United Nations researchers project that AI workloads will double global data center power consumption and double global data center water consumption by 2030, putting hard numbers on the resource curve that hyperscalers have been climbing since the launch of ChatGPT. The forecast frames the five-year window as the period in which AI moves from a meaningful share of data center demand to the dominant driver of it.
The doubling applies to both inputs at once. Electricity feeds the racks of accelerators that train and serve large models; water, in most current designs, feeds the cooling systems that keep those racks within thermal spec. The UN's framing treats the two as joint constraints rather than independent ones, because the design choices that reduce one tend to raise the other.
The projection lands in a year when the biggest model developers have been pre-committing capacity years out. Alphabet's $85B stock sale to fund Google's AI buildout, which AI Chat Daily covered last week, is the kind of capital move that only pencils if power and water can actually be sourced at the sites where new campuses are planned.
Key facts
- 01UN researchers project AI workloads will double global data center power consumption by 2030.
- 02The same analysis projects data center water consumption will also double by 2030.
- 03The forecast covers the five-year window from today through the end of the decade.
- 04Power and water are now the two binding constraints on where new AI capacity can be built.
Power is the more visible constraint. Grid interconnect queues in the US, Ireland, and parts of the Nordics already run multiple years, and several US utilities have warned that load growth from AI campuses is outpacing planned generation. Doubling demand by 2030 means either accelerating new generation, including nuclear restarts and small modular reactors, or rationing connections among the largest customers.
Water is the more contested one. The number that matters at a site level is gallons per day of evaporative loss, which depends heavily on cooling design. Closed-loop and air-cooled facilities consume meaningfully less water than open-loop evaporative ones, and several recent hyperscaler builds have moved toward closed-loop precisely because water rights, not capital, are the binding local constraint.
The UN figures are global aggregates, which means the local picture varies. A campus in Iceland drawing on geothermal power and ambient cooling has a very different profile from one in Arizona drawing on a stressed aquifer. National regulators are starting to ask for facility-level reporting on both inputs, and the EU's data center reporting rules under the Energy Efficiency Directive are the most developed example.
For the AI companies themselves, the projection is less a warning than a planning input. OpenAI, Anthropic, Google, Meta, and xAI have all signed multi-gigawatt power deals over the past 18 months, and the deals increasingly include water-use commitments and recycling provisions. The companies that get to 2030 with the most contracted power and the cleanest water story will have a structural cost advantage on inference.
There are credible counterweights to the doubling forecast. Model efficiency continues to improve, with inference cost per token falling sharply each year as architectures, quantization, and specialized chips mature. If efficiency gains outrun demand growth, the curve flattens. The UN researchers' projection assumes current efficiency trends hold but does not assume a step-change improvement of the kind that DeepSeek-style training optimizations could deliver.
The 2030 doubling figure is a useful planning anchor for anyone building or financing AI infrastructure, but it is not destiny. The bigger story is that power and water are now first-order variables in AI strategy, on par with chip supply. Companies that treat them as procurement problems rather than PR problems will keep shipping; those that don't will run into siting fights that delay capacity by quarters, not weeks. The UN's contribution here is to put a number on a constraint the industry has already started pricing in.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
