Among the costs of running a data centre day to day, electricity is usually the largest. But for an AI data centre, the running costs are only part of the picture. The servers cost far more to buy than the power they use, and they have to be replaced every few years. Counted over its life, an AI data centre spends more on its computers than on its electricity.
Electricity is the largest running cost
Digital Realty, a large data centre owner, calls utilities its largest expense category. In 2024 it reported utilities of about US$1,333 million out of property-level operating expenses of about US$2,519 million, around 53% (calculated). It adds that the vast majority of this cost is passed directly through to its customers, which limits its own exposure to power prices. Equinix lists utility costs including electricity among the largest parts of its cost of revenues, alongside depreciation, rent for leased sites, bandwidth, staff, repairs and security.
But the computers cost more
Epoch AI's model of a one-gigawatt AI data centre estimates operating costs of about US$907 million a year, of which energy is the largest part at about US$594 million. Yet it estimates that servers dominate total cost at US$5 billion per year, or 60% of the total, because they are expensive and are replaced every few years.
| Cost in Epoch AI's model of a 1 GW AI data centre | US$ million per year |
|---|---|
| Energy | 594 |
| Taxes | 143 |
| Maintenance | 120 |
| Labour | 40 |
| Water | 6 |
| Total operating cost | 907 |
| Servers, spread over their life | about 5,000 |
The model assumes electricity at 8.34 US cents per kWh and a PUE of 1.14.
Efficiency still matters
Even if electricity is not the biggest lifetime cost, saving it matters to whoever pays the bill. The Uptime Institute's 2025 survey reports an average annual PUE of 1.54 among respondents, almost unchanged for six years, and that costs remain operators' leading concern.
What is not certain
- Electricity prices differ greatly between countries and contracts; the share of electricity in running costs follows them.
- The Epoch AI figures are from one model of a large US facility, not a measured site.
Sources and method
- Digital Realty Trust, Inc., Annual Report on Form 10-K for the year ended 31 December 2024. https://www.sec.gov/Archives/edgar/data/1297996/000155837025001424/dlr-20241231x10k.htm
- Equinix, Inc., Annual Report on Form 10-K for the year ended 31 December 2024. https://www.sec.gov/Archives/edgar/data/1101239/000162828025005126/eqix-20241231.htm
- Epoch AI (Michael and Cottier), One-gigawatt AI data center up-front CapEx breakdown, 14 May 2026. https://epoch.ai/data-insights/ai-datacenter-cost-breakdown
- Uptime Institute, Global Data Center Survey 2025, July 2025. https://datacenter.uptimeinstitute.com/rs/711-RIA-145/images/2025.Annual.Survey.Report.pdf?version=0
Method: each link was opened on 29 October 2026 and every figure was taken from the page or filing cited beside it, in the currency stated there; shares marked as calculated are simple arithmetic on the cited figures. Company and consultancy figures describe their own data and models; they are not recommendations.
Read next
This article provides general information, not investment advice or an offer of securities.