datacentre websites inside its area. This stretches from the datacentre hotspots of west London and Docklands, south-eastwards to Kent, Surrey and Sussex, and consists of all of Essex and East Anglia.
Pc Weekly analysis carried out in March 2026 analysed Electrical energy Efficiency Certificates knowledge to establish datacentre areas and capacities, discovering 80 datacentres within the UKPN area with a mixed capability of 798MW.
The UKPN dataset begins at first of 2023 and runs to April 2026. Altogether, it includes nearly 5.4 million rows and covers datacentres categorised by the voltage they import from the grid: extra-high voltage (12 websites), excessive voltage (60), and low voltage (24).
Utilisation ratios are calculated by evaluating precise electrical energy import – measured half-hourly by good meter – in opposition to the utmost capability booked by the client.
Website voltage corresponds to the seemingly dimension of the datacentre. Low-voltage (LV) websites are sometimes 400V connections for smaller enterprises, edge datacentres and server rooms. Excessive voltage (HV) within the UKPN knowledge seemingly refers to 11kV or 33kV connections to colocation hubs and mid-range datacentres. Additional-high-voltage (EHV) connections are 33kV, 66kV, or 132kV, and canopy hyperscale campuses and rising synthetic intelligence (AI) factories.
The common utilisation fee for all UKPN knowledge is simply over 20% of booked capability. Additional-high-voltage websites use the least of their allotted provide (12%), whereas low-voltage websites use extra (18%).
Now we have taken the information factors and break up them by voltage ranges that correspond to datacentre dimension. These are then proven in a chart the place dips, plateaus, spikes, and so forth are seen. These correspond to real-world occasions that embody activation of recent datacentre capability, will increase in booked capability prematurely of AI GPU deployments, the heatwave of July and August 2025, precise deployment and “burning in” of AI datacentre infrastructure, and a rush to beat Ofgem’s “use it or lose it” directive in early 2026.
Keep in mind that the chart exhibits utilisation fee, so whereas in some instances the reason for a spike is likely to be apparent – equivalent to elevated energy draw for cooling throughout a heatwave – different adjustments may not be so apparent, equivalent to a rise in booked capability that adjustments the ratio.
Now let’s take a look at a few of the key occasions that present up within the knowledge.
Small websites can’t cope with the heatwave: July and August 2025
One of the vital pronounced spikes within the inexperienced (low-voltage website) knowledge occurred in July and August 2025, when meteorological knowledge exhibits the UK confronted 4 distinct heatwaves between June and August. Temperatures reached 35.8°C in Kent on 1 July, whereas August noticed sustained excessive night-time temperatures.
The spikes within the chart present {the electrical} signature of smaller air-cooled datacentres and server rooms desperately making an attempt to maintain temperatures down. Not like giant hyperscale websites that use liquid cooling, smaller websites are probably the most weak to local weather stress.
They sometimes depend on legacy direct enlargement air-conditioning. When ambient temperatures exceed 30°C, these items draw two or 3 times their regular energy simply to keep up the established order. Electrical energy utilisation ratio spikes right here as a result of complete facility energy skyrockets whereas IT load stays static, leading to a short lived collapse of energy utilization effectiveness (PUE).
Capability will increase: Ratios decline – late 2023 into 2024
A lot of the story revealed by the information is that of large-scale websites including capability, and reserving extra electrical energy provide in anticipation of deliveries and deployment of GPUs. This began to occur in late 2023 and into 2024.
The sudden drop-off in September and October 2023 for the biggest datacentres – the crimson line – is probably going the results of capability coming on-line.
So, when a hyperscale website prompts a brand new part, its import capability – the whole energy booked from the grid (the denominator) – jumps immediately. However as a result of the IT load (the numerator) solely populates as servers are bodily racked and “burned in” over subsequent months, the utilisation proportion seems to crash.
On the identical time, we see gradual (blue line) and extra pronounced (inexperienced line) declines in late 2023 and into 2024. That’s a probable indication that bigger, newer, extra environment friendly services are pulling general-purpose workloads away from the older small and mid-range services.
CBRE’s forecast for large-scale colocation take-up in London in 2024 was forecast to hit 130MW, and as this new, environment friendly capability got here on-line, it “emptied” the much less environment friendly websites.
The truth that the smallest datacentres (seemingly enterprise-owned or older retail colocation websites) took till Could 2024 to settle from utilisation ratios of 0.2 to 0.15 signifies an extended migration interval. Not like the hyperscalers, which transfer workloads in huge, software-defined blocks, smaller organisations usually tend to be sure by bodily {hardware} lifecycles.
