Data drought: AI’s water problem

Water scarcity has become one of the most pressing environmental and economic risks of the 21^st Century. No longer only a humanitarian concern, it is a material business issue affecting supply chains, operating costs, regulation, and corporate licence to operate.

According to the Carbon Disclosure Project (CDP), an estimated 60% of countries are at risk of having unsustainable water resource usage by 2050.

At present, around 2.4bn people live in water stressed regions – areas withdrawing 25% or more of renewable freshwater supplies – and within this number, approximately 800mn people live under high or critically high-water stress, where more than 75% of available water is used each year.

These conditions inevitably lead to water insecurity where basic freshwater needs cannot be met reliably, and the effects of water scarcity are often immediate, local, and financially disruptive.

Technology is thirsty

Against this backdrop, the technology sector has become a surprisingly large and growing water consumer. Semiconductor manufacturing relies on vast volumes of ultrapure water, while data centres withdraw billions of gallons annually for cooling.

A United Nations study estimates that AI could account for up to 6.6bn cubic metres of water use by 2027, which is nearly two thirds of England’s annual water consumption.

The supply of water and electricity could prove the limiting factor for AI development, rather than the supply of high-end semiconductors. How companies – including our portfolio holdings Alphabet, Microsoft, and Amazon – address these needs will determine which reap the benefits of the new technology.

Cooling is the primary driver of this demand. Data centres use water for on-site evaporative cooling but also consume embedded water through the electricity they rely on, most of which still comes from thermoelectric power plants.

In the US, almost a third of freshwater withdrawals is for thermoelectric power, which accounts for about 78% of net electricity generation.

This figure is expected to grow as data centre electricity demand rises; the US Department of Energy forecasts US data centre direct water use to increase between 17–33% per annum to 140-275bn litres in 2028. However, Barclays analysts suggest that excludes nearly 800bn litres of indirect water use.

The global energy demand of hyperscalers has grown approximately 25%+ per annum for six consecutive years and Alphabet, Meta, and Microsoft now account for just over 1% of total US electricity demand, with data centres accounting for >4% of US electricity demand. This is why it is important to consider indirect water use related to powering the data centres.

In response, Microsoft, Amazon, and Alphabet have committed to replenishing more water than they consume by 2030, but progress remains uneven.

For example, Microsoft reports a leading Water Usage Effectiveness (WUE) figure of 0.3 L/kWh, significantly below the industry average of 1.8 L/kWh. These targets typically exclude indirect water linked to electricity, understating real consumption.

As power demand becomes the bottleneck for data centre expansion, nuclear has returned to the discussion, particularly small modular reactors (SMRs).

Microsoft recently joined the World Nuclear Association and is collaborating on SMRs as well as other advanced nuclear technologies. While this may provide energy relief, SMRs are not expected to come online in the US until 2030.

Stakeholder scrutiny is also intensifying. Community opposition, drought-linked resource tensions, and local policy intervention have delayed or halted an estimated $64bn in data centre investment.

For example, Alphabet’s projects in Chile and Uruguay were forced into redesign or regulatory review due to environmental concerns. Emerging legislation in certain US states proposes shifting utility cost burdens from households to operators, signalling a new phase of governance risk.

The convergence of AI expansion, rising water stress, and community scrutiny means the operating environment for hyperscalers will look very different over the next 2–5 years. The key issue is not simply whether companies have targets, but whether they can deliver operational models that scale sustainably without triggering local resistance, regulatory intervention, or reputational damage.

Balancing ambition

In September 2025, OpenAI announced deployment plans that would equate to the power need of ~17 nuclear plants or nine Hoover dams – roughly the same amount of electricity needed to power >13mn US homes. This raises a fundamental question: how will society and infrastructure systems absorb demand without accelerating water scarcity and carbon emissions?

The rapid expansion of AI and cloud infrastructure is placing new levels of pressure on finite water resources. While efficiency gains and corporate pledges are welcome, absolute consumption is rising. Without credible solutions, the sector risks missing environmental commitments, facing regulatory and legal challenges, and encountering physical limits on growth.

Some argue AI will help solve environmental problems; others note it may worsen them by accelerating energy and water demand. Reality will be a combination, but only if companies act at the pace of the problem, not the pace of their ambitions.