Data Centres and Power Transformers: Why a Flat Load Ages Insulation Faster

Most transformers that now feed data centres were specified for a different kind of customer. A utility substation or an industrial supply sees a daily cycle: high in the day, low at night, with hours of cooling between peaks. A large data centre does not sleep. Its servers, its cooling plant and its power conversion run at a steady, high share of capacity for weeks, and since 2024 its largest sites have also shown a new behaviour: power that swings in seconds with the rhythm of AI workloads, and that can disappear from the grid all at once.

This article puts numbers on what that means for a power transformer. It starts with the demand data, Europe and the United States side by side, then works through the thermal arithmetic with the loading guide’s own equations, and ends with the evidence an owner can collect to know whether a unit is coping.

How fast is data centre demand growing, and where?

The International Energy Agency estimates that data centres used about 415 TWh in 2024, around 1.5% of world electricity, and that this will more than double to about 945 TWh by 2030 in its Base Case. The growth is concentrated: the United States and China account for nearly 80% of it. The comparison below uses the same sources for both sides of the Atlantic.

The two markets differ in scale more than in direction. The United States builds more, faster, and spends several times more. Europe grows from a smaller base, with tighter grids in its traditional hubs and a policy push that is still being negotiated. In the US, the Lawrence Berkeley National Laboratory puts data centres at 4.4% of national electricity in 2023 and at 6.7% to 12% by 2028.

Italy shows how fast requests can outrun the grid. Terna, the transmission operator, counted about 30 GW of data centre connection requests at the end of 2024, more than twenty times the figure of 2021, with about 18.6 GW in Lombardy alone and nearly half of that in the province of Milan. Its July 2026 results put the total at 70 GW in December 2025 and 94 GW in June 2026. Most of these requests will never be built, but each one that is will need transformers, and many will lean on substations that already exist.

Why does the shape of the load matter more than the peak?

Paper insulation ages with temperature, and the relationship is steep. IEC 60076-7:2018 expresses it as a relative ageing rate referred to the winding hot spot: for non-thermally upgraded paper the rate doubles every 6 K, with a value of 1.0 at 98 °C; for thermally upgraded paper the standard uses an Arrhenius expression with 1.0 at 110 °C. At 122 °C, upgraded paper ages 3.3 times faster than at its reference; at 140 °C, 17 times faster.

What matters for life is the integral of that rate over time. A conventional load spends its night hours well below peak, and those hours age the paper very little. A data centre load removes the night. To see the size of the effect, we ran the difference-equation model of IEC 60076-7 (clause 8.2.3) with the parameters of the standard’s own worked example in Annex I: an ONAF transformer with a 45 K top-oil rise, a 35 K hot-spot gradient, a loss ratio of 8, and time constants of 150 minutes for the oil and 7 minutes for the winding. Both profiles peak at 0.9 p.u., the ambient averages 20 °C with a ±5 °C daily swing, and the model runs in one-minute steps until the daily cycle repeats.

The peak temperatures are almost identical, 1.2 K apart, so a check based on the maximum hot spot would see no difference. The ageing is 2.3 times higher. Both values are still below 24 equivalent hours per day, the rate at which a unit uses its insulation life as designed, so neither profile is dangerous at these conditions. The point is how much margin the flat profile spends, and how quickly the rest disappears when other factors are added.

What do ambient temperature and harmonics add?

Every kelvin of ambient goes straight into the hot spot. IEC 60076-7 also warns that the yearly average is the wrong input for ageing: because ageing is exponential, the standard uses a weighted ambient, and in its own example a site with a 15.0 °C yearly average has a weighted ambient of 20.4 °C. Data centres are often built where land and grid capacity are available, not where the climate is mild, and our article on ambient temperature and climate scenarios shows how much the input can move.

Harmonic currents add losses that the nameplate does not see. IEEE C57.110-2018 treats winding eddy losses as rising with the square of the harmonic order, and other stray losses with an exponent of 0.8 as a conservative estimate. Its harmonic loss factor, F_HL, measures how much a given current spectrum multiplies the eddy losses. The standard covers liquid-immersed and dry-type power and distribution transformers, not rectifier transformers, which have their own IEEE C57.18.10.

The harmonic case uses a spectrum chosen for the arithmetic, not a measurement: 20% fifth, 12% seventh, 7% eleventh and 5% thirteenth harmonic. It gives F_HL = 3.5 and an rms current 3% above the fundamental. If eddy losses are 10% of the I²R losses at rated current, the total load losses rise by about 23%, and in this model that costs as much as ten degrees of extra ambient. Modern server power supplies and active front-end converters can keep distortion much lower. The lesson is to measure the spectrum at the transformer secondary rather than assume it, and our note on harmonic loading and transformer heating explains how to use the result.

What happens when the load moves in seconds?

AI training does not draw power smoothly. Measurements published by the Lawrence Berkeley National Laboratory show the characteristic square-wave pattern of training jobs, with all nodes dropping power together at synchronisation points. A 2026 study of measured workloads at a US national laboratory found that a 16-node configuration swung between about 12 kW and 48 kW. Recent power-system studies show that such fluctuations, multiplied across a large site, can excite poorly damped oscillation modes of the grid.

The faster risk is a sudden loss of load. On 10 July 2024, a fault on a 230 kV line in Northern Virginia, cleared normally with six automatic reclosing attempts in about 82 seconds, coincided with the loss of about 1,500 MW of load. The North American Electric Reliability Corporation found that it was entirely data centre load, disconnected by the customers’ own protection and controls, and that local voltage rose to 1.07 p.u. before operators switched out capacitor banks. A similar fault in April 2026 caused 500 MW of sustained loss, and another event in July 2026 was still being analysed in September.

For a transformer, these events are not thermal problems. They are voltage swings, switching operations and repeated fault clearing and reclosing in a short time, and they put work on tap changers and on the insulation of units near the site. They also make the load history harder to anticipate, which is one more reason to keep measured data rather than assume a profile.

Which evidence shows that a unit is coping?

A transformer that feeds a data centre needs the same evidence as any other unit, read with the duty in mind. The thermal margin, the paper and the tap changer deserve more attention than they usually get.

Two habits make this evidence useful. Read trends, not single samples, because the duty is steady and its effect builds slowly. And sample more often than the routine plan when a unit moves from a conventional customer to a data centre: the first year under the new duty is the baseline for the next twenty.

How does Seetalabs help owners of these transformers?

Seetalabs works on the evidence side of the problem: turning the results an owner already has into decisions that can be checked. Three free tools address the questions above directly, in the browser and without sending data, except for the climate lookup:

• Site Climate & Loading brings historical climate data for the site, from NASA POWER, into the loading check, so the ambient is not a guess.
• Moisture & Loading relates paper moisture to the loading margin, the link explained in our article on bubble inception temperature.
• Fleet DGA Triage ranks a fleet by IEEE C57.104-2019 status from a laboratory CSV, to see which units need a closer look first.

For the fleet as a whole, RONIN by Seetalabs reads the laboratory reports you already receive, with a review before calculation, and returns each reading with the standard, edition and clause behind it: DGA status to IEEE C57.104-2019, fault interpretation with Duval methods and IEC 60599:2022 ratios, oil properties read one by one against IEC 60422:2024, paper condition from 2-FAL as a range for the degree of polymerisation, and trends drawn on real sampling dates. For a unit that has just taken on a data centre, that history is what lets you see a change early enough to plan the outage in your window, not in the middle of the customer’s busiest month. With large power transformers taking up to four years to replace, that window is worth protecting.

What can be done this quarter?

• List the transformers that serve or back up data centre load, with their rating, cooling, paper type and the share of their capacity the data centre uses.
• For each, compare the current daily profile with the profile the unit was specified for, and recalculate ageing with the weighted ambient of the site.
• Measure the harmonic spectrum at the secondary of the units with the highest load factor.
• Add furans and a moisture estimate to the next oil sample, and shorten the sampling interval for the first year under the new duty.
• Check tap changer counters and diverter gases on units near large sites, and review how protection and reclosing interact with the customer’s ride-through settings.

A data centre is the most stable customer a transformer can have. That is exactly why it uses insulation life without ever giving it a night off.

Sources

• IEA, Energy and AI, World Energy Outlook Special Report, 2025 (chapters 2 and 5, tables 2.4 and figures 2.12, 5.13).
• Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report, December 2024; US Department of Energy, DOE releases new report evaluating increase in electricity demand from data centers, 20 December 2024.
• European Commission, Commission sets course for Europe’s AI leadership with an ambitious AI Continent Action Plan, 9 April 2025; Proposal for the Cloud and AI Development Act, COM(2026) 502, 3 June 2026.
• The White House, Executive Order 14318, Accelerating Federal Permitting of Data Center Infrastructure, 23 July 2025.
• Terna, Piano di Sviluppo 2025, programmazione territoriale e interventi di connessione, section 2.2; H1 2026 consolidated results, 29 July 2026.
• NERC, Incident Review: Considering Simultaneous Voltage-Sensitive Load Reductions, 8 January 2025; Reliability Insights: large computational load risks due to voltage sensitivity, September 2026.
• X. Chen et al., Electricity demand and grid impacts of AI data centers: challenges and prospects, arXiv:2509.07218, 2025 to 2026.
• R. Vercellino et al., Measurement of generative AI workload power profiles for whole-facility data center infrastructure planning, arXiv:2604.07345, 2026.
• J. Krajacic et al., Forced oscillations in power systems induced by data centers hosting AI workloads, arXiv:2609.27698, 2026.
• IEC 60076-7:2018, Loading guide for mineral-oil-immersed power transformers (clauses 6.3, 7.1, 8.2.3 and Annex I); IEEE C57.110-2018, Recommended practice for establishing transformer capability when supplying nonsinusoidal load currents.