There is a growing narrative in the media — and increasingly on LinkedIn — that AI-driven data centres are pushing us toward “water bankruptcy.”

In some versions, this is framed as an existential crisis driven almost entirely by AI’s rapid growth.

We wanted to understand whether this claim stands up to scrutiny. Rather than relying on headlines, we spent time looking at the underlying mechanics of data centre water usage — specifically water used for cooling, which is where most of the concern is focused.

Scope of This Analysis

Data centre water usage is complex and influenced by multiple factors. For clarity, this analysis looks only at water used for cooling operations. It explicitly excludes:

1. Water used in semiconductor manufacturing
2. Water used in electricity generation

Those are important topics, but they deserve separate treatment.

Data Centres in Scope

Using publicly available information, we looked at ten of the largest US data centres by square footage and estimated power capacity:

1. Google — Columbus Cluster (Ohio): 500 MW
2. Google — Omaha Cluster (Nebraska): 500 MW
3. Meta AI — Columbus Site (Ohio): 500 MW
4. Amazon AWS — Project Rainier (Indiana): 420 MW
5. Microsoft Azure — Atlanta Site (Georgia): 350 MW
6. xAI — Colossus 2 (Tennessee): 400 MW
7. Microsoft Azure — Fairwater Campus (Wisconsin): 350 MW
8. AWS — Mississippi AI (Mississippi): 300 MW
9. OpenAI / Crusoe Energy — Stargate Project (Texas): 200 MW
10. xAI — Colossus 1 (Tennessee): 300 MW

This represents approximately 3.8 gigawatts of capacity.

For context, total US data centre capacity is estimated at 40–50 GW and is expected to roughly double by the end of the decade. This sample therefore represents less than 10% of current US capacity.

Key Assumptions

Because operators do not publicly disclose detailed operating data, we made several high level assumptions:

1. Average utilisation: Facilities operate at 50% of maximum power capacity over a year.
2. Cooling efficiency:

• Traditional systems used ~2.97 litres of water per kWh.
• Modern systems have reduced this to ~0.3 litres per kWh.

3. Water recycling: Modern data centres operate on approximately 90% closed loop cooling, meaning most water is reused rather than consumed.

These assumptions materially change the outcome and are critical to interpreting the results.

What the Numbers Suggest

Under these assumptions, the ten data centres analysed require:

1. A significant initial water intake to fill cooling systems
2. Ongoing annual water replenishment to account for losses

Even when extrapolated across the broader US data centre footprint — and then globally — the numbers are undeniably large. The US is estimated to account for around 50% of global data centre capacity.

However, two things stand out:

1. Cooling efficiency improvements have already reduced water intensity dramatically.
2. Ongoing water consumption is far lower than headline figures often imply, due to closed loop reuse.

There will, of course, be many older, less efficient facilities still in operation. This analysis should therefore be viewed as a minimum baseline, not a best or worst case scenario.

Water vs Energy: The Bigger Constraint

If AI driven data centre capacity continues to double every few years, water will certainly remain an important constraint.

But it may not be the first one we hit.

Generating renewable energy at the required scale — reliably, affordably, and fast enough — appears to be a significantly larger challenge than cooling water availability alone.

The Question We’re Not Asking

One final point rarely discussed: how AI capacity is actually being used.

Public usage data suggests that a large share of consumer AI activity is focused on chat based interactions — effectively a more powerful version of traditional search.

If consumers were required to pay an explicit environmental impact cost for AI usage, would demand remain this high? Or is current behaviour driven largely by convenience, with little visibility of the underlying resource trade offs?

So, Are We Headed for Water Bankruptcy?

AI-driven data centres do use a lot of water. That is not in dispute.

But the picture is more nuanced than many headlines suggest. Cooling efficiency has improved substantially, water reuse is now standard in modern facilities, and energy generation may prove to be the tighter constraint.

The real risk may not be AI itself — but how thoughtlessly we choose to use it.