The AI buildout has a water problem that almost no forecast prices in. Hyperscale cooling and the upstream chip supply chain both draw heavily on freshwater, and they are concentrating that draw in exactly the basins least able to absorb it.
This report models datacenter water demand to 2030 at basin resolution — direct cooling, plus the embedded water in the semiconductors inside each rack — and maps it against regional water stress to show where growth runs into a hard physical ceiling.
What’s inside
- A bottom-up demand model covering 240 announced and operational datacenter sites
- Embedded-water accounting that links each site back to its chip supply chain
- Basin-level stress overlays for North America, Europe, and East Asia
- Three demand scenarios to 2030, with siting implications for each
- A short-list of basins where new capacity is effectively capped today
The core finding
Roughly a third of planned hyperscale capacity is sited in basins already classed as high- or extreme-stress — before the chips inside it are even accounted for.
Once embedded water is included, the picture worsens: cooling is the visible draw, but fabrication is the larger one over a chip’s life. Treating the two separately, as most public models do, understates the true exposure of an AI siting decision.
Who it’s for
Infrastructure investors underwriting datacenter or fab assets, developers selecting sites, and operators stress-testing existing portfolios against water risk. If a siting or capacity decision depends on water, this is the briefing it should start from.
Need this scoped to a specific portfolio or region? Commission a custom briefing →
Related reports
Advanced-Node Reshoring: Capacity, Subsidy & Risk
What the announced fabs actually deliver once power, water, and labor constraints are priced in. With a 2027 scenario set.
Interconnection Queues & the Industrial Buildout
Why the queue — not generation — is now the binding constraint, and which regions clear fast enough to host new heavy load.