Websites that activated capability in late 2023 included:
- Iron Mountain’s LON-2, with the primary part of its eventual 27MW of capability in Slough, was confirmed as operational on the finish of 2023. Its grid capability was seemingly booked into the UKPN system in September 2023 as a part of its pre-commissioning part.
- Equinix’s LD11x/LD13 expansions, in the meantime, have been particularly designed to lure the massive three hyperscalers, and moved from building to “obtainable capability” in late 2023.
GPU provide constraint and the ‘West London land seize’
From late 2023, the lead occasions on Nvidia GPU clusters turned very prolonged, with Omdia reporting 36 to 52 weeks for H100-based servers. On the identical time, datacentre operators scrambled for grid provide, usually reserving approach past what they might instantly use in order that they could possibly be able to deploy Hopper GPUs once they lastly arrived in mid- to late-2024. That’s another excuse grid utilisation seems to plummet in late 2023 within the chart.
In July 2022, the Larger London Authority (GLA) despatched a warning to builders, stating that main new planning purposes in Hillingdon, Ealing and Hounslow would face delays of as much as a decade, with some connection dates pushed again so far as 2035 or 2037.
That breaking level was triggered by the intense focus of datacentres alongside the M4 hall. By mid-2023, datacentres accounted for 18% of complete demand in West London. Transmission-level capability and native distribution reached full capability as a result of builders had legally “reserved” future energy capability and left zero headroom for brand spanking new housing or industrial tasks.
Monetary stories from datacentre Actual Property Funding Trusts (REITs) like Equinix and Digital Realty again this up. Of their 2023 annual stories, these corporations famous document “backlog” ranges, the place capability was signed and dedicated however not but billing.
Within the knowledge, a excessive backlog means the distribution community operator (UKPN) has allotted the ability, however the servers aren’t spinning, and this matches the 2023-2024 trough the place utilisation ratios settled at a decrease baseline in contrast with the pre-AI land seize.
The explanation the ratio didn’t bounce again instantly is that AI density is extra environment friendly than legacy density. A rack of H100s would possibly draw 40kW, but it surely replaces dozens of legacy racks drawing 5kW every. As hopper GPUs lastly arrived in mid- to late-2024, they stuffed that phantom capability, however as a result of grid capability had been aggressively over-booked in 2023, utilisation ratios remained low. The trade successfully constructed a buffer that it’s nonetheless filling as we speak.
GPU deployment, AI datacentre burn-in: 2024 and 2025
The peaks of mid-2024 match with the seemingly deployment of Nvidia Hopper (H100/H200) GPUs. The Hopper era was the primary GPU to hit a 700W Thermal Design Energy (TDP) – ie, the wattage for which its cooling needed to be designed. An HGX H100 node of eight GPUs attracts roughly 10.2kW. The spikes within the knowledge from late-2024 seemingly signify the initiation of large-scale coaching runs the place hundreds of those items synchronise their energy draw.
These represented a shift in datacentre energy dynamics, from the steady-state attracts of the earlier Ampere (A100) era to extremely risky, high-density profiles.
Late 2025 marked a pivot from Hopper to Blackwell’s high-density liquid-cooled necessities. This transition is mirrored within the UKPN telemetry as a definite shift from steady-state energy draw to the risky, “peaky” plateaus of large-scale Blackwell coaching epochs.
The “mountain vary” within the chart starting in November 2025 marks the power-on month for the UK’s first Nvidia Blackwell (B200) clusters. That is the signature of preliminary mannequin coaching, which is a particularly intensive, continuous course of.
Every B200 GPU has a base 1,000W TDP, configurable as much as 1,200W. The GB200 Grace-Blackwell Superchip, in the meantime – which shipped from late 2024 – mandated direct-to-chip liquid cooling to handle its excessive density.
The smoking gun right here is the launch of the Nebius AI Cloud cluster at Ark Information Centres’ Longcross Park in Surrey. This went reside in November 2025 with a number of thousand Nvidia Blackwell GPUs and a 16MW signature.
The EHV line stays elevated and jagged via March 2026, reflecting the excessive, sustained draw of “epochs” and “checkpoints” throughout frontier mannequin pre-training.
One last spike: Use it in order that they don’t lose it?
The large spike in late March 2026 coincides with the Ofgem Demand Connections Reform deadline of 13 March 2026.
Within the face of huge will increase in electrical energy demand, not least by datacentre operators – and with the demand queue hovering to 125GW by June 2025 – Ofgem had proposed more durable monetary checks and “use it or lose it” guidelines to clear the queue. Giant-scale operators with parked capability have been incentivised to point out energy draw to show their tasks have been “viable” and “strategically vital” earlier than the brand new guidelines might claw again their unutilised megawatts.
