Artificial intelligence (AI) is driving a global construction boom in data centers, but the enormous computing power required to train and operate AI models comes with a hefty price tag, from expensive computer chips and electricity to water-intensive cooling systems.
An AI data center is essentially a giant computer factory, housing thousands of specialized processors that work together to train artificial intelligence models, answer questions and generate images, videos and other content.
Unlike conventional data centers, which mainly store information and run websites and business applications, AI facilities need enormous computing capacity to process billions of calculations simultaneously, creating substantial demands for electricity, cooling and physical infrastructure.
GPUs top the expense list
The first major expense is the computer hardware, particularly graphics processing units (GPUs), which are designed to handle the parallel calculations required by AI systems.
Advanced AI processors can cost tens of thousands of dollars each, while a large facility may contain tens of thousands of them, alongside servers, high-speed networking equipment and storage systems.
Building the facility itself is another substantial investment, requiring large sites, reinforced buildings, specialist electrical equipment, backup generators, transformers and connections to the electricity grid.
However, the biggest challenge does not end when construction is completed: keeping thousands of powerful processors running around the clock creates enormous ongoing demands for energy and cooling.
Electricity: the biggest running expense
AI processors consume electricity continuously, and the larger the computing system, the greater the demand on the power grid.
The International Energy Agency (IEA) estimates that data centers worldwide consumed about 415 terawatt-hours (TWh) of electricity in 2024, equivalent to approximately 1.5% of global electricity consumption.
The IEA’s latest outlook projects that data center electricity consumption could almost double, from 485TWh in 2025 to 950TWh in 2030, with AI-focused facilities accounting for a particularly rapid increase. IEA
To understand the scale, consider a hypothetical AI data center with an average electricity demand of 100 megawatts (MW), roughly comparable to the continuous power consumption of a small town.
Running such a facility around the clock would require 2.4 million kilowatt-hours (kWh) of electricity every day, or 876 million kWh annually.
At an illustrative industrial electricity price of US$0.05 to US$0.10 per kWh, its annual electricity bill would range from US$43.8 million to US$87.6 million, before additional charges and other operating expenses.
A facility requiring 1 gigawatt (1,000MW) would consume 10 times as much electricity, potentially generating an annual power bill of US$438 million to US$876 million at the same tariffs.
These figures illustrate why technology companies are negotiating long-term electricity supply agreements, investing in renewable energy and exploring nuclear power to secure reliable supplies.
Water: the hidden cost of keeping AI cool
Electricity is not the only resource consumed by AI infrastructure, as powerful processors generate so much heat that they need sophisticated cooling systems to prevent overheating.
Some facilities use conventional air conditioning, while others circulate chilled water through cooling towers or use liquid cooling systems that transfer heat directly away from processors.
Water is particularly useful because evaporation removes large quantities of heat, but the process means some of the water is lost rather than returned to the local supply.
The IEA estimates that a 100MW hyperscale data centre in the US consumes approximately 2 million litres of water daily when direct cooling and indirect water use associated with electricity generation are included. BUILD UP
That is equivalent to roughly 800 Olympic-sized swimming pools of water over a year, assuming a pool holds 2.5 million liters, or the daily water consumption of approximately 6,500 households.
The distinction between direct and indirect water use is important because a data center also has a water footprint through the power stations supplying its electricity and the manufacturing of its computer chips.
Google’s 2025 environmental report, covering 2024, disclosed that its data centers consumed approximately 7.8 billion US gallons (29.5 billion liters) of water during the year. Google Sustainability+1
Water consumption varies considerably according to the climate, cooling technology and electricity supply, with facilities in hot, water-stressed regions potentially facing greater environmental and operational challenges.
A useful comparison is that the water consumed by one large AI data center every day could supply thousands of households, making its location particularly important in regions already experiencing water shortages.
To put the annual figure into perspective, 730 million litres would fill approximately 290 Olympic-sized swimming pools, rather than 800.
The cost of building an AI data center
The initial investment can be just as daunting as the running costs.
According to property consultancy JLL, the average global construction cost of a data center was forecast to reach US$11.3 million per MW in 2026, covering the building and its basic infrastructure but excluding land and active computing equipment.
The technology fit-out, including servers and AI hardware, can add substantially more, with JLL estimating that it could cost as much as US$25 million per MW for AI infrastructure.
This means a 100MW AI data centre could involve billions of dollars in construction and equipment investment, depending on its design and technology requirements.
Why the investment matters
The economics of AI data centers present technology companies with a difficult balancing act: they must spend billions upfront, secure reliable electricity and water supplies, and generate enough revenue to recover their investments.
The processors also become outdated over time, requiring expensive upgrades, while depreciation, maintenance, staffing and financing add to the overall cost.
The challenge is therefore not simply building the world’s most powerful AI infrastructure, but ensuring that the services running on it generate enough income to justify the enormous expense.
For investors, this raises a fundamental question: will the revenue generated by AI grow quickly enough to justify the unprecedented investment in the infrastructure powering it?
That question could ultimately determine whether the current AI infrastructure boom produces sustainable profits or leaves technology companies with expensive facilities that are difficult to operate profitably.