Energy and Sustainability

How much energy data centers use

A data center turns electricity into computation and heat — in practice, almost all of the electricity becomes heat that must be removed (see cooling and airflow). The amounts are large enough to matter at national level. Statistics Netherlands (CBS) estimated that Dutch data centers used 5.1 billion kWh in 2024, 4.6% of all electricity used in the country (preliminary figures, published December 2025), up from 3.3% in 2021. Of the roughly 200 data centers, about 45 large ones account for around 90% of that use.

The pressure is rising. AI training and inference hardware packs far more power into a rack than ordinary servers do, and cooling and power distribution have to follow (see power). At the same time, the electricity grid in many regions cannot supply new large connections quickly. That combination — growing demand, limited grid capacity and public attention — is why energy is now measured, reported and regulated, and why operators work on efficiency in every layer.

Power usage effectiveness (PUE)

PUE (power usage effectiveness) is the total energy used by a data center divided by the energy used by its IT equipment, measured over a full year. A PUE of 1.0 would mean every kilowatt-hour goes to the IT equipment; a PUE of 1.5 means that for every 1 kWh the servers, storage and network use, another 0.5 kWh goes to cooling, power conversion losses, lighting and the rest of the building. The metric was created by The Green Grid and is standardized as ISO/IEC 30134-2; the second edition, ISO/IEC 30134-2:2026, replaced the 2016 version.

The Uptime Institute's annual survey gives a weighted industry average of 1.56 in 2024, 1.54 in 2025 and 1.52 in 2026; Uptime noted in 2025 that the headline figure had virtually stood still for six consecutive years. Respondents' largest sites average 1.45, and 23% work at a site below 1.3. Older sources quoting "about 1.8" describe the industry of around 2014.

PUE must be measured over a year because it changes with the weather: in summer the cooling works harder. This program computes the monthly and annual PUE, and the water metric explained below, from a year of meter readings at a mid-size site:

# Monthly meter readings for one year at a mid-size data center.
# total_mwh: utility meter (everything); it_mwh: UPS output to the IT load;
# water_m3: water used by the cooling towers.
MONTHS = ["Jan", "Feb", "Mar", "Apr", "May", "Jun",
          "Jul", "Aug", "Sep", "Oct", "Nov", "Dec"]
total_mwh = [1012, 915, 1010, 990, 1045, 1090, 1150, 1145, 1060, 1020, 975, 1005]
it_mwh = [720, 652, 718, 700, 712, 715, 722, 720, 705, 710, 690, 715]
water_m3 = [610, 560, 700, 900, 1450, 2100, 2650, 2600, 1700, 950, 640, 600]

for month, total, it in zip(MONTHS, total_mwh, it_mwh):
    if month in ("Jan", "Apr", "Jul", "Oct"):
        print(f"{month}  PUE {total / it:.2f}")

pue = sum(total_mwh) / sum(it_mwh)
wue = sum(water_m3) * 1000 / (sum(it_mwh) * 1000)   # liters per kWh of IT energy
overhead = sum(total_mwh) - sum(it_mwh)
print(f"Annual PUE {pue:.2f}  (IT {sum(it_mwh):,} MWh, overhead {overhead:,} MWh)")
print(f"Annual WUE {wue:.2f} L/kWh  ({sum(water_m3):,} m3 of water)")

Jan  PUE 1.41
Apr  PUE 1.41
Jul  PUE 1.59
Oct  PUE 1.44
Annual PUE 1.46  (IT 8,479 MWh, overhead 3,938 MWh)
Annual WUE 1.82 L/kWh  (15,460 m3 of water)

Picking the best month (1.41) for a brochure is a classic way to flatter a site. The annual figure is the honest one, and it is computed from the energy totals, not by averaging the monthly ratios.

What PUE does not show

• IT efficiency. PUE treats every kilowatt-hour that reaches a server as useful. Idle servers, forgotten VMs and ten-year-old hardware all count as "IT load". Consolidating them saves real energy but can make the PUE worse, because the fixed overhead is now divided by a smaller IT load.
• Moving load across the line. Raising the inlet temperature saves cooling energy, but above a certain point the server fans spin faster, and fan energy counts as IT energy. The PUE improves while the total may not.
• Water and carbon. A site can reach a low PUE with evaporative cooling that uses a lot of water, or run on a coal-heavy grid. That is why PUE is reported next to WUE and renewable energy figures.
• Comparison between sites. Climate, measurement points and load all differ; PUE is most useful for tracking one site over time.

Water usage effectiveness (WUE)

WUE (water usage effectiveness), also defined by The Green Grid, is the annual water used on site for cooling and humidification, in liters, divided by the energy used by the IT equipment in kilowatt-hours. The unit is L/kWh and the ideal value is 0. In the example above, evaporative cooling uses about four times as much water in July as in January, which is exactly when water is scarcest.

Water and energy pull in opposite directions. Evaporating water is a very efficient way to remove heat, so evaporative systems lower the PUE but raise the WUE; dry coolers and closed liquid loops use little water but more electricity on hot days. Which trade-off is right depends on the local climate and water supply. Note also that site WUE leaves out the water used to generate the electricity in the first place.

Carbon intensity and renewable energy

The climate impact of a data center depends less on its kilowatt-hours than on where they come from. The carbon intensity of electricity is the amount of greenhouse gas emitted per unit produced, usually given in grams of CO₂ (or CO₂-equivalent) per kWh. It differs between countries and changes hour by hour on the same grid: low when the sun shines and the wind blows, high when gas or coal plants cover the evening peak.

Data center operators often claim "100% renewable energy". How they count that matters. Renewable energy matching is the accounting that pairs a consumer's electricity use with renewable generation it has bought, through certificates (in the EU, guarantees of origin) or long-term power purchase agreements (PPAs) with wind and solar parks. Annual matching only requires the yearly totals to be equal; hourly (time-based) matching requires renewable supply in the same hour as the consumption, which is far harder and much closer to what the grid physically delivers.

This program shows the difference for one illustrative day: a data center with a steady 1 MW load and a solar contract that almost covers the day's total:

# One illustrative day, in 3-hour blocks starting at midnight.
load_mwh = [3.0] * 8                                   # a steady 1 MW data center
solar_mwh = [0.0, 0.0, 2.0, 6.5, 8.0, 5.0, 0.5, 0.0]   # contracted solar park
grid_g_per_kwh = [180, 170, 210, 150, 120, 200, 330, 260]  # grid carbon intensity

total_load = sum(load_mwh)
total_solar = sum(solar_mwh)
matched = sum(min(load, sun) for load, sun in zip(load_mwh, solar_mwh))
from_grid = [load - min(load, sun) for load, sun in zip(load_mwh, solar_mwh)]
grid_tonnes = sum(mwh * g for mwh, g in zip(from_grid, grid_g_per_kwh)) / 1000
plain_tonnes = sum(mwh * g for mwh, g in zip(load_mwh, grid_g_per_kwh)) / 1000

print(f"Load {total_load:.1f} MWh, solar bought {total_solar:.1f} MWh")
print(f"Matched over the day:     {min(total_solar / total_load, 1):.0%} renewable")
print(f"Matched per 3-hour block: {matched / total_load:.0%} renewable")
print(f"Grid power still used: {sum(from_grid):.1f} MWh, "
      f"about {grid_tonnes:.1f} t CO2 (without solar: {plain_tonnes:.1f} t)")

Load 24.0 MWh, solar bought 22.0 MWh
Matched over the day:     92% renewable
Matched per 3-hour block: 48% renewable
Grid power still used: 12.5 MWh, about 2.9 t CO2 (without solar: 4.9 t)

On paper the day is 92% solar; in reality less than half of the consumption is covered at the moment it happens, and the evening and night run on grid power, including the evening peak when the grid is at its most carbon-intensive. Real hourly matching uses 24 hours, not eight blocks, and a mix of wind, solar and storage to fill the gaps — but the lesson is the same: ask which kind of matching a "renewable" claim uses.

Residual heat reuse

Almost all the energy a data center uses comes out as heat, which is normally released into the air by cooling towers or dry coolers. Residual heat reuse (waste heat reuse) means capturing that heat and delivering it to someone who needs it: a district heating network, offices, homes, greenhouses or swimming pools.

The obstacle is temperature. Air-cooled data centers deliver heat at a low temperature, usually too low for a heating network to use directly, so a heat pump has to raise it, which costs electricity. Liquid-cooled racks return warmer water, which makes reuse easier. A heat customer also needs to be nearby, and must be able to take the heat year-round — including summer, when nobody heats their home. That is why heat reuse works best where a data center is planned together with a heating network from the start.

Some countries now require it. Germany's Energy Efficiency Act (EnEfG), for example, sets a minimum share of reused heat for new data centers, rising over time (reported as 10% for sites starting operation from July 2026, 15% from July 2027 and 20% from July 2028), together with PUE limits for new and existing sites; check the current text before relying on the exact figures.

Reporting and rating

Reporting duties for data centers differ per region. In the EU, the Energy Efficiency Directive (EED), Directive (EU) 2023/1791, is the EU law that sets energy efficiency targets for the EU and its member states and, in its Article 12, requires data centers with an installed IT power demand of at least 500 kW to publish and report key information on their energy and water performance every year to a European database.

The Dutch context

Grid congestion

In large parts of the Netherlands the electricity grid is full: grid congestion means large consumers wait, sometimes for years, for a new or heavier connection. Since October 2024 a prioritization framework of the regulator ACM (Authority for Consumers and Markets) has given socially important projects priority on the waiting list; it was widened on 1 January 2026, and from 1 July 2026 small consumers in congested areas can also end up on a waiting list. Data centers around Amsterdam are named among the causes of congestion in the region. For new data centers, available grid capacity is now a harder limit than available land; for existing ones, a planned expansion of IT load can be blocked by the connection, which is why it belongs in capacity planning years ahead.

Siting rules for hyperscale data centers

Since 1 January 2024, a national rule (an amendment of the Besluit kwaliteit leefomgeving, Staatsblad 2023, 492) restricts where hyperscale data centers may be built. It defines them as data centers with more than 10 hectares of built area and a connection capacity of at least 70 MW, and allows them only at two locations: Eemshaven (Het Hogeland, Groningen) and Agriport A7 (Hollands Kroon, Noord-Holland).

Because both criteria must be met, a compact, multi-story data center with a large connection can fall outside the definition: news reports in January 2026 described a planned 78 MW project in Amsterdam on about 2.3 hectares. In March 2026 the Tweede Kamer (House of Representatives) adopted a motion asking the government to tighten the definition. As of October 2026 no new rule is in force; the definition is under review.

What operators can do

Most of the levers are ordinary operational work, and many of them save money as well as energy:

• Remove what is not used. Idle servers and forgotten VMs still draw power; proper decommissioning and consolidation through virtualization reduce the IT load itself, which no PUE improvement can match.
• Stop over-cooling. ASHRAE recommends server inlet temperatures of 18–27 °C for common air-cooled equipment. Many rooms still run at the bottom of that range or below it, which wastes cooling energy.
• Separate hot and cold air with aisle containment, blanking panels and sealed cable openings, so the cooling does not have to compensate for mixing.
• Refresh hardware. Newer servers often do the same work with far fewer machines and watts (see hardware refresh).
• Measure well. Meters at the utility intake, the UPS output and per rack PDU are the basis for PUE, for EU reporting and for finding where the energy goes.
• Plan for heat and water when building or expanding: liquid cooling, heat reuse and dry cooling are much cheaper to design in than to retrofit